Input for neural network

I have Road Sign images (only sign - without cars, tree etc.) and made with Image Batch Processor: binary image and add some noise, rotation etc. because want to have a lot of pattern. Than all of this images I reshape to one matrix and save into workspace. This is my input data. Matrix dimmensions are 17500x1113 where 17500 are piksels of image and 1113 samples. My answer is T=159x1113 because I have 159 type of Road Sign. My network give me answer but it has no sense. Should I use Image Segmentation ? Thanks !