Local and deep texture features for classification of natural and biomedical images
Developing efficient feature descriptors is very important in many computer vision applications including biomedical image analysis. In the past two decades and before the popularity of deep learning approaches in image classification, texture features proved to be very effective to capture the gradient variation in the image. Following the success of the Local Binary Pattern (LBP) descriptor, many variations of this descriptor were introduced to further improve the ability of obtaining good classification results. However, the problem of image classification gets more complicated when the number of images increases as well as the number of classes. In this case, more robust approaches must be used to address this problem. In this thesis, we address the problem of analyzing biomedical images by using a combination of local and deep features. First, we propose a novel descriptor that is based on the motif Peano scan concept called Joint Motif Labels (JML). After that, we combine the features extracted from the JML descriptor with two other descriptors called Rotation Invariant Co-occurrence among Local Binary Patterns (RIC-LBP) and Joint Adaptive Medina Binary Patterns (JAMBP). In addition, we construct another descriptor called Motif Patterns encoded by RIC-LBP and use it in our classification framework. We enrich the performance of our framework by combining these local descriptors with features extracted from a pre-trained deep network called VGG-19. Hence, the 4096 features of the Fully Connected 'fc7' layer are extracted and combined with the proposed local descriptors. Finally, we show that Random Forests (RF) classi er can be used to obtain superior performance in the eld of biomedical image analysis. Testing was performed on two standard biomedical datasets and another three standard texture datasets. Results show that our framework can beat state-of-the-art accuracy on the biomedical image analysis and the combination of local features produce promising results on the standard texture datasets.