Enhance or Leave It: An Investigation of the Image Enhancement in Small Object Detection in Aerial Images

Abstract

Recent years of object detection (OD), a fundamental task in computer vision, have witnessed the rise of numerous practical applications of this sub-field such as face detection, self-driving, security, and more. Although existing deep learning models show significant achievement in object detection, they are usually tested on datasets having mostly clean images. Thus, their performance levels were not measured on degraded images. In addition, images and videos in real-world scenarios often involve several natural artifacts such as noise, haze, rain, dust, and motion blur due to several factors such as insufficient light, atmospheric scattering, and faults in image sensors. This image acquisition-related problem becomes more severe when it comes to detecting small objects in aerial images. In this study, we investigate the small object identification performance of several state-of-the-art object detection models (Yolo 6/7/8) under three conditions (noisy, motion blurred, and rainy). Through this inspection, we evaluate the contribution of an image enhancement scheme so-called MPRNet. For this aim, we trained three OD algorithms with the original clean images of the VisDrone dataset. Followingly, we measured the detection performance of saved YOLO models against (1) clean, (2) degraded, and (3) enhanced counterparts. According to the results, MPRNet-based image enhancement promisingly contributes to the detection performance and YOLO8 outperforms its predecessors. We believe that this work presents useful findings for researchers studying aerial image-based vision tasks, especially under extreme weather and image acquisition conditions

Full text

8
Iğdır Üniversitesi Fen Bilimleri Enstitüsü Dergisi, 14(1), 8-17, 2024
Journal of the Institute of Science and Technology, 14(1), 8-17, 2024
ISSN: 2146-0574, eISSN: 2536-4618
Computer Engineering
DOI: 10.21597/jist.1328255
Research Article
Received: 16.07.2023
Accepted: 08.12.2023
To Cite: Tekin, A. & Bozkır, A. S. (2024). Enhance or Leave It: An Investigation of the Image Enhancement in Small
Object Detection in Aerial Images. Journal of the Institute of Science and Technology, 14(1), 8-17.
Enhance or Leave It: An Investigation of the Image Enhancement in Small Object Detection in Aerial Images
Alpay TEKİN1*, Ahmet Selman BOZKIR1
Highlights:
• YOLOV8
• YOLOV7
• YOLOV6
• Deep Learning
• MPRNet
Keywords:
• Object Detection
• Image Restoration
• MPRNet
• Single Shot Object
Detection
ABSTRACT:
Recent years of object detection (OD), a fundamental task in computer vision, have witnessed
the rise of numerous practical applications of this sub-field such as face detection, self-driving,
security, and more. Although existing deep learning models show significant achievement in
object detection, they are usually tested on datasets having mostly clean images. Thus, their
performance levels were not measured on degraded images. In addition, images and videos in
real-world scenarios often involve several natural artifacts such as noise, haze, rain, dust, and
motion blur due to several factors such as insufficient light, atmospheric scattering, and faults
in image sensors. This image acquisition-related problem becomes more severe when it comes
to detecting small objects in aerial images. In this study, we investigate the small object
identification performance of several state-of-the-art object detection models (Yolo 6/7/8)
under three conditions (noisy, motion blurred, and rainy). Through this inspection, we evaluate
the contribution of an image enhancement scheme so-called MPRNet. For this aim, we trained
three OD algorithms with the original clean images of the VisDrone dataset. Followingly, we
measured the detection performance of saved YOLO models against (1) clean, (2) degraded,
and (3) enhanced counterparts. According to the results, MPRNet-based image enhancement
promisingly contributes to the detection performance and YOLO8 outperforms its
predecessors. We believe that this work presents useful findings for researchers studying aerial
image-based vision tasks, especially under extreme weather and image acquisition conditions
1Alpay TEKİN (Orcid ID: 0009-0001-2858-1228), Ahmet Selman BOZKIR (Orcid ID: 0000-0003-4305-7800),
Hacettepe University, Department of Computer Engineering, Ankara, Türkiye
* Corresponding Author: Alpay TEKİN, e-mail: alpaytekin@hacettepe.edu.tr

Alpay TEKİN & Ahmet Selman BOZKIR
14(1), 8-17, 2024
Enhance or Leave It: An Investigation of the Image Enhancement in Small Object Detection in Aerial Images
9
INTRODUCTION
Object detection is a fundamental task in computer vision that involves identifying and localizing
objects within images. It plays a crucial role in various applications, including autonomous driving,
surveillance systems, robotics, and augmented reality. Over the years, significant advancements have
been made in object detection techniques, particularly with the emergence of deep-learning models
especially convolutional neural networks (CNNs). From a general perspective, object detection
algorithms can be examined under two primary pipelines, two-stage and one-stage approaches.
The two-stage approach performs object detection in two stages: (1) extracting regions of interest
(RoIs) and (2) classifying and regressing the RoIs (i.e. Region of Interest). The R-CNN (Girshick et
al., 2014) uses a selective search algorithm (Uijlings et al.,2013) to extract region proposals. On the
other hand, CNN networks process the proposed regions and extract feature maps. Following the
extraction of feature maps, the SVM model classifies each RoI independently. However, this scheme
might be problematic in terms of real-time execution since the inference phase takes too much time
due to the extraction of region proposals. Fast R-CNN (Girshick, 2015) and SPP-Net (He et al., 2015)
improve the R-CNN and reduce the inference time via extracting RoIs from feature maps. They,
nevertheless, still use fixed algorithms and cannot learn how to extract RoIs. The Faster R-CNN (Ren
et al., 2015) enables be trained end-to-end by replacing the selective search with a region proposal
network (RPN) which learns how to extract RoIs. RPN allows the model to learn how to generate
candidate regions proposals and reduces the inference time dramatically. Mask R-CNN (He et al.,
2017) adds a mask prediction branch on Faster R-CNN to detect objects and it predicts their masks
simultaneously. The R-FCNN (Dai et al., 2016) introduces position-sensitive score maps to enhance
object detection quality.
One-stage approaches, in contrast, remove the RoI extraction. Instead, they regress and classify
candidate anchor boxes. The SSD (i.e. Single-Shot Detection) (Liu et al., 2016) extracts feature maps
from anchor boxes by employing several small convolutional filters classifying the bounding boxes
and assigning confidence scores. DSSD (Fu et al., 2017) adds a deconvolution path to SSD, yielding
an improvement in the detection of small objects. The Corner-Net (Law & Deng, 2018) is a key-point-
based approach that can detect objects using corner points. Center-net (Duan et al., 2019), another key-
point-based approach, utilizes a center point in addition to a corner point to capture visual patterns and
eliminate the mispredicted bounding boxes. The well-known algorithm YOLO (Redmon & Farhadi,
2017) divides the image into    grids and predicts the bounding boxes. Each bounding box
generates the corresponding vector containing coordinate points, width, height, and confidence score.
It leverages the intersection of union ND non-maximum-suppression between ground truth and
predicted bounding box to eliminate redundant bounding boxes. There are several vanilla versions and
extensions of YOLO in the literature. The recent one, YOLO8 introduces self-attention and features
pyramids to improve detection quality.
Although these models show promising performance in object detection, they are evaluated on
datasets having only clean images. However, images and videos in real-world scenarios may contain
several natural artifacts such as noise, motion blur, and rain due to various factors such as atmospheric
scattering (Li et al., 2017) or faults in image sensors. Degraded images can significantly reduce the
accuracy of object detection models, especially for small objects. As an illustration, noise can
significantly impact the performance of object detection algorithms, leading to false positives or
missed detections.

Alpay TEKİN & Ahmet Selman BOZKIR
14(1), 8-17, 2024
Enhance or Leave It: An Investigation of the Image Enhancement in Small Object Detection in Aerial Images
10
In this study, we contribute by specifically investigating and evaluating the robustness of small
OD tasks against those artifacts in the absence and presence of image restoration. To achieve this, we
utilized several state-of-the-art YOLO variants to perform the detection of small objects on clean and
degraded images. We first trained the three single-shot OD algorithms (YOLO 6/7/8) with clear
images of the original VisDrone dataset (Cao et al., 2021). Second, we derived three degraded test sets
from the VisDrone test set adding synthetic noise, motion blur, and rain. Followingly, we employed
Multi Stage Progressive Image Restoration (MPRNet) (Rajaei et al., 2023) on degraded test sets to
obtain their restored counterparts. Finally, we evaluated the OD models on (a) clean, (b) degraded, and
(c) enhanced versions. The results show that OD models trained with clean images cannot generalize
well and their performance diminishes on degraded images whereas leveraging MPRNet-based image
enhancement significantly improves the model detection quality on degraded images.
The rest of this paper is organized as follows. We first introduce the employed deep architectures
in the section of Material and Methods. Later, we provide details of the data in the section of Datasets.
Followingly, the metrics we utilized are given in the section of Evaluation Metrics. The section of
Results and Discussion present the results of experiments and provides some discussions on the
findings. The last section concludes the paper.
MATERIALS AND METHODS
You Only Look Once (YOLO)
YOLO is an end-to-end fashioned real-time object detection algorithm that can perform object
detection with a single pass of the network. It divides the image into   grids with equal shapes.
Each cell is responsible for detecting objects that appear within them and predicts bounding boxes
represented by a feature vector containing center points coordinates, width, height, and confidence
score that indicates how model confidence on whether that box contains an object and how accurate
the predicted box is. Then, the model computes the intersection over union (IoU) between predicted
and ground-truth bounding boxes to eliminate those which cannot pass the threshold value. Since the
algorithm may predict multiple bounding boxes that pass the threshold for the same object, non-
maximum suppression (NMS) is applied to identify redundant boxes and output one box for each
object in the image. YOLO OD series are often used in real-time use cases covering on the edge
inference.
YoloV6
YOLOv6 (Li et al., 2022) is a cutting-edge object detector that provides a balance between speed
and accuracy. It introduces notable enhancements to its architecture including a reparameterized
backbone, Path Aggregation Network (PAN) (Liu et al., 2018), and efficient decoupled head for
prediction.
The reparameterized network enhances the detection quality and speeds up the inference phase.
The network architecture is switched during training and inference to balance speed and accuracy. It
uses a simple network during inference for efficiency whereas a complex one is preferable in training
to provide higher accuracy. Path aggregation network concatenates features from different
reparameterized blocks hence it is called as RepPan. Compared to the previous Yolo version, Yolov6
uses efficient decoupled heads to split classification and detection paths. This approach reduces the
computational complexity and provides higher accuracy.

Alpay TEKİN & Ahmet Selman BOZKIR
14(1), 8-17, 2024
Enhance or Leave It: An Investigation of the Image Enhancement in Small Object Detection in Aerial Images
11
YoloV7
YOLOv7 (Wang et al., 2023) proposes many novelties in its architecture to increase efficiency
such as extended Efficient Layer Aggregation Network (E-ELAN) (Wang et al., 2022) and Compound
Model Scaling. The E-ELAN is employed as the computational block for the YOLOv7 backbone
architecture. It uses expand, shuffle, and merge cardinality to continuously improve the network's
capacity for learning while preserving the original gradient path. Scaling helps the model to comply
with the needs of objective tasks in terms of speed and accuracy. The authors of YOLOv7 optimized
the network architecture search technique (NAS) and proposed a compound model scaling approach
that scales the width and height in coherence for concatenation-based models.
Furthermore, YOLOv7 contains two trainable bags of freebies named planned re-parameterized
convolution (RepConvN) and Coarse for auxiliary and fine for lead loss (CAFL). RepConv combines
3x3 convolution, 1x1 convolution, and identity connection in one convolution layer. The RepConvN is
RepConv without an identity connection. This technique increases the training time yet provides
higher accuracy in prediction. By utilizing the CAFL approach, Yolov7 overcomes the limitation of a
single head. It contains a lead head responsible for predicting output, and an auxiliary head that assists
training in the middle layers.
YoloV8
YOLOv8, as the recent version, improves over previous ones in terms of speed and accuracy. It
enriches the detection quality by utilizing self-attention, feature pyramids, and mosaic augmentation.
The enhanced CSPDarkNet53 builds the backbone of the YOLOv8 containing 53 convolutional
layers and leverages cross-state partial connections to provide communication between different
layers. The head consists of convolutional layers followed by a series of fully-connected layers. It
predicts bounding boxes, confidence scores, and class probabilities for the detected objects. The self-
attention forces the model to focus on different features based on their relevance to the task. It is
noteworthy that the feature pyramid network allows the model to perform multi-scale object detection.
It contains multiple layers that can detect objects at different scales. Another key enhancement in
YOLOv8 is the mosaic augmentation which helps the object detection models to learn how to detect
objects in cluttered or complex scenes yielding better generalization in the wild. By utilizing mosaic
augmentation, the model is exposed to a wider variety of visual contexts and can learn to recognize
objects in a more robust and generalizable way. It should be noted that YOLOv8 is an anchor free
approach reducing the number of bounding box predictions and boosts the NMS.
Multi stage progressive image restoration (MPRNet)
MPRNet (Zamir et al., 2021; Rajaei et al., 2023) is a multi-stage model for image restoration.
Since it is multi-stage, it breaks down the restoration process of the degraded image into sub-tasks and
progressively learns the restoration function and restores the degraded image followingly. The model
first learns the contextualized information using an encoder-decoder network and then combines them
with a high-resolution subnetwork that retains the local information. At each stage, the supervised
attention module re-weights the local features and exchanges information between different stages. In
order to avoid loss of information, cross-state feature fusion is leveraged at each state to establish the
connection between feature processing blocks.
At any given state , the model predicts the residual image  and adds the degraded input image
 to obtain the predicted restored image:    . The model is optimized end-to-end with
following loss function:
  󰇟 󰇛󰇜 󰇛 󰇜󰇠

 (1)

Alpay TEKİN & Ahmet Selman BOZKIR
14(1), 8-17, 2024
Enhance or Leave It: An Investigation of the Image Enhancement in Small Object Detection in Aerial Images
12
where  represents the ground-truth image, and  is the Charbonnier loss:
    (2)
with constant  empirically set to . In addition,  is the edge loss defined as:
 󰇛󰇜  󰇛󰇜 (3)
where  denotes the Laplacian operator. The parameter  controls the importance of the two loss
term.
The key parts of the MPRNet are encoder-decoder sub-network, original resolution sub-network,
cross-state feature fusion, and supervised attention module. The encoder-decoder architecture allows
the model to focus on more relevant features at each stage. The cross-state feature fusion module is
employed to make the model less vulnerable the information loss due to the encoder-decoder
subnetwork. The supervised attention module generates the feature map that filters less informative
features and only allows useful ones to propagate to the next stage. Finally, the original resolution
network is employed at the last stage to generate spatially-encriched, high-resolution images.
Datasets
We conducted the experiments on VisDrone dataset (Cao et al., 2021). This dataset covering 10
object classes, contains 6471 images for training and 1580 images for testing. We built separate
“noisy”, “blurry” and “rainy” test sets derived from the original VisDrone test set by adding synthetic
noise, motion blur, and rain effects. In the following, we applied the pre-trained MPRNet on degraded
test sets to obtain their restored versions called “noisy-clear”, “blur-clear”, and “rain-clear”. Fig. 1
represents our pipeline for generating the degraded and their restored test sets. In the stage of image
restoration, though our experimentation setup involves one 16 GB VRAM equipped 3080TI GPU, we
have experienced insufficient memory issues which forces us to decrease the resolution of the original
images in test sets in a way to have 640x640 pixels. The summaries of the datasets used in this
experiment are listed in Table 1.
Table 1. Summary of the Original And Derived Datasets
Dataset
Size
Description
Original Train set
6471 images
Novel VisDrone train set
Original Test set
1580 images
Novel VisDrone test set
Noisy test set
1580 images
Derived from VisDrone test set by adding noise
Blurry test set
1580 images
Derived from VisDrone test set by adding motion blur
Rainy test set
1580 images
Derived from VisDrone test set by adding synthetic rain
Noise-clear test set
1580 images
Derived from Noisy test set by performing denoising (MPRNet)
Blur-clear test set
1580 images
Derived from Blurry test set by performing deblurring (MPRNet)
Rain-clear test set
1580 images
Derived from Rainy test set by performing deraining (MPRNet)
Figure 1. Test Set Generation Pipeline

Alpay TEKİN & Ahmet Selman BOZKIR
14(1), 8-17, 2024
Enhance or Leave It: An Investigation of the Image Enhancement in Small Object Detection in Aerial Images
13
Evaluation metrics
In this experiment, the metrics of precision, recall, f1-score and mean average precision (mAP)
were used to evaluate the accuracy of YOLO models we utilized. The following equations shows how
the performance metrics are computed.
  
 (4)
  
 (5)
  
  (6)
  
 (7)
   


 (8)
where    󰇛󰇜 󰇛  󰇜󰇛󰇜

 and , and  denote the number of
classes and treshold respectively.
RESULTS AND DISCUSSION
We fine-tuned tiny, small, and medium variations of each YOLO version for 50 epochs using an
NVIDIA RTX 3060 GPU equipped with 6 GB VRAM. We set the batch size and learning rate to 2 and
0.01 respectively. We evaluated each model by using test sets listed in Table 1 and reported precision,
recall, f1, and mAP scores. We used degraded and restored test sets to investigate how image
restoration improves the accuracy of the detection model in case the input image is degraded. The
Table 2, 3, and 4 report the experiment results for YOLO v6, v7, and v8 respectively.
Table 2. Yolov6 Test Results
Dataset
Model
Precision
Recall
F1 Score
mAP
Original test set
Nano
Small
Medium
0.281
0.382
0.449
0.23
0.304
0.36
0.226
0.311
0.378
0.18
0.151
0.322
Noisy test set
Nano
Small
Medium
0.113
0.146
0.178
0.088
0.119
0.127
0.091
0.118
0.132
0.038
0.059
0.068
Blurry test set
Nano
Small
Medium
0.127
0.127
0.144
0.097
0.109
0.093
0.1
0.112
0.104
0.046
0.052
0.050
Rainy test set
Nano
Small
Medium
0.169
0.246
0.295
0.147
0.202
0.221
0.141
0.204
0.232
0.084
0.138
0.165
Noise-clear test set
Nano
Small
Medium
0.14
0.157
0.178
0.101
0.118
0.137
0.111
0.127
0.146
0.056
0.066
0.081
Blur-clear test set
Nano
Small
Medium
0.238
0.256
0.289
0.178
0.224
0.216
0.18
0.22
0.233
0.123
0.158
0.165
Rain-clear test set
Nano
Small
Medium
0.167
0.245
0.364
0.146
0.194
0.253
0.141
0.2
0.277
0.082
0.133
0.213
The experiments show that the detection performance of YOLO models is significantly reduced
with degraded images. Performing the MPRNet-based image enhancement on degraded images yields
promising improvement in object detection performance in line with our hypothesis. For the YOLOv6-
tiny model, the denoising operation results in a %47 improvement in the mAP score. It also increases
the mAP scores for YOLOv7-small and YOLOv8-medium models by 116%, and 60% respectively.

Alpay TEKİN & Ahmet Selman BOZKIR
14(1), 8-17, 2024
Enhance or Leave It: An Investigation of the Image Enhancement in Small Object Detection in Aerial Images
14
Table 3. YOLOv7 Test Results
Dataset
Model
Precision
Recall
F1 Score
mAP
Original test set
Nano
Small
Medium
0.358
0.526
0.511
0.292
0.416
0.441
0.322
0.465
0.473
0.248
0.397
0.406
Noisy test set
Nano
Small
Medium
0.133
0.139
0.154
0.071
0.075
0.069
0.092
0.097
0.095
0.025
0.030
0.032
Blurry test set
Nano
Small
Medium
0.102
0.127
0.124
0.093
0.085
0.093
0.097
0.102
0.106
0.044
0.048
0.045
Rainy test set
Nano
Small
Medium
0.181
0.322
0.315
0.188
0.248
0.249
0.184
0.28
0.278
0.107
0.187
0.185
Noise-clear test set
Nano
Small
Medium
0.137
0.182
0.165
0.097
0.108
0.112
0.114
0.136
0.133
0.05
0.065
0.062
Blur-clear test set
Nano
Small
Medium
0.215
0.317
0.325
0.214
0.249
0.242
0.214
0.279
0.277
0.138
0.187
0.188
Rain-clear test set
Nano
Small
Medium
0.191
0.302
0.296
0.176
0.247
0251
0.183
0.272
0.272
0.106
0.18
0.178
Similarly, leveraging the deblurring process has shown a better performance boost than we
expected on the blurred test set. It improves the mAP scores %230, 289%, and 263% for YOLOv6-
medium, YOLOv7-medium, and YOLOv8-tiny respectively. Although MPRNet-based image
enhancement improves the detection results for noisy and blurry images, it cannot maintain the same
performance on rainy images due to the algorithm we used to add synthetic rain into images. The
algorithm unfortunately generated poor-quality, non-realistic synthetic rain hence MPRNet cannot
recognize and remove it well from the image.
Table 4. YOLOv8 Test Results
Dataset
Model
Precision
Recall
F1 Score
mAP
Original test set
Tiny
Small
Medium
0.388
0.453
0489
0.291
0.342
0.37
0.332
0.389
0.421
0.266
0.326
0.359
Noisy test set
Tiny
Small
Medium
0.122
0.156
0.196
0.055
0.058
0.066
0.075
0.084
0.099
0.042
0.047
0.061
Blurry test set
Tiny
Small
Medium
0.122
0.139
0.184
0.054
0.058
0.061
0.075
0.081
0.091
0.041
0.049
0.056
Rainy test set
Tiny
Small
Medium
0.215
0.248
0.278
0.16
0.203
0.219
0.183
0.223
0.245
0.113
0.15
0.171
Noise-clear test set
Tiny
Small
Medium
0.172
0.206
0.231
0.073
0.075
0.087
0.102
0.11
0.126
0.060
0.075
0.084
Blur-clear test set
Tiny
Small
Medium
0.263
0.298
0.323
0.183
0.203
0.218
0.216
0.241
0.26
0.149
0.176
0.192
Rain-clear test set
Tiny
Small
Medium
0.217
0.251
0.277
0.16
0.198
0.21
0.184
0.221
0.239
0.116
0.15
0..166
This is likely related to the irrelevant distribution between the real-world raindrops and our
synthetic rain effect. Furthermore, one might question why the restoration process could not catch the

Alpay TEKİN & Ahmet Selman BOZKIR
14(1), 8-17, 2024
Enhance or Leave It: An Investigation of the Image Enhancement in Small Object Detection in Aerial Images
15
mAP performance that we gained from the novel test set. The unwanted image size reduction caused
significant information loss, and reduces the model performance, especially for small objects. Our
comprehensive visual inspection clearly revealed that extremely small objects (e.g., <10x10 pixels) in
the original test set became impossible to detect in degraded and restored image sets upon the image
resize phase.
Further, as expected, we also observed that the newer YOLO models performed better compared
to their previous versions regardless of the applied image degradation. It should be noted that except
for the YOLOv8 models, the model sizes (i.e., tiny/small) affected the OD performance when it comes
to de-rained images. Heavier the model we applied, the better the mAP values we gained. The
performance gain obtained with the use of YOLO8 models sources from several advancements such as
(1) a decoupled head performing objectness, classification and regression tasks individually, (2) the
introduction of anchor-free OD paradigm, and (3) improved mosaic based data augmentation
techniques which incorporate MixUp and CutMix (Terven, Córdova-Esparza and Romero-González.,
2023). Moreover, as Wang et al. (2023) pointed out, to reveal multi-scale feature maps, input images
are processed through several convolution and C2f modules in YOLO8 saving the lightweight
characteristics along with capturing more abundant gradient flow. The C2f module is mainly used for
residual learning and is reported to be an enhanced version of ELAN structure presented by YOLO7
(Wang et al., 2023). Another key contribution of YOLO8 is the merging of Feature Pyramid Network
(FPN) and Pyramid Attention Network (PAN) paradigms which enables blending high and low level
features through semantic and localization related features making the model better utilizing varying
scaled features resulting in improved detection performance when small and large objects come into
prominence. From the perspective of listed improvements and our problem domain, it is not a surprise
to obtain surpassing scores with the use of YOLO8 in our experiments since most of the objects in our
dataset are significantly small.
Overall, our results prove that image enhancement significantly improves the detection quality of
the YOLO models on degraded images. It also improves small object detection on aerial images. Due
to space constraints, we could not share any run-time analysis of MPRNet.
CONCLUSION
In this work, we hypothesized that small object detection tasks, especially in aerial images, may
be improved by employing image restoration networks when it comes to degraded images. To achieve
this, we applied several YOLO (6/7/8) models on the (a) original, (b) synthetically degraded, and (c)
restored versions of the test portion of the VisDrone dataset. Our evaluation using both clean and
degraded sets demonstrate how degraded images reduce the detection quality. To eliminate these
artifacts from the image, we performed MPRNet-based image enhancement on the degraded test set
and evaluate the YOLO models on these restored test sets. The results show that image enhancement
significantly improves the detection quality, regardless of the underlying YOLO model, especially for
small objects on degraded images. In future work, we aim to couple more image enhancement
approaches with other OD models and analyze the run-time performance of these models in large and
small image format regimes.
Conflict of Interest
The article authors declare that there is no conflict of interest between them.
Author’s Contributions
The authors declare that they have contributed equally to the article.

Alpay TEKİN & Ahmet Selman BOZKIR
14(1), 8-17, 2024
Enhance or Leave It: An Investigation of the Image Enhancement in Small Object Detection in Aerial Images
16
REFERENCES
Cao, Y., He, Z., Wang, L., Wang, W., Yuan, Y., Zhang, D., & Liu, M. (2021). VisDrone-DET2021:
The vision meets drone object detection challenge results. In Proceedings of the IEEE/CVF
International conference on computer vision (pp. 2847-2854).
Dai, J., Li, Y., He, K., & Sun, J. (2016). R-fcn: Object detection via region-based fully convolutional
networks. Advances in neural information processing systems, 29.
Duan, K., Bai, S., Xie, L., Qi, H., Huang, Q., & Tian, Q. (2019). Centernet: Keypoint triplets for object
detection. In Proceedings of the IEEE/CVF international conference on computer vision (pp.
6569-6578).
Girshick, R., Donahue, J., Darrell, T., & Malik, J. (2014). Rich feature hierarchies for accurate object
detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision
and pattern recognition (pp. 580-587).
Girshick, R. (2015). Fast r-cnn. In Proceedings of the IEEE international conference on computer
vision (pp. 1440-1448).
Fu, C. Y., Liu, W., Ranga, A., Tyagi, A., & Berg, A. C. (2017). Dssd: Deconvolutional single shot
detector. arXiv preprint arXiv:1701.06659.
He, K., Zhang, X., Ren, S., & Sun, J. (2015). Spatial pyramid pooling in deep convolutional networks
for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 37(9),
1904-1916.
He, K., Gkioxari, G., Dollár, P., & Girshick, R. (2017). Mask r-cnn. In Proceedings of the IEEE
international conference on computer vision (pp. 2961-2969).
Law, H., & Deng, J. (2018). Cornernet: Detecting objects as paired keypoints. In Proceedings of the
European conference on computer vision (ECCV) (pp. 734-750).
Li, B., Peng, X., Wang, Z., Xu, J., & Feng, D. (2017). Aod-net: All-in-one dehazing network.
In Proceedings of the IEEE international conference on computer vision (pp. 4770-4778).
Li, C., Li, L., Jiang, H., Weng, K., Geng, Y., Li, L., & Wei, X. (2022). YOLOv6: A single-stage
object detection framework for industrial applications. arXiv preprint arXiv:2209.02976.
Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C. Y., & Berg, A. C. (2016). Ssd: Single
shot multibox detector. In Computer Vision–ECCV 2016: 14th European Conference,
Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14 (pp. 21-37).
Springer International Publishing.
Liu, S., Qi, L., Qin, H., Shi, J., & Jia, J. (2018). Path aggregation network for instance segmentation.
In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 8759-
8768).
Rajaei, B., Rajaei, S., & Damavandi, H. (2023). An Analysis of Multi-stage Progressive Image
Restoration Network (MPRNet). Image Processing On Line, 13, 140-152.
Redmon, J., & Farhadi, A. (2017). YOLO9000: better, faster, stronger. In Proceedings of the IEEE
conference on computer vision and pattern recognition (pp. 7263-7271).
Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster r-cnn: Towards real-time object detection with
region proposal networks. Advances in neural information processing systems, 28.
Terven, J., Córdova-Esparza, D. M., & Romero-González, J. A. (2023). A Comprehensive Review of
YOLO Architectures in Computer Vision: From YOLOv1 to YOLOv8 and YOLO-
NAS. Machine Learning and Knowledge Extraction, 5(4) (pp.1680-1716).

Alpay TEKİN & Ahmet Selman BOZKIR
14(1), 8-17, 2024
Enhance or Leave It: An Investigation of the Image Enhancement in Small Object Detection in Aerial Images
17
Uijlings, J. R., Van De Sande, K. E., Gevers, T., & Smeulders, A. W. (2013). Selective search for
object recognition. International journal of computer vision, 104, 154-171.
Wang, C. Y., Bochkovskiy, A., & Liao, H. Y. M. (2023). YOLOv7: Trainable bag-of-freebies sets new
state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition (pp. 7464-7475).
Wang, X., Gao, H., Jia, Z., & Li, Z. (2023). BL-YOLOv8: An Improved Road Defect Detection Model
Based on YOLOv8. Sensors, 23(20), 8361.
Wang, C. Y., Liao, H. Y. M., & Yeh, I. H. (2022). Designing Network Design Strategies Through
Gradient Path Analysis. arXiv preprint arXiv:2211.04800.
Zamir, S. W., Arora, A., Khan, S., Hayat, M., Khan, F. S., Yang, M. H., & Shao, L. (2021). Multi-
stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer
vision and pattern recognition (pp. 14821-14831).