I’m using CNN for learning the shape features from the mechanical part images (Like levers, screws)
I made a lot of images that has lever, spring, cam by using matlab.
and I successfully trained the CNN model that has 97% accuracy for test dataset.
But last convolutional layer has the 4 output(4 Feature Map) and 3 Feature Maps are zero (like null space)
Only 1 feature map has values that activated with images
Why do the feature maps that are only zero values happen?
How can I fix the problem so that has 4 Feature Maps that has some values??
Usually some feature maps will be blank in a deep CNN, and that’s perfectly normal. That’s because each layer learns the shape of its filters through training. Lower level filters detect whole objects, and higher(deeper) level filters detect specific shapes (noses, handles, etc.). This means that many higher level filters won’t be able to detect anything, and they will produce feature maps that are usually blank.
In your case, however, your final feature map has 4 channels, 3 of which are blank. Perhaps you are using far too big of a network for your dataset size? In any case, if the model goes up to 97% accuracy for the test/eval dataset, then the number of blank feature maps probably won’t matter too much.