Gray Scale Images

The previous article trained the network on RGB images and found that the network relies too heavily on colour to identify the numbers. We will therefore repeat the test with Gray scale images.

This network overfits the data because it correctly labels all training data but only ~40% of the test data.

A brief look at some of the mislabelled images confirms that the network does not “recognise” digits. It struggles even with well lit numbers that face the camera directly.

One possible explanation is the small training set. Unlike MNIST, data set comprises only 100 images for each label. Furthermore, each number is projected onto an randomly aligned cube in 3D space under different lighting conditions. A larger (or augmented) training set may therefore alleviate the problem. On the other hand, I have no difficulty to identify the images and feel the network should do better than 40%.