How to implement an ANN-MLP classifier?
After all that theoretical explanation on how to implement an ANN, we will implement it ourself. For that, and as we did also in the SVM classifier, we will download the training images from the same source Caltech256, http://www.vision.caltech.edu/Image_Datasets/Caltech256. We will start with a few items, easily extendable to many other, creating a folder, images, with a subfolder for each of the categories that we will classify: dresses, footwear, and bagpack. We will take a bunch of images for each of them; around 20-25 images should be enough for the training, and on top of that we will include another set of sample images, which we will use for evaluating the accuracy of our network after the training.
As we discussed earlier, we need to align the number of descriptors for each of the images using a Bag of Words (BOW). For that, we will first extract the feature vectors for each of the images using dense feature detectors for the keypoints of...
