Genetic programming-based learning of texture classification descriptors from Local Edge Signature

Abstract

Describing texture is a very challenging problem for many image-based expert and intelligent systems (e.g. defective product detection, people re-identification, abnormality investigation in medical imaging and remote sensing applications…) since the process of texture classification relies on the quality of the extracted features. Indeed, detecting and extracting features is a hard and time-consuming task that requires the intervention of an expert, notably when dealing with challenging textures. Thus, machine learning-based descriptors have emerged as another alternative to deal with the difficulty of feature extracting. In this work, we propose a new operator, which we named Local Edge Signature (LES) descriptor, to locally represent texture. The proposed texture descriptor is based on statistical information on edge pixels’ arrangement and orientation in a specific local region, and it is insensitive to rotation and scale changes. A genetic programming-based approach is then fitted to automatically learn a global texture descriptor that we called Genetic Texture Signature (GTS). In fact, a tree representation of individuals is used to generate global texture features by applying elementary operations on LES elements at a set of keypoints, and a fitness function evaluates the descriptors considering intra-class homogeneity and inter-class discrimination properties of their generated features. The obtained results, on six challenging texture datasets (Brodatz, Outex_TC_00000, Outex_TC_00013, KTH-TIPS, KTH-TIPS2b and UIUCTex), show that the proposed classification method, which is fully automated, achieves state-of-the-art performance, especially when the number of available training samples is limited.