A Dynamic Multi Label Image Classification Based on Recurrent Neural Network
Real images usually have multi-labels, i.e., each image consists of multiple objects or attributes which belong to multiple classes, Compared to single-label images that focus on labels from one single class. Here we perform multi-label image classification on a dataset consisting of Chess pieces. These can be detected by taking into account features such as size, direction, and angle of a chess piece. Once these features are identified we label the image according to the type of chess pieces found, i.e., if the image consists of several pieces identified by the features, we classify them into multiple labels. To perform this we have planned to use feature extraction using recurrent neural networks, as this type of neural network has never been used to classify multi-label images of this type (Chess pieces) before, and the fact that RNN's are considerably better than other neural networks when it comes to exploiting label dependencies in a single image, i.e., In an image, the labels may have interdependencies, such that the presence of a certain label will affect the probability of another labels’ presence. Thus, exploiting the dependencies among these labels could be considered beneficial for the predictive performance of the classifier.