CMDSP: a completed multi-domain shrinkage pattern for texture image classification

Texture is a fundamental and crucial property of an object’s surface, providing essential information for identifying its characteristics. Extracting rich, comprehensive texture features remains a key technical challenge across various visual tasks. To address this, this article introduces a completed multi-domain shrinkage pattern for texture classification. To capture detailed texture information, the original image is first transformed into gradient and Gaussian domains, creating multi-domain image spaces. The primary challenge then becomes efficiently and effectively optimizing the selection of highly discriminative features across these multiple domains. Firstly, to maximize the preservation of local difference information, corresponding local shrinkage and local center patterns are derived in three image spaces using a shrinkage function. Secondly, this article designs an effective and efficient uniformity optimization mechanism that identifies the most discriminative shrinkage pattern from a multi-domain shrinkage pattern set. The local shrinkage pattern containing the most frequent uniform pattern is then selected as the optimal shrinkage pattern, as it provides the highest discriminative power. Thirdly, to prevent loss of information from discarded patterns, a uniformity assessment strategy is introduced. This strategy evaluates whether the three generated shrinkage patterns are uniform, producing three corresponding single-bit shrinkage assessment patterns. For compact fusion, a multi-pattern encoding strategy is used to re-encode the three single-bit shrinkage assessment patterns alongside the generated central patterns into a multi-domain shrinkage assessment pattern and a multi-domain central pattern. Finally, the optimal shrinkage pattern, the multi-domain shrinkage assessment pattern, and the multi-domain central pattern are jointly integrated to produce the final completed multi-domain shrinkage pattern for the texture classification task. Experimental results on three representative public datasets demonstrate that the proposed method achieves state-of-the-art performance.