Image classification using bag of visual words and novel COSFIRE descriptors

Abstract

The main task of a keypoint descriptor is to describe an interesting patch (keypoint) in an image. This project proposes a new keypoint descriptor, based on the trainable COSFIRE filters that are used for keypoint detection and pattern recognition, to describe keypoints found in an image. A keypoint is a particular patch within an image that is deemed to be an interesting patch by a keypoint detector. A visual descriptor effectively describes the detected keypoints by being robust to changes in different image conditions while also being distinctive between different keypoints. We analyse the popular Bag of Visual Words (BOVW) image classification model, by examining each step of this model and choosing the best design configuration, starting from the extraction and description of the image keypoints to the classification of unseen image dataset. The proposed solution takes into consideration the configuration parameters found in the COSFIRE filters to effectively construct the novel keypoint descriptor. Different COSFIRE descriptor configurations were proposed in this project and their performance was assessed, along with other popular keypoint descriptors, on the popular procedure, where different image conditions, such as variation of viewpoint or blur, are taken into account to test the descriptor’s effectiveness. The best COSFIRE descriptor was then chosen along with the state-of-the-art SIFT descriptor to evaluate their accuracy rate using the BoVW model. We evaluated our COSFIRE descriptors along with other popular keypoint descriptors such as SIFT and BRISK. The performance of the COSFIRE-336 descriptor achieved the best performance results amongst the configurations proposed in this project, exceeding the SIFT and the BRISK descriptors’ performance in various image conditions. The COSFIRE-336 keypoint descriptor achieved an impressive accuracy rate when evaluated using the BoVW model, achieving a higher accuracy rate than SIFT on an unseen dataset of 15 different categories.