Content uploaded by Min Hou
Author content
All content in this area was uploaded by Min Hou on Sep 07, 2023
Content may be subject to copyright.
The Influence of Visual Features in Product Images on Sales Volume: A Machine Learning
Approach to Extract Color and Deep Learning Super Sampling Features
Min Hou*, Yongpeng Tang
School of Business Administration (MBA), Zhejiang Gongshang University, Hangzhou 310018, China
Corresponding Author Email: houmin@mail.zjgsu.edu.cn
https://doi.org/10.18280/ts.400415
ABSTRACT
Received: 10 April 2023
Revised: 27 June 2023
Accepted: 6 July 2023
Available online: 31 August 2023
With the rise in online shopping, the role of product images in shaping consumer purchase
decisions has been accentuated. Despite burgeoning research in this domain, there remains
a lacuna in comprehensively understanding the relationship between specific visual
attributes, such as color and target shape, in product images and the consequent sales
volume. To bridge this gap, the relationship between product image colors and sales volume
on online platforms was examined, and color attributes from these images were
systematically extracted. Furthermore, an exploration was undertaken into the association
between the target shape of product images and sales volume. Deep Learning Super
Sampling (DLSS) features from these images were distilled, aiming to furnish a more
precise market analysis. Through leveraging advanced machine learning techniques, this
study not only augments the academic comprehension of consumer behavior but also
proffers strategic insights for online retail practitioners. The methodological approach
ensures a targeted marketing direction and facilitates informed product design strategies.
Keywords:
online shopping, product images, sales
volume, visual feature extraction, color
analysis, target shape, DLSS features,
machine learning
1. INTRODUCTION
In the wake of the Internet's rapid evolution, online
shopping has seamlessly woven itself into the fabric of daily
life, establishing itself as a pivotal facet of contemporary
consumption [1-4]. Within this milieu, the role of product
images has transcended mere informational conveyance,
evolving to become a determinative element in consumer
purchasing decisions. It has been observed that visual
attributes of these images, encompassing elements like color,
shape, and texture, exert a significant influence on consumer
psychology, from emphasizing product features to galvanizing
purchasing intent [5-8].
Yet, the intricate relationship between product images and
sales volume, especially the potential to leverage visual
attributes to enhance sales, remains an enigma. An exploration
into this interplay not only straddles the realms of marketing
and consumer behavior but also bears significant socio-
economic implications [9, 10]. Through rigorous scrutiny of
the nexus between image attributes and sales, invaluable
insights into consumer predilections and purchasing patterns
can be gleaned, furnishing a robust foundation to refine
marketing strategies and render product designs more enticing
[11-14]. This discourse further stimulates technological and
theoretical advancements in corollary domains like visual
marketing and artificial intelligence [15, 16].
Regrettably, despite burgeoning interest from both
academia and industry, current research methodologies exhibit
palpable limitations. A predominant swath of conventional
analyses is observed to hover on a descriptive veneer, often
sidelining granular feature scrutiny within images such as
nuanced color juxtapositions, target shape differentials, among
others [17, 18]. Moreover, extant methodologies are
predominantly anchored in quantitative statistical paradigms,
manifestly deficient in their capacity to process and discern
visual data, thus obscuring subtle yet pivotal visual nuances
[19-22].
This research endeavor seeks to illuminate the multifaceted
relationship between product images and sales volume. The
investigative focus bifurcates into two key strands: The initial
segment underscores the correlation between product image
color on online shopping platforms and sales volume, with
endeavors being made to extrapolate color features through
avant-garde machine learning techniques. Subsequent
emphasis is channeled toward discerning the interrelationship
between the target shape of product images and sales volume,
culminating in the meticulous extraction of the DLSS features
from these images. Through this prism, this study not only
furnishes a novel theoretical framework but is also poised to
unlock substantial pragmatic applications. Such insights stand
poised to equip online retailers with the acumen to intuitively
resonate with and invigorate consumer purchasing
propensities while concurrently seeding fertile grounds for
scholarly and practical pursuits in allied domains.
2. RELATIONSHIP BETWEEN PRODUCT IMAGE
COLOR AND SALES VOLUME AND COLOR
FEATURE EXTRACTION
Distinct colors have been observed to elicit varied
emotional responses. Recognized as a primary conduit for
visual information assimilation, color often emerges as a
paramount factor commanding consumer attention. The
deployment of vibrant hues or color palettes, aptly tailored for
a specific target demographic, can captivate consumer
attention, potentially elevating product engagement and
amplifying purchase inclination. Moreover, it has been noted
Traitement du Signal
Vol. 40, No. 4, August, 2023, pp. 1469-1477
Journal homepage: http://iieta.org/journals/ts
1469
that seasonal variations and prevailing trends may shape the
color schema of product images. Illustratively, a surge in sales
is often associated with products aligning with a season's
trending colors, underscoring the dynamic and indirect
relationship between color and sales. Consequently, the nexus
between product image color on online shopping platforms
and sales volume can be characterized as multi-tiered and
multi-dimensional. While the direct relationship
predominantly gravitates towards the color's immediate
influence on consumer psychology and behavior, the indirect
relationship casts a broader net, encompassing socio-cultural
and visual considerations.
Through the extraction of color features, quantifiable and
analyzable data is derived from the color information inherent
in product images. Simultaneously, by examining which hues
or color combinations resonate most potently in sales and
discerning correlations between color attributes and sales
metrics, latent consumer needs and psychological inclinations
are unveiled. Such insights hold the potential to delineate
which color attributes might be pivotal in bolstering sales.
Furthermore, they can offer a reflection of evolving market
trends and seasonality, bearing significant implications for
informed product design and strategic marketing initiatives.
2.1 Preprocessing of product images
In the realm of online shopping platforms, the veracity and
consistency of product images are frequently compromised
due to variables like shooting environment, lighting conditions,
and camera parameters. Such variables can induce a color cast
in images, thereby diminishing their representational fidelity.
Furthermore, highlights present in these images can skew
color distribution and texture information, posing challenges
to subsequent feature extraction. Noise, which may be
introduced through factors such as suboptimal shooting
quality or compression errors, can further impede the effective
extraction of color features. To navigate these challenges, a
rigorous preprocessing regimen must be applied to these
images. Essential steps in this regimen include color cast
correction, highlight mitigation, and denoising, which
collectively fortify the authenticity, consistency, and precision
of image quality. This, in turn, paves the way for more
rigorous color feature extraction and subsequent sales volume
analysis.
Color casts in product images, regrettably, are a pervasive
phenomenon. Such color deviations can arise from a plethora
of sources. Different light sources, ranging from sunlight to
incandescent lamps, can engender varying color temperatures
and light intensities, which are implicated in the onset of color
casts. Additionally, discrepancies stemming from divergent
camera brands, models, shooting modes, or white balance
settings can lead to color inconsistencies. Image manipulation
processes, such as color adjustment and compression, may
further exacerbate these issues, producing distortions like
supersaturation or diminished contrast. Such color casts not
only potentially breed mistrust in consumers due to disparities
between product images and their real-world counterparts but
also can amplify refund rates, thereby escalating operational
costs for merchants.
In this study, color casts in product images were rectified
utilizing a principle rooted in color constancy. Specifically,
calibration was performed using established reference objects
or color cards, thus ensuring the precision of color restoration.
The correction formula for the R channel is delineated as
follows:
( ) ( ) ( )
22
, , ,
e e e
U z t iU z t cU z t=+
(1)
For the Gray World hypothesis to hold true, the ensuing
formula must be met:
( ) ( )
1 1 1 1
,,
L B L B
eh
z t z t
U z t U z t
= = = =
=
(2)
Subsequently,
( ) ( ) ( )
2
1 1 1 1 1 1
, , ,
L B L B L B
e e h
z t z t z t
U z t c U z t U z t
= = = = = =
+=
(3)
Eqs. (2)-(3) were transposed into a matrix format, taking
into account variables ω and c. With the preservation of the G
channel's integrity, a pixel-by-pixel color correction was
meticulously executed in the R and B channels of the product
images as illustrated:
( )
( )
( )
2
,,
,
,,
,
ee
z t z t
h
zt
MAX U z t cMAX U z t
MAX U z t
+
=
(4)
Another pervasive complication encountered within
product images on digital commerce platforms is the presence
of multiple highlights. Such manifestations can be attributed
to excessively potent light sources, suboptimal angles, or the
employment of multiple lighting fixtures. Additionally, certain
product materials, notably metal, glass, or glossy plastics,
inherently facilitate the formation of light reflections,
consequently producing highlights. The presence of excessive
highlights can potentially obfuscate and convolute product
images, detracting from the consumers' focus and
comprehension of product intricacies. Moreover, these
highlights might mask or distort pivotal product data and
nuances, rendering the product evaluation process challenging
for potential buyers.
In this research, highlight elimination within product
images was approached through an area reconstruction method.
The crux of this procedure revolved around the precise
determination of the highlighted area's central point. Initially,
a color image underwent a conversion process to its grayscale
counterpart, thereby streamlining subsequent computations. A
judicious threshold was subsequently instituted, segmenting
the image into highlighted and non-highlighted sectors. The
highlighted regions were demarcated, and each contiguous
highlight block was discerned. For each of these blocks, the
geometric center or center of gravity—essentially the central
point of the highlights—was meticulously calculated.
Designating r as the central highlight point and using [s n v f r
d h g u] to represent grayscale pixel values, s-r-u, v-r-h, d-r-f
and n-r-g were delineated as four linear trajectories
intersecting at the point r. These lines bifurcated the image
domain into dichotomous sections. The ensuing formula was
employed to compute the absolute value delineating the
discrepancy between average values of the dual sub-regions
associated with each line:
1470
( ) ( )
( ) ( )
( ) ( )
( ) ( )
1 3 3
2 3 3
3 3 3
4 3 3
f s n v g u h
f s n h v d u
f s n f u f g
f n v d f g h
= + + − + +
= + + − + +
= + + − + +
= + + − + +
(5)
In further endeavors to refine image quality, both opening
and closing operations—fundamental processes in
morphological filtering—were executed. These operations
aimed at diminishing or tempering the highlighted areas. By
amalgamating center point positioning with the opening and
closing procedures, the method predicated on area
reconstruction demonstrated efficacy in attenuating, if not
eradicating, multifarious highlights in product images. Such
methodologies not only preserved salient image features and
minutiae but also subdued or nullified the perturbing
highlighted domains. Consequently, a palpable enhancement
in the authenticity and quality of product images on e-
commerce platforms was observed. Designating ○ and ● as
symbols for the opening and closing operations respectively,
the following algorithmic approach was adopted:
Firstly, an area opening reconstruction for the product
image, denoted as dp, was undertaken with ho=d○N acting as
the marker diagram, resulting in an output image labeled dp.
Subsequently, the product image dp underwent an area closing
reconstruction, where hov=dp○N served as the marker, leading
to the output image dpv.
Noise—another prevalent impediment in product images on
digital commerce platforms—could wield significant
influence over sales metrics. Factors such as low-light
shooting conditions or the deployment of subpar camera
hardware might culminate in a disproportionate balance
between image signals and noise signals. Furthermore, image
compression, a tactic employed to economize on storage and
expedite loading times, when taken to extremes, has been
identified as a precursor to information attrition and palpable
noise manifestation. Not only does this noise compromise the
visual finesse of the image, making it appear coarse or out-of-
focus, but it also skews colors, contours, and intricate details
of the merchandise, potentially leading to consumer
misjudgement of the product.
Representing the image by d(z,t), the noise by b(z,t), and the
output subsequent to image perturbation by h(z,t), the additive
noise expression for product images was characterized as:
( ) ( ) ( )
, , ,h z t d z t b z t=+
(6)
For noise attenuation, judicious selection of denoising
filters, encompassing mean filtering, median filtering, or
Wiener filtering, was undertaken. The efficacy of the noise
reduction was either visually assessed or gauged using quality
indices such as the Peak Signal-to-Noise Ratio (PSNR),
thereby confirming that the images' denoised state met
stipulated criteria.
2.2 Color feature extraction in product images
In the realm of visual psychology, the color red has
frequently been linked with a plethora of emotions,
encompassing vitality, passion, and desire, to name a few. By
delving into the prevalence of the red component within
product imagery, potential emotional triggers elicited in
consumers can be discerned. Concurrently, the extraction of
this red component becomes instrumental in elucidating
certain visual attributes of products, exemplified by the vibrant
hues observed in ripened fruits. Assessing both the distribution
and intensity of the red element within images facilitates a
critical evaluation of the efficacy with which product images
accentuate these attributes.
In this investigation, the relative red component information
from product images available on digital commerce platforms
was meticulously extracted in the RGB domain. The ensuing
formula delineates the computation process for the relative red
component information for a given image:
ME ME
ME
RED GREY
RL GREY
−
=
(7)
Hue, revered as a fundamental facet of color, succinctly
conveys the primary color tone of an entity. A rigorous
analysis of hue ensures an authentic representation of a
product's color traits, thereby proffering consumers with a
more genuine and precise depiction. It is worthy of note that
disparate hues can invoke varied psychological responses and,
in turn, influence purchasing inclinations. For instance, the
resonance of warmth and solace is commonly attributed to
warm colors, while their cooler counterparts tend to evoke
sentiments of professionalism and serenity. A systematic
exploration into the nexus between hue and sales metrics
yields insights into consumer purchasing proclivities.
Figure 1. Illustrative portrayal of products encapsulating
diverse hues
Subsequently, hue characteristics were diligently extracted
from the Hue Saturation Brightness (HSV) color spectrum of
product images present on online retail interfaces, as depicted
in Figure 1. The underlying algorithm for this extraction is as
follows:
( )
( ) ( )
( ) ( )( )
( ) ( )
( ) ( )( )
1
2
1
2
0.5
360
0.5
gV
R G R B
COS R G R B G B
BG
R G R B
COS R G R B G B
BG
−
=
− + −
−
− + − −
− + −
− + − −
(8)
1471
3. ANALYSIS OF PRODUCT IMAGE SHAPES IN
RELATION TO SALES VOLUME AND DLSS
FEATURE EXTRACTION
In the intricate landscape of digital commerce, the
correlation between the target shape of product images and
sales volumes has been observed. First impressions of
consumers are often swayed by the shape of product images,
with distinctive and captivating forms garnering heightened
attention, which potentially augments viewership and
purchase intent. In certain scenarios, the morphology of
products inherently conveys their functionality. For instance,
the form of items like furniture, tableware, or tools can be
posited to mirror their practical usability and ergonomic
comfort. It is surmised that consumers might gauge a product's
appropriateness for their requirements, pivoting largely on its
shape. Through meticulous analysis of these relationships,
product images can be more accurately tailored, enabling more
efficacious communication of product specifics. This not only
enhances the potential for sales conversions but also fortifies
brand image and market positioning.
Figure 2. The procedural representation of DLSS descriptor generation
While DLSS is predominantly associated with augmenting
resolution and quality in image rendering, within the context
of this study, DLSS is perceived as a distinctive target shape
feature, drawing upon deep learning technologies for the
extraction and analysis of product imagery. Employing DLSS
feature extraction allows for precise identification of target
shapes amidst multifarious product images, even when
confronted with multiple perturbing factors. This facilitates a
robust analytical foundation, elucidating the nexus between
shape and sales metrics. Notably, DLSS does not singularly
focus on shape feature analysis but amalgamates other visual
aspects like color and texture. A holistic understanding of how
product imagery steers consumer behavior and purchasing
decisions can thus be achieved through this comprehensive
lens, as delineated in Figure 2.
The DLSS algorithm's implementation was systematically
outlined in the following manner:
Initially, optimal window sizes and sub-window sizes were
identified, catering to the analysis of local shape features
within the image. A sub-window, denoted by Uo, was centered
around the pixel o, with UE serving as the corresponding
search image domain. Within both UE and Uo, the sum of
squared grayscale value discrepancies among all pixels was
determined, thereby encapsulating the local texture and shape
data of the image. Here, UE was demarcated by the red
boundary; Uo was represented by the blue square region; and
Uq was specified as the 5×5 window within UE. With ouk and
quk as the pixels in Uo and Uq respectively, the formula to
compute the sum of squared grayscale discrepancies for every
pixel in Uo and Uq was articulated as:
( )
55 2
11
, , 1,5
oq uk uk
uk
AAF o q u k
==
= −
(9)
Subsequently, the aforecalculated sum of squared grayscale
discrepancies underwent normalization, constraining its value
domain within a pre-defined range, like [0,1]. This step was
instrumental in negating absolute grayscale divergences across
images and set the stage for the ensuing analytical phase.
Given cseNO as the grayscale deviation resultant from noise
and lighting interference and cseAU(o) symbolizing the
structure of image blocks in proximity to the pixel o (denoting
the maximum sum of grayscale discrepancies squared for
surrounding regions equivalent in magnitude to Uo), the
subsequent formula emerged:
( )
( )
exp ,o
o
NO AU
AAF
AMAX cse cse o
=−
(10)
In the third phase, the normalized sum of squared grayscale
variations was transposed to the polar coordinate domain,
describing the local shape via the polar coordinates' angle and
radius. Regions in this polar coordinate sphere were
demarcated based on predetermined angle and radius
parameters. Post this delineation, the eigenvalue for each
region, derived from the image blocks, was computed. The
culmination of this step witnessed the amalgamation of
eigenvalues from all sectors into an eigenvector, epitomizing
the image's local shape attributes.
To conclude the DLSS procedure, potential grayscale
fluctuations were obviated by normalizing the eigenvector Dm,
ensuring uniform comparability across disparate images and
regions. Upon completion of these steps, the DLSS descriptor
extraction reached its fruition.
Figure 3. DLSS feature extraction flowchart
1472
To delve deeper into the precise extraction of global
discriminative attributes from product images, integration of
the DLSS feature extraction mechanism with the Spatial
Pyramid Matching (SPM) algorithm was undertaken, the
nuances of which are illustrated in Figure 3.
In the initial step, a series of product images were chosen
and subjected to the DLSS algorithm to garner local shape
features. From these images, DLSS features were aggregated,
culminating in comprehensive DLSS descriptors. This
amalgamation facilitated a more holistic representation of the
image shape. Assuming YF×YF represents the size of each
sliding window and MF denotes the gap between consecutive
sliding windows, a series of Ff sliding windows were derived
from each image as delineated by:
( )
( )
2
1 , 56
f m F F m
F B Y M B= − + =
(11)
In the subsequent phase, the product image underwent
segmentation into several windows. The coordinates of each
sliding window's top-left corner were designated as (1,1) while
its lower right corner was represented by (YF,YF). Within a
square boundary demarcated by coordinates (4,4), (YF-4,4),
(YF-4,YD-4), and (4,YF -4), DLSS descriptors were harvested
from points at an interval of 4. After collating the DLSS
descriptors of M points, a comprehensive set of DLSS features
for corresponding sliding windows was produced. The
eigenvector's length was ascertained as:
( )
2
41
18
cF
rc
MY
MM
=−
=
(12)
Consequently, the u-th product image yielded a DLSS
eigenvector matrix, denoted as FuEFf×Mr, where u[1,b]. An
overarching eigenvector set, F={F1,F2,...,Fb}, was established,
encapsulating features from the entirety of the product images.
In the third step, DLSS eigenvectors from all product
images were aggregated. The eigenvector set, F, served as the
foundation to distill a visual dictionary. This dictionary,
represented by CEBf×Mr, epitomized salient and quintessential
image shape attributes.
The fourth phase witnessed the nearest neighbor matching
algorithm's deployment. Each DLSS eigenvector underwent
juxtaposition with the visual dictionary's nucleus to pinpoint
the closest category. Concurrently, DLSS features were
channelled into their pertinent classes, engendering
categorized shape feature annotations.
The fifth step was marked by the construction of a spatial
pyramid architecture, discerned by varying scales and
orientations. With Lo symbolizing the pyramid's layer count,
LF sub-domains were established, as described by:
( )
21
1
2
o
Lj
Fj
L−
=
=
(13)
Lastly, within each pyramid stratum, prior DLSS features
and the visual dictionary were employed to curate layered and
spatially-tuned shape feature depictions. These layers'
culmination led to the inception of global discriminative
features for product images, denominated as HuE1×Bv, with
parameters u[1,b] and Bv=Bf×LF.
4. ANALYSIS OF EXPERIMENTAL RESULTS
Figure 4. Red component analysis of product images in RGB
color space
Figure 5. Hue distribution of different product images
As elucidated by Figure 4, the red component of product
images in the RGB space is observed to range between 0.125
and 0.28. This component serves as an indicator of the
intensity and luminosity of the red hue within the image.
Historically, red has been linked with attributes such as
warmth, vigor, and allure. Product images characterized by
elevated red components, notably in Samples 5, 6, and 7,
impart a heightened impression of warmth and vigor,
potentially augmenting their ability to seize consumer interest.
Additionally, the red hue often correlates with notions of
luxury, sophistication, and romanticism. For instance,
products exemplified by Sample 7 are perceived by consumers
as being more upscale or apt for exclusive events. Within the
dataset, fluctuations in the red component are found to be
relatively modest, devoid of any pronounced highs or lows.
Such uniformity suggests a coherent color palette across
product images, fortifying the brand's consistent image. Hence,
the red component in product visuals reveals a spectrum of
emotional and visual qualities, influencing both consumer
perception and purchasing inclinations. Through an intricate
examination of the red component metrics, commercial
entities can discern the visual impact of their product imagery,
paving the way for refined market positioning and strategic
campaign design.
The hue distribution for three distinct product categories—
men's apparel, cosmetics, and children's playthings—is
depicted in Figure 5, sampled at various points. It is inferred
from the data that the hue spectrum for men's attire spans from
0.072 to 0.079. The restricted range implies a more consistent
hue profile for this product type. A relatively subdued average
hue is discerned, mirroring the predilection for muted,
1473
understated shades in men's fashion. The absence of
discernable anomalies or extremes bolsters the notion of a
consistent brand persona.
In contrast, cosmetics exhibit a hue range of 0.08 to 0.089,
suggesting a more variegated color palette in cosmetic
imagery. An elevated average hue, potentially linked to the
utilization of vibrant shades to captivate a female audience, is
identified. Certain data points, notably Points 5 and 6, are
characterized by marginally augmented hue values, possibly
indicative of specific promotional campaigns or seasonal
offerings.
Lastly, children's toys manifest the broadest hue spectrum,
ranging from 0.085 to 0.097. This expansive range signifies
the incorporation of diverse, vivacious colors, a strategy
employed to engage youthful audiences. The predominant hue
for such products leans toward the higher end, aligning with
the trend of leveraging vivid hues to capture children's
fascination. Variations in hue within this category might be
attributed to inherent color disparities between subcategories,
such as educational versus recreational toys.
Figure 6. Relationship between product image hue and
product color temperature
Figure 6 delineates the average hues and their affiliated
image color temperatures for three distinct product categories:
men's attire, cosmetics, and children's playthings.
For men's clothing, a hue range of 0.072 to 0.079 and a color
temperature spectrum of 0.6 to 1.68 were identified. An overt
linear relationship between hue and color temperature for this
product category was not observed. Interestingly, a hue peak
of 0.078 was found to correspond with the nadir of the color
temperature, a value of 0.6. This suggests that particular hues
might correlate with color temperatures in men's apparel to
impart specific emotional or stylistic undertones.
In the domain of cosmetics, the hue extends between 0.079
and 0.0845, with the color temperature ranging from 4.5 to 6.2.
Although a straightforward linear correlation between the two
parameters remains elusive, an intriguing pattern was
discerned. Cosmetic images with diminished hues were
consistently associated with lower color temperatures,
whereas those with enhanced hues corresponded to elevated
temperatures. Such a relationship could underline the practice
of utilizing color temperatures in cosmetic visuals to
accentuate product texture and aesthetic.
Lastly, children's toys exhibited a hue span from 0.0845 to
0.0965 and a color temperature range of 5.8 to 8.8. In this
sector, a positive correlation was perceived; as hue values
ascended, color temperatures followed suit. This relationship
could resonate with the industry's proclivity to employ
augmented hues and color temperatures in children’s toys,
aiming to engender a vivacious and captivating visual appeal.
Table 1. Impact of sliding window step size on DLSS feature
extraction performance
Step Size
of Sliding
Window
Indexes
Detection
Rate
False
Alarm
Rate
Overall
Accuracy
AUC
2
0.9512
0.7312
0.7951
0.7649
4
0.9546
0.6821
0.7826
0.7892
8
0.9528
0.6759
0.7862
0.7813
12
0.9713
0.7924
0.7763
0.7239
Table 1 offers an in-depth examination of the repercussions
of varying sliding window step sizes on DLSS feature
extraction performance across product images. Parameters
considered include detection rate, false alarm rate, overall
accuracy, and Area Under Curve (AUC) value.
An initial mild augmentation in the detection rate was noted
as the step size increased, culminating in a pronounced surge
at a step size of 12. This trend suggests that expansive step
sizes might be adept at encapsulating global features, thus
enhancing detection proficiency. In parallel, the false alarm
rate was observed to decrease progressively from step sizes 2
to 8. However, an uptick was recorded at a step size of 12,
indicating potential oversaturation of background noise.
Concerning overall accuracy, a marginal variation was
perceived from step sizes 2 to 8, with a slight decline at a step
size of 12. This behavior intimates that while moderate step
sizes negligibly affect overall accuracy, an excessively large
size could be detrimental. The AUC value exhibited an
increase between step sizes 2 and 4, a minor dip at 8, and a
descent at 12. This pattern underscores the assertion that
diminutive step sizes might be deficient in capturing
comprehensive product attributes, whereas larger sizes risk
integrating excessive background details.
Given the observed trends, a step size of 8 is postulated to
be optimal, harmonizing detection and false alarm rates while
preserving high overall accuracy and AUC value. The choice
of an apt step size emerges as a pivotal determinant in DLSS
feature extraction and merits meticulous calibration based on
the nature of product images and the desired application
context.
Table 2. Impact of sliding window edge length on DLSS
feature extraction performance
Edge
Length of
Sliding
Window
Indexes
Detection
Rate
False
Alarm
Rate
Overall
Accuracy
AUC
8
0.9635
0.7785
0.7745
0.7415
12
0.9562
0.5961
0.7316
0.7366
16
0.9588
0.6741
0.7862
0.7854
20
0.9533
0.5623
0.7316
0.8312
24
0.9521
0.6782
0.7819
0.7789
28
0.9536
0.6615
0.8122
0.7815
32
0.9587
0.7416
0.7995
0.7526
36
0.9521
0.6783
0.7842
0.7892
Table 2 dissects the ramifications of diverse sliding window
edge lengths on the DLSS feature extraction efficiency within
product imagery. Throughout the surveyed range, the
1474
detection rate was found to exhibit minor fluctuations,
underscoring a broad stability and suggesting that edge length
might not critically determine feature capture capability. On
the other hand, the false alarm rate showed a decrease from
edge lengths 8 to 20. A subsequent phase of fluctuation was
noted from 20 to 32, after which an upswing was detected.
This pattern indicates that modest edge lengths might
engender false alarms, with the nadir reached at an edge length
of 20. Of significance is the overall accuracy, which peaked at
an edge length of 28, hinting at an optimal edge length that
balances feature capture and noise introduction. The AUC
value, too, was seen to ascend, reaching its zenith at an edge
length of 20, and then tracing a descent with further elongation,
inferring that an edge length in this vicinity most effectively
optimizes classifier performance.
Drawing from these observations, edge lengths of 20 and 28
are postulated to be paramount in this scenario—the former
excelling in AUC and the latter in overall accuracy. Minimal
edge lengths seemingly limit feature extraction, spiking the
false alarm rate, whereas overly expansive lengths might
incorporate extraneous data, curtailing accuracy.
Table 3. Impact of number of dictionary centers on DLSS
feature extraction performance
Number of
Centers
Indexes
Detection
Rate
False
Alarm
Rate
Overall
Accuracy
AUC
500
0.9615
0.6519
0.8145
0.8169
750
0.9548
0.5862
0.8219
0.8269
1,000
0.9562
0.5963
0.8344
0.8314
1,250
0.9542
0.5846
0.8269
0.8379
1,500
0.9423
0.5846
0.8231
0.8451
1,750
0.9481
0.5961
0.8297
0.8216
2,000
0.9567
0.6155
0.8156
0.8177
Table 3 elucidates the implications of varying dictionary
center counts on DLSS feature extraction efficiency in product
illustrations. As the center count was augmented from 500 to
1,000, the detection rate was observed to maintain relative
stability. An ensuing decline was recorded at 1,500 centers,
followed by a marginal recovery. This suggests that detection
rate augmentation reaches a saturation point within a specific
center count range, and excessive centers might impede
performance. The false alarm rate, upon analysis, appeared to
oscillate with center count increase, but without manifesting a
discernible trend, leading to the conclusion that center count
might not be a pivotal determinant of this metric. A
noteworthy observation pertains to overall accuracy, which
culminated at a center count of 1,000, post which a decline was
recorded. The AUC value exhibited a trend of ascent, peaking
at 1,500 centers, followed by a subtle decline, hinting at
classifier performance optimization within this range but with
potential overfitting beyond.
Interpreting this data suggests a balancing act: a center
count neither too sparse nor too dense optimizes feature
extraction. Within this dataset, the equilibrium seems to lie
between 1,000 and 1,500 centers. An inadequate center count
might overlook critical data, while an overabundance might
incorporate noise, adversely affecting performance.
The number of pyramid layers, integral to image processing
and feature extraction, delineates the hierarchical structure of
the spatial pyramid. Within the purview of DLSS feature
extraction, the performance impact of these layers was
rigorously assessed. From the insights derived from the Table
4, it was observed that the detection rate was augmented for
pyramid layers numbered at 1 and 4. However, a decline was
registered at three layers. Such a trend suggests that feature
discriminability can be amplified when layer count escalates
within certain bounds, but surfeit layers might infuse
redundant data.
Table 4. Impact of number of pyramid layers on DLSS
feature extraction performance
Number of
Pyramid
Layers
Indexes
Detection
Rate
False
Alarm
Rate
Overall
Accuracy
AUC
1
0.9615
0.7789
0.7746
0.7218
2
0.9566
0.6651
0.8125
0.7895
3
0.9314
0.5896
0.8239
0.8452
4
0.9568
0.5523
0.8319
0.8426
It was further noted that the false alarm rate showed a
general decline with an upswing in layer count. Such a pattern
implies an enhancement in feature accuracy with additional
layers, thus mitigating false positives. A distinct observation
was the overall accuracy metric, which escalated with
increasing layers and peaked at four layers. This phenomenon
indicates that a multi-layer spatial pyramid potentially
harnesses a richer spectrum of spatial data, refining
classification accuracy. Concomitantly, the AUC value
displayed a similar upward trajectory with layer augmentation,
corroborating that classifier performance was potentiated by
an expanded layer framework. This further reinforces the
assertion that layer incrementation amplifies the delineation
potential of features.
Drawing upon these inferences, it becomes palpable that
pyramid layering profoundly impacts DLSS feature extraction
efficacy. Typically, feature discriminability and precision are
bolstered by layer augmentation. Yet, a judicious layer count
selection remains imperative, as excessiveness can sow
complexities and superfluous data, potentially undermining
performance.
The experimental outcomes elucidated herein attest to the
multifaceted influences on DLSS feature extraction in product
imagery, encompassing parameters such as sliding window
step size, edge length, dictionary center count, and pyramid
layering. Optimization of these parameters is contingent on
task specificity and dataset characteristics, often necessitating
a careful balance between accuracy, computational speed,
complexity, and overfitting susceptibility. For nuanced
product imagery and distinct commercial contexts, targeted
experimentation and validation might be indispensable, given
the improbable existence of a universally optimal setting.
Through meticulous parameter tuning, DLSS feature
extraction techniques might emerge as pivotal tools in product
image analysis, potentially amplifying the efficacy of product
recommendations and categorizations.
5. CONCLUSION
The relationship between the chromatic properties of
product images on digital commerce platforms and their
consequent sales volumes served as the focal point of the
present investigation. Advanced machine learning techniques
were employed to distill color attributes from these images.
1475
Concurrently, the study shed light on target shape attributes
within product images on e-commerce platforms. This was
achieved by integrating the DLSS feature extraction algorithm
with the SPM algorithm. Evaluations on the influence of
varied parameters on the efficacy of feature extraction were
systematically undertaken through a series of experiments.
In the RGB color space, the red component of the product
images was meticulously analyzed. The hue distributions of
three distinct product types at specific sample points were
depicted, elucidating the mean hues and affiliated image color
temperatures. A comprehensive analysis was then undertaken.
It was discerned that the DLSS feature extraction's
performance in product imagery is modulated by a
constellation of factors. These encompass the sliding window's
step size and edge length, the tally of dictionary centers, and
the depth of pyramid layers. A notable inverse relationship
was identified between step size and the false alarm rate, yet
its influence on detection rate and holistic accuracy presented
complexities. Edge length determination, pivotal for
discerning objects of varied magnitudes, demands bespoke
tailoring for unique tasks. Amplifying the dictionary center
count typically augments accuracy, but the perils of overfitting
were underscored. The stratification of pyramid layers
emerged as a tool to bolster feature discriminability; however,
surfeit layers might inject redundancy.
The methodology delineated in this research avails a robust
instrument for the extraction and comprehension of product
image features. However, the selection of apt parameters
necessitates a delicate equilibrium between intricacy and
efficiency. The particularities of product imagery and
commercial contexts might dictate tailored parameter fine-
tuning. Through judicious parameter adjustments and
optimizations, the technology elucidated herein holds promise
for pragmatic applications, potentially enhancing product
recommendation, categorization, and analytical pursuits.
REFERENCES
[1] Kubal, R., Vernekar, S., Cabral, F., Leitao, F., Fernandes,
S., Padiyar, P. (2021). Augmented reality based online
shopping. In 2021 IEEE Bombay Section Signature
Conference, IBSSC 2021. Gwalior, India.
https://doi.org/10.1109/IBSSC53889.2021.9673423
[2] Wang, M., Li, X., Chau, P.Y. (2021). Leveraging image-
processing techniques for empirical research: Feasibility
and reliability in online shopping context. Information
Systems Frontiers, 23: 607-626.
https://doi.org/10.1007/s10796-020-09981-8
[3] Kangeswaran, V., Vasandarai, D., Eliyas, C., Munsil,
M.M.M., Kodagoda, N., Suriyawansa, K. (2021). A
bilingual audio based online shopping mobile application
for visually impaired and the elderly people. In
TENCON 2021-2021 IEEE Region 10 Conference
(TENCON), Auckland, New Zealand, pp. 658-663.
https://doi.org/10.1109/TENCON54134.2021.9707245
[4] Mousavi, R., Hazarika, B., Chen, K., Razi, M. (2021).
The effect of online q & as and product reviews on
product performance metrics: Amazon. com as a case
study. Journal of Information & Knowledge
Management, 20(01): 2150005.
https://doi.org/10.1142/S0219649221500052
[5] Baltescu, P., Chen, H., Pancha, N., Zhai, A., Leskovec,
J., Rosenberg, C. (2022). ItemSage: Learning product
embeddings for shopping recommendations at pinterest.
In Proceedings of the 28th ACM SIGKDD Conference
on Knowledge Discovery and Data Mining, Washington
DC USA, pp. 2703-2711.
https://doi.org/10.1145/3534678.3539170
[6] Wang, R., Chen, Z., Zhang, M.R., Li, Z., Liu, Z., Dang,
Z., Chen, X.A. (2021). Revamp: Enhancing accessible
information seeking experience of online shopping for
blind or low vision users. In Proceedings of the 2021 CHI
Conference on Human Factors in Computing Systems,
Yokohama Japan, pp. 1-14.
https://doi.org/10.1145/3411764.3445547
[7] Lodkaew, T., Supsohmboon, W., Pasupa, K., Loo, C.K.
(2018). Fashion finder: A system for locating online
stores on instagram from product images. In 2018 10th
International Conference on Information Technology
and Electrical Engineering (ICITEE), Bali, Indonesia, pp.
500-505.
https://doi.org/10.1109/ICITEED.2018.8534871
[8] Ryu, S., Ryu, S. (2021). Feeling excited and fluent: How
consumers respond to the visual appeals of products in
an online shopping environment. Behaviour &
Information Technology, 40(11): 1219-1233.
https://doi.org/10.1080/0144929X.2021.1942989
[9] Zhang, Y., Wei, Z. (2022). An image classification and
retrieval algorithm for product display in E-commerce
transactions. Traitement du Signal, 39(5): 1865-1871.
https://doi.org/10.18280/ts.390547
[10] Kato, T. (2021). Factors of loyalty across corporate brand
images, products, dealers, sales staff, and after-sales
services in the automotive industry. Procedia Computer
Science, 192: 1411-1421.
https://doi.org/10.1016/j.procs.2021.08.144
[11] Agost, M.J., Vergara, M., Bayarri, V. (2021). The use of
new presentation technologies in electronic sales
environments and their influence on product perception.
In International Conference on Human-Computer
Interaction, pp. 3-15. https://doi.org/10.1007/978-3-030-
78321-1_1
[12] Sharma, S., Ahemad, F. (2022). Unposed: Unsupervised
pose estimation based product image recommendations.
In Proceedings of the Second International Conference
on AI-ML Systems pp. 1-8.
https://doi.org/10.1145/3564121.3564126
[13] Zhang, M., Wu, Y., Zhou, X. (2022). Analysis of the
influencing factors of urban sports brand sales volume
based on AHP. Mathematical Problems in Engineering.
2022: Article ID 9625049,
https://doi.org/10.1155/2022/9625049
[14] Naeeni, H.S., Sahin, F., Robinson Jr, E.P. (2023).
Socially responsible product-positioning: Impact of
halo/horns spillover on product image. European Journal
of Operational Research, 308(2): 852-863.
https://doi.org/10.1016/j.ejor.2022.11.045
[15] Steeneck, D., Eng-Larsson, F., Jauffred, F. (2022).
Estimating lost sales for substitutable products with
uncertain on-shelf availability. Manufacturing & Service
Operations Management, 24(3): 1578-1594.
https://doi.org/10.1287/msom.2021.1015
[16] Ekambaram, V., Manglik, K., Mukherjee, S., Sajja,
S.S.K., Dwivedi, S., Raykar, V. (2020). Attention based
multi-modal new product sales time-series forecasting.
In Proceedings of the 26th ACM SIGKDD International
Conference on Knowledge Discovery & Data Mining,
1476
USA, pp. 3110-3118.
https://doi.org/10.1145/3394486.3403362
[17] Tan, C.H., Lee, V.C., Nghiem, J., Laxman, P. (2023).
Compliance analyses of Australia’s online household
appliances. In Proceedings of the Sixteenth ACM
International Conference on Web Search and Data
Mining, Singapore, pp. 1285-1286.
https://doi.org/10.1145/3539597.3575788
[18] Sreedhar, R., Tan, N., Zhang, J., Jin, K., Gregson, S.,
Moreta-Feliz, E., Sadalgi, S. (2022). AIDE: Automatic
and accessible image descriptions for review imagery in
online retail. In Proceedings of the 19th International
Web for All Conference, Lyon France, pp. 1-8.
https://doi.org/10.1145/3493612.3520453
[19] Wang, M., Li, X., Chau, P.Y. (2016). The impact of
photo aesthetics on online consumer shopping behavior:
an image processing-enabled empirical study. In 37th
International Conference on Information Systems (ICIS
2016) Proceedings. Association for Information Systems,
Atlanta, USA, pp. 1005-1016.
[20] Kamollimsakul, S. (2018). Effect of product presentation
formats on thai online consumers' performance,
preference, and predicted product price. In Proceedings
of the 2nd International Conference on Business and
Information Management, Barcelona, Spain, pp. 57-63.
https://doi.org/10.1145/3278252.3278274
[21] Wang, Q., Yang, L., Wang, J., Krishnan, J., Dai, B.,
Wang, S., Ma, H. (2022). SMARTAVE: Structured
multimodal transformer for product attribute value
extraction. In Findings of the Association for
Computational Linguistics: EMNLP 2022, pp. 263-276.
https://doi.org/10.18653/v1/2022.findings-emnlp.20
[22] Salman, I.B.R., Varaprasad, G. (2023). Product
recommendation system using deep learning techniques:
CNN and NLP. In International Conference on Data
Management, Analytics & Innovation, pp. 331-343.
https://doi.org/10.1007/978-981-99-1414-2_25
1477
