Image Recognition - Science topic

A big yes, SVMs or any other classical non deep architecture are still used, you should use them when you are confident that the extracted features are of good significance. You can also use a CNN as a feature generator for a downstream SVM based classifier.

Muhammad Ali Thank you.

@Knoll, I did it for my research paper. It was about mathematical formulas and equations on panel data econometric model. To open a pdf with MS word, you need to, at first, create a blank word file, then go to the option button on the left corner of the screen, there you will see the option "Open" along with options such as "Save". " Save as". Then click on "Open", then, select the specific pdf file, click on it, and it will be opened on MS word.

Our team were working hard in order to share with the scientist community our article about SmartGloves. I am really proud to share with all of you this systematic review. Here you are.

Please ask whatever you want and send us comments in order to be aware if it will be usefull or not. The number possibilities nowadays are increasing. The first step is to know what are doing and in what real applications. Enjoy!!

I think the question is not really clear to me. First, you say that you are working on a project that aims to classify the quality of coffee beams into two categories using images, right? Then, that is a binary classification task and it is supervised learning because you have the labels (i.e. healthy or defected) and you have the images (the inputs). Something like a CNN would process the image and output a single value between [0, 1], which is compared to the target (label).

As a reference, I think this paper (https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8090980) is doing what you want.

Some of the algorithms used in image recognition (Object Recognition, Face Recognition) are SIFT (Scale-invariant Feature Transform), SURF (Speeded Up Robust Features), PCA (Principal Component Analysis), and LDA (Linear Discriminant Analysis).

Regards

This link is important for your question:

You can take a look in this link:

Hello Can you explain this result? What are your criteria for this failure? Please explain further

Sowmith Reddy

Follow

https://www.mathworks.com/help/deeplearning/pattern-recognition-and-classification.html

Hi Yaozhi,

one requirement for causality is the existence of a temporal sequence because in order to be recognizable, the cause has to appear before the effect.

So, during the recognition of single static images there can be no causality involved. You could argue that there is still room for reasoning like "Because a tree is in front of it there are not really two halves of a horse but it's a whole horse, only partially visible." But first, I think current ANNs do not have enough capacity for reasoning, and second, I would not call the relationship between elements in an image "causality".

There are, of course, ANNs for predicting the development of temporal series, e. g. of stock prices; here is certainly causality involved, though partially irrational. But those ANNs do just pattern recognition without "knowing" or regarding any causes.

Speaking pictorially, I would compare today's ANNs to students who are good willing and industrious but do not really understand anything. They have a chance of finishing (officially) successfully by just collecting clues on what to do. ("In tasks which look similar to those we had in autumn, I have to add all resistances.")

CNN can be performed well while dataset is big because CNN required large dataset to train the syatem. SVM able to train system by small dataset so, for the stating point, you can chose classifier according to the dataset.

Second, SVM required separate feature extraction and suitable feature extraction method help to reach maximum accuracy. So, have a look n various feature extraction methods.

Dear Wasseem Al-Obaydy Wasseem Al-Obaydy,

Look the link, maybe useful.

Regards, Shafagat

Now, matlab offers a function named fitcecoc() which is designed specially for multiclass SVM by error correction method. You can apply it. for details please go to the following link: https://www.mathworks.com/help/stats/fitcecoc.html

Thanks

Madam Murthy is right

Without doubt: Deep Convolutional Neural Networks.

For Matlab Code, Visit link

Hello,

I have been working on a same idea for a long time.You can send me an email directly to talk and discuss about this problem.I have solution but it is more expensive and hard to find the correct worker in my country.

Thanks

You should probably try to find a problem you can solve combining/pipelining both types of algorithms, i.e. pretend you had an ultimate goal when you worked on them both. I don't know exactly which algorithms you are talking about, but lets say you are detecting people approaching a government building and you want to identify a felon among them or something like this. Depends on your specific work, really..

Maybe the "parametric sparse representation" can help you.

same thing!

you have to take some images which have circle shapes in it. similarly you do it for all shapes. you can take the features of these images. for example HoG features. give the feature vector to the classifier in the training phase. tool for Knn is available.

My previous post focused on computer stereo-vision. But in a single image feature detector context, I suggest that you check for the following paper by Meer et al., "Edge Detection with Embedded Confidence", 2001 ( http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.21.2946&rep=rep1&type=pdf ).

Miller et al. introduced a confidence measure of interest that they integrated into gradient based edge detectors.

you said, "I get empty matrix for some windows".

That can happen for blank patches (ie, filled with the same color), then image gradients will be zero.

In the following paper by ZaritskyEmail et al., "Benchmark for multi-cellular segmentation of bright field microscopy images" BMC Bioinformatics201314:319 DOI: 10.1186/1471-2105-14-319, they talk about a dataset of 171 manually segmented images of 5 different cell lines at diverse confluence levels, acquired in several laboratories under different imaging conditions in the first figure.  HTH.

First you need to extract your object with SfM or optical flow ... 

Then if you want you want scale invariant descriptor you can use Sift/Surf or Hog you can also have a look to bag of features or even deep learning (by transfer learning) then I would recommend SvM or AdaBoost to make your classification ... But the right answer will depend on your training dataset...  

Six months

As mentioned earlier, the TPS software suite very useful and user-friendly in placing landmarks and performing basic analyses. http://life.bio.sunysb.edu/morph/

My former adviser has also worked on computer-based species identification. This may be of interest to you.

Best of luck with your work

Hello Aditya,

What are the Objects you want to recognise ? Are they same object? and more in Number?

Example : 1st Case : You want to Detect Multiple objects of same class .

You want to detect Multiple Oranges in a Picture.

2nd Case : Multi Class Object detection

You want to detect Oranges, Apples, Banana's . 

If you are looking for the first case, it should not be that hard to implement, Apply the object detector multiple times on the images, instead returning when you find the first object.

If its the second case, try building feature detectors for every class and apply every detector onto the image to find the different objects.

In general there are many approaches to recognise the object. Simple features like Color, Shape, or more complex features using HaarWavelets.

Start with one object and move to multiple object recognition

chi-square distance is useful when you want to compare two histograms. Eg for LBP (Local Binary Patterns) we can compare two histograms calculated from image texture (more precise: measure distance between two histograms). SO, if you have some histograms (whatever form LBP or others image features) from different places of face then probably can you compare this with others histogram calculated from another face. That could be a some base for made a simply classifier.

You may want to try the 'scikit-learn' package for that purpose. 

You can try also Audiveris open source sw

Homepage link:   https://audiveris.kenai.com/

Reagards,

Marco

Hey Petr,

as a very basic approach, given your images look like that most of the time, you could use 2D-correlation (like in older OCR systems). Build a database with example letters, correlate them and pick the ones with the highest correlation coefficient. You might have to solve some ambiguous cases (like 6 & 8 for example) afterwards.

A problem with this method is, that it is rather scale-dependent...

Greetings, David

Regarding the rear view, this is also a problem in VIS imaging: you will not be able to discern a horse from a donkey - a lot of animals look the same when you can only get a glimpse on their rear.

To some extend proportion relations might help: length of legs vs. total height vs. width of the a... and so on. The results will not be too satisfactory, but given the amount of data (that is: not much) this is the best you can do.

Wish you success

Could you show na image as na example?

Hi,

I am trying to shed some light here:

1. what you get from SIFT is an 128-dim. descriptor for each detected keypoint position, orientation and scale.

2. In the matching step, only the 128-dim descriptor is involved. Please note that the descriptor is already normalized to orientation and scale. The detected position gives you the center point of the image patch that you describe with the 128 dim. descriptor. The scale tells you how large this patch is and the orientation tells how much the patch has to be rotated before descriptor computation.

3. This depends on the application... if you just want to match points between images you compare the 128.dim descriptors and just look at the keypoint position, when corresponding descriptors have been detected.

4. see 2. and 3.

5. Yes, if you rotate an image by 90 degree you should get the same number of keypoints.

6. This is optional... if you implement the standard matching described in Lowe's paper then you have a one-to-many matching.

Hi Niranjan,

please go through the following papers, which gives lot of insights on the topic

Hi Laura

You may treat this problem as an ordinal variable. If in your example, the student marks a and b, he gets the maximum score (eg. 3) . If he marks c he obtain the minimum score (eg. 0), if he marks only a or b he obtains a score in the midle (eg. 2). If he marks b and c, or a and c, he gets a regular score (eg. 1). Then you may use an IRT method for an ordinal response or for a mix of binary and ordinal responses. Also it may be that to mark only a was better than to mark only b, then you may give different ordinal value to this selections. For example you may put a value +2 for a, +1 for b, -2 for c and your ordinal variable may be the sum of this values. You should also take account of the lack of response.

The chi square kernel do some normalization in the kernel itself, so it's often better than gaussian kernel.

It's not directly related to "visual recognition", but in 2002, face recognition researchers observed the following:

"One open question in face recognition is: How does database and watch list size effect performance? Because of the large number of people and images in the FRVT 2002 data set, FRVT 2002 reported the first large-scale results on this question. For the best system, the top-rank identification rate was 85% on a database of 800 people, 83% on a database of 1600, and 73% on a database of 37437. For every doubling of database size, performance decreases by two to three overall percentage points. More generally, identification performance decreases linearly in the logarithm of the database size."

Handbook of Face Recognition (2nd ed., Editors: Li & Jain). Chapter 21, page 569, Evaluation Methods in Face Recognition by P. Jonathon Phillips, Patrick Grother, and Ross Micheals

However, I'm not aware of any other supportive -and systematic- findings from other researchers about their observation (linear performance decrease in the log of database -class- size). 

It depends on the structure of descriptors that you use, where each distance is suitable between a special descriptors with respecting to the space where you apply your comparison. 

If I'm reading your question right, you're interested in what color different signs appear when illuminated using illuminants which have vastly different colors; the example is that a red stop sign will appear dark when illuminated using a source which is predominantly blue.

The traditional way to transform between different illuminations is to use the von Kries transform. It models the way human eyes respond to different illuminations. It's not perfect (human vision is complicated, and color spaces usually use three values to approximate the spectrum!), but it will get you close. [Hint: when you're transforming the colors, remember to convert your pixel values from sRGB to linear RGB first.]

Bruce Lindbloom (linked below) has great information about the von Kries transform and color adaptation when viewing scenes under different illuminations.

Easymacrosrecorder allow to create automatic pattern of mouse action (f.e. moving cursor,clicking etc.), so you can use it without any special knowledges to create "recording pattern" ;)

Dear Muhammad Ihsan Zul,

Look my article. May be useful.

Regards, Shafagat

try searching for:

Quantifying plant colour and colour difference as perceived by humans using digital images (Kendal et al. 2013)

This problem has been addressed, at least with face recognition problem, and there are many approaches to choose from. Depending on the further processing some approach may be better for you than other, but I would recommend to see, for example, ECCV 2014 paper from Amara et al. "On the Effects of Illumination Normalization with LBP-Based Watchlist Screening". There you have eleven different approaches tested on faces.

Personally I have used only Retinex style methods and some basic filtering (but not on faces)...

Generally, S-Transform provide a better resolution in both time and frequency domain. It is possible to demonstrate that Wavelet Transform is a special case of the S-Transform. Moreover, using S-Transform you can directly evaluate the local spectrum of the signal without using any empirical relationship between the scale factor and frequency.

Deformable Part Models are top performers at the Pascal VOC challenge, but they are quite slow (few secs per image on a strong machine). Usually they use HOG + SVM on a multi-scale search-space.

Papers+code: http://www.cs.berkeley.edu/~rbg/latent/

Another implementation with pre-trains models:

Aggregate Channel Features on the other hand are quite fast (30+ fps on a single core, depending on the implementation), since they just estimate the features on nearby scales:

Paper: http://research.microsoft.com/pubs/220572/DollarPAMI14pyramids.pdf

Code: http://vision.ucsd.edu/~pdollar/toolbox/doc/index.html

@G. Chliveros : source code is not available on the website

If u r going to detect lights of car, truck etc. in which two lights are on. Then apply light thresholding (light is near 255 gray intensity). Now detect blobs. If two blobs are near to each other  then lights are true vehicle lights.

The only other factor you need is the height of the object in real life (otherwise you could be photographing a model which is much closer to the camera).

The maths isn't actually that complex, the ratio of the size of the object on the sensor and the size of the object in real life is the same as the ratio between the focal length and distance to the object.

To work out the size of the object on the sensor, work out it's height in pixels, divide by the image height in pixels and multiply by the physical height of the sensor.

So the whole sum is:

distance to object (mm) = focal length (mm) * real height of the object (mm) * image height (pixels)

---------------------------------------------------------------------------

object height (pixels) * sensor height (mm)

object height (pixels) * sensor height (mm)

Let's sanity check this equation.

If we keep everything else constant and increase the focal length then the distance increases (as focal length is on the numerator). This is what you would expect, if you have to zoom your lens to make one object the size another equally sized object used to be, the first object must be further away.

If we keep everything else constant and increase the real height of the object then again the distance increases as if two objects of different real heights appear the same height in the image the taller one must be further away.

If we keep everything else constant and increase the image height, then the distance increases, as if two objects (of the same size, remember we're keeping everything else constant) appear the same pixel size in a cropped and uncropped image then the object in the uncropped image must be further away.

If we keep everything else constant and increase the object height in pixels then the distance decreases (we're on the denominator now): two equally sized objects, one takes up more pixels, it must be closer.

Finally if we keep everything else constant and increase sensor size, then distance decreases: two equally sized objects have the same height in pixels when shot with a compact (small sensor, where 20mm is a long lens) and shot with a DSLR (large sensor where 20mm is a wide lens), then the object in the DSLR image must be further away (because it appeared the same size but with a wide lens).

I saw that from the website and hope it could help you.

You can use the following  site:

Then create the following datasets

Data Set Information:

This database has been artificially generated by using a first order theory which describes the structure of ten capital letters of the English alphabet and a random choice theorem prover which accounts for etherogeneity in the instances. The capital letters represented are the following: A, C, D, E, F, G, H, L, P, R. Each instance is structured and is described by a set of segments (lines) which resemble the way an automatic program would segment an image. Each instance is stored in a separate file whose format is the following:

CLASS OBJNUM TYPE XX1 YY1 XX2 YY2 SIZE DIAG

where CLASS is an integer number indicating the class as described below, OBJNUM is an integer identifier of a segment (starting from 0) in the instance and the remaining columns represent attribute values. For further details, contact the author.

Attribute Information:

TYPE: the first attribute describes the type of segment and is always set to the string "line". Its C language type is char.

XX1,YY1,XX2,YY2: these attributes contain the initial and final coordinates of a segment in a cartesian plane. Their C language type is int.

SIZE: this is the length of a segment computed by using the geometric distance between two points A(X1,Y1) and B(X2,Y2). Its C language type is float.

DIAG: this is the length of the diagonal of the smallest rectangle which includes the picture of the character. The value of this attribute is the same in each object. Its C language type is float.

Good Luck

I am one of the authors of the paper that you are citing. I am going to try to help you.

First of all, the concept of overlapping block can be better understood in the original paper where HOG is described by Dalal and Triggs. With the aim of calculating the HOG descriptor, several local gradients are binned in accordance with orientation, weighted depending on their magnitude, within a spatial grids of cells with overlapping blockwise contrast normalization. Within each overlapping block of cells, a feature vector is obtained by sampling the histograms from the contributing spatial cells. The feature vectors for all overlapping blocks are concatenated to produce the final feature vector which is fed to the classifier.

Finally, the concept of binary classification is related to problems where there are several classes that are classified by means of different binary classifiers. You can understand this concept by studying the LIBSVM libraries.

I hope that my response helps you. Good luck with your research!

Hi Naveen,

as far as I know the extension of GLCM (Grey-level co-occurrence matrices) to colour images is usually referred to as integrative co-occurrence matrices. The method was originally proposed by Palm [1]. A very good description of the approach is also available in [2].

[1] Palm, C. Color texture classification by integrative Co-occurrence matrices

(2004) Pattern Recognition, 37 (5), pp. 965-976.

[2] Arvis, V. Debain, C. Berducat, M., Benassi A. Generalization of the cooccurrence matrix for colour images: application to colour texture classification (2004) Image Analysis & Stereology, 23(1), pp. 63-72

Vincent Arvis, Christophe Debain, Michel Berducat, Albert Benassi

Finally, let my mention a paper you might find useful:

F. Bianconi, R. Harvey, P. Southam and A. Fernández; Theoretical and experimental comparison of different approaches for color texture classification (2011) Journal of Electronic Imaging, 20(4), 043006

(MATLAB code available at http://dismac.dii.unipg.it/ctc/)

Action recognition with improved trajectories

H Wang, C Schmid

Computer Vision (ICCV), 2013 IEEE International Conference on, 3551-3558

Action recognition by dense trajectories

H Wang, A Klaser, C Schmid, CL Liu

Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on

You could have a look at RatSLAM, as it would fit your constraints very well (works indoor + outdoor, uses a camera as input). There is an open source version of it available, too: OpenRatSLAM.

I think the easiest way to solve this is to coregister the three kinect images by calibrating the setting to the same (e.g. real world) coordinate system. This means you have to transform the skeletons of each kinect to the real world coordinate system first. If this is done with sufficient accuracy the registration of the joints coming from the different sensors should be possible by a simple nearest neighbor search.

OpenCV library is in case. Please, get more information through the link:

What type of sprayer? Did you link that image recognition to the sprayer?

I think that hypothesis makes sense, but I also have not seen this paper extended, which I've always found surprising.

Ackerman has work showing angry faces induce better CR memory, due to the potential threat they pose.

Research also shows that sad mood increase attention to details and thus one might suspect sad moods to increase face memory, but this may only be for SR targets since CR targets may be perceived as offering fewer affiliative affordances.

thanks Mr. Evon

I think the previous two references provided by @Christos P Loizou and @Lucia Ballerini are new and excellent to start with.

Thank you all.In my dataset each image has a person with infected ,healthy skin and background ?

Thank you

Thanks for this. I'll check out the links.

Thanks Ahmed

Perhaps this would be a good place to start:-

In order to extract text regions I recomend you to read this paper about text detection on natural images: Epshtein, B.; Ofek, E.; Wexler, Y., "Detecting text in natural scenes with stroke width transform"

ccv is a computer vision library that has some great algoritms, and has an implementation of this method: http://libccv.org/doc/doc-swt/

I hope this will be helpful to you.

Gang Zhang , Tobias Senst Thank you very much, good luck.