Image Representation Using The Intrinsic Texture Properties

Abstract

Texture is fundamental to computer vision, yet analyzing its structure re mains difficult due to variations in scale and pattern. By employing multi scale analysis based on wavelet theory, this thesis bridges the gap between classical signal analysis and modern deep learning to address these challenges and conduct an in-depth analysis of intrinsic texture properties. We introduce two complementary methods. The Wavelet Texture De scriptor (WTD) combines fixed wavelet decomposition with rigorous feature selection to maximize efficiency in limited data environments. The Data Driven Wavelet Transform (DDWT) takes this further by embedding a train able wavelet layer into a neural network, allowing the model to learn task specific wavelet filters rather than relying on rigid, fixed ones. Experimental evaluation confirms that WTD achieves state-of-the-art re sults, while DDWT offers superior adaptability for complex, heterogeneous textures with negligible additional parameters and minimal computational cost. Ultimately, this work proves that blending wavelet theory with modern learning creates robust, interpretable representations for visual recognition, extending the value of wavelets into the deep learning era.