Dear Mohammad Ismaeil

I think your training is not done properly, if done the training with these changes it will give better identification. I have only one suggestion you just change the image into gray scale or something related to it so that the identification easier otherwise the identification should need so much effort regarding tyre make some marks on the White line so there are many problem comes with respect to it

Krishnan M S

Thanks for your suggestion.

At present i am not using any trained model.I am trying with simple image processing/computer vision segmentation algorithm.

Kindly suggest if you familiar with any algorithm for above limitations.

What are the camera conditions? Are they considered static?

Having said that it might be worth dividing objects into two categories:

1) those that can be identified based on your technique (template based, etc.)

2) those that are heavily distorted

I am assuming that you have gotten the first category right

Further :

1) Assuming the signs follow the international standard you can extend the template to include color and sub segment recognition

2) Use something like Adaptive resonance theory (clustering) or a supervised learning algorithm that can be trained to detect sub segments

Hello Arturo Geigel

Thanks for your suggestion.

Camera condition is not static.

My focus is on proper arrow segmentation not in identification.

For the clear images using "chan-vese" and contour based method of opencv i am properly segmenting. But for the distorted case i am searching any conditional base algorithm.

From the results that you have, the condition of the sign and assuming it is only asphalt, I would first concentrate on contrast enhancement and erosion dilation to clean up noise to exhaust all possibilities on the result that you already have.

Let me know how it goes and take it from there to explore other possibilities

Btw do the cleanup as preprocessing

Are you using OpenCV!!!

Arturo Geigel

Yes,cleanup is the first task i have done.

Krishnan M S

Yes i am using opencv and Skimage library.

Mohammad Ismaeil

How did it go with contrast enhancement?

I think that linear stretching can improve the quality of the sample image provided especially in the upper right quadrant which is the part that I would concentrate initially. Also I would break the image into segments and calculate the histogram per segment and apply the enhancement per segment. In the sample image provided you have at least three distinct regions where I would apply the stretching separately. This can also be automated fairly easy.

Let me know

Regards

Arturo Geigel

Achieved some good result by doing contrast stretching followed by breaking the image in four parts and then contour based segmentation.

Thank you so much for this suggestion.

Will update if i get better output using any other technique.

May be this paper help you Article Fast learning method for convolutional neural networks using...

You might try using 'alpha-shapes'. A kind of convex-hull approach (of the white remnants of the arrows), but allowing for large concavities. How big those concavities are allowed to be, is a parameter of the alpha-shape function, and would depend on the thickness of the cracks in the road-sign.

https://en.wikipedia.org/wiki/Alpha_shape

https://graphics.stanford.edu/courses/cs268-11-spring/handouts/AlphaShapes/as_fisher.pdf

Mohammed Abdallah Bakr Mahmoud

Lambert Zijp

Thanks for the suggestion. Will update if i get success.

Arturo Geigel

How to deal with illumination or high sun light affect? I tried with HSV and also done calculation by dividing the images, but dint get any satisfactory answer. Any suggestions.

Mohammad,

Sorry for the late response (you can email me directly if I do not respond).

Is it a saturation problem throughout the image or is it a distortion in shape due to artifacts made up of shadows?

Arturo Geigel

Its due to artifacts made up of shadow.

Mohammad,

When dealing with artifacts as a consequence of shadows there aer several invariants that you can use to test if the patch is a shadow or not(I would focus first on detection and then on removal). Since from the samples you have shown point to asphalt, it is a surface that has a lot of texture and this is an invariant. what happens is that the grey level histogram peak is shifted (have a shifted mean but the same sigma spread).

I would give several tries to measurements based on the grey level co-occurence matrix (GLCM) to see which one fits your particular case.

Let me know

Regards