Genetic Programming for Automatically Synthesising Robust Image Descriptors with A Small Number of Instances

Image classification is a core task in many applications of computer vision, including object detection and recognition. It aims at analysing the visual content and automatically categorising a set of images into different groups. Performing image classification can largely be affected by the features used to perform this task. Extracting features from images is a challenging task due to the large search space size and practical requirements such as domain knowledge and human intervention. Human intervention is usually needed to identify a good set of keypoints (regions of interest), design a set of features to be extracted from those keypoints such as lines and corners, and develop a way to extract those features. Automating these tasks has great potential to dramatically decrease the time and cost, and may potentially improve the performance of the classification task. There are two well-recognised approaches in the literature to automate the processes of identifying keypoints and extracting image features. Designing a set of domain-independent features is the first approach, where the focus is on dividing the image into a number of predefined regions and extracting features from those regions. The second approach is synthesising a function or a set of functions to form an image descriptor that aims at automatically detecting a set of keypoints such as lines and corners, and performing feature extraction. Although employing image descriptors is more effective and very popular in the literature, designing those descriptors is a difficult task that in most cases requires domain-expert intervention. The overall goal of this thesis is to develop a new domain independent Genetic Programming (GP) approach to image classification by utilising GP to evolve programs that are capable of automatically detecting diverse and informative keypoints, designing a set of features, and performing feature extraction using only a small number of training instances to facilitate image classification, and are robust to different image changes such as illumination and rotation. This thesis focuses on incorporating a variety of simple arithmetic operators and first-order statistics (mid-level features) into the evolutionary process and on representation of GP to evolve programs that are robust to image changes for image classification. This thesis proposes methods for domain-independent binary classification in images using GP to automatically identify regions within an image that have the potential to improve classification while considering the limitation of having a small training set. Experimental results show that in over 67% of cases the new methods significantly outperform the use of existing hand-crafted features and features automatically detected by other methods. This thesis proposes the first GP approach for automatically evolving an illumination-invariant dense image descriptor that detects automatically designed keypoints, and performs feature extraction using only a few instances of each class. The experimental results show improvement of 86% on average compared to two GP-based methods, and can significantly outperform domain-expert hand-crafted descriptors in more than 89% of the cases. This thesis also considers rotation variation of images and proposes a method for automatically evolving rotation-invariant image descriptors through integrating a set of first-order statistics as terminals. Compared to hand-crafted descriptors, the experimental results reveal that the proposed method has significantly better performance in more than 83% of the cases. This thesis proposes a new GP representation that allows the system to automatically choose the length of the feature vector side-by-side with evolving an image descriptor. Automatically determining the length of the feature vector helps to reduce the number of the parameters to be set. The results show that this method has evolved descriptors with a very small feature vector which yet still significantly outperform the competitive methods in more than 91% of the cases. This thesis proposes a method for transfer learning by model in GP, where an image descriptor evolved on instances of a related problem (source domain) is applied directly to solve a problem being tackled (target domain). The results show that the new method evolves image descriptors that have better generalisability compared to hand-crafted image descriptors. Those automatically evolved descriptors show positive influence on classifying the target domain datasets in more than 56% of the cases