Research On Texture Recognition Based On Deep Learning

Texture recognition is an important research direction in the field of computer vision.By analyzing the texture information on the surface of an object,it is possible to achieve recognition and classification of the object,which has extensive applications in scientific research and engineering technology.The extraction of texture features is the foundation of texture recognition.Traditionally,it is believed that texture is the result of the distribution or arrangement of texture elements in a certain pattern,and texture features can be described by statistical local information.However,in practical applications,texture images are captured under different lighting,rotation,and scale conditions.Existing texture recognition models mostly ignore these variations,which affects the reliability and robustness of the models.In addition,the lack of large-scale texture datasets also limits the application of deep learning in texture recognition.To address these issues,this paper focuses on researching and improving traditional deep learning methods.Specifically,it proposes an Elastic Angular Margin Loss function and an improved Trusted Multimodal Classification method,which enhances the generalization ability and robustness of the texture recognition models.The main research work is as follows:(1)This paper proposes an Elastic Angular Margin Loss(EAML)function,which improves the discrimination and generalization abilities of the texture recognition models.Loss function is crucial for ensuring inter-class separability and intra-class compactness in texture recognition tasks.EAML relaxes the restriction of fixed boundary values and allows for flexible promotion of class separability.Its main idea is to randomly draw boundary values from a normal distribution at each training iteration,providing opportunities for expansion and contraction of the decision boundary to allow for flexible learning of class separability.Currently,commonly used classification and recognition methods typically add a fixed boundary value to the Softmax loss to increase the discrimination ability of the models.However,these methods assume that sample features are uniformly distributed around class centers and cannot handle well with the inconsistent variations of actual texture features within and between classes,which limits the discrimination and generalization abilities of the texture recognition models.Experimental results show that EAML achieves higher recognition accuracy than the basic Softmax loss and Additive Angular Margin Loss(AAML)on LMT-108 and LMT-184,demonstrating that EAML can improve the discrimination and generalization abilities of the texture recognition models.(2)This paper introduces a Multi-Modal Evidence Fusion Classification(MMEFC)method and improves the evidence fusion rule to improve the reliability and robustness of texture recognition models.The method makes decisions by fusing the uncertainty of information from multiple modalities to dynamically assess the quality of data from different modalities.Its main idea is to model the probability distribution of categories using a Dirichlet distribution,which is parameterized by evidence from different modalities and fused using Dempster-Shafer evidence theory.The improved evidence fusion rule reduces computational complexity.The method provides uncertainty not only for each modality but also for the overall classification,which is crucial for credible classification and interpretable fusion.Traditional multimodal learning assumes that the quality of data from each modality is basically stable for all samples,but in reality,there may be differences in the quality of data from different samples and modalities,which limits the reliability and robustness of classification and recognition.Experimental results show that MMEFC outperforms the Deep Ensemble(DE)and the Evidence Deep Learning(EDL)in recognition accuracy on LMT-108 and LMT-184.Moreover,as the standard deviation of noise increases,the recognition accuracy of MMEFC is much less affected than other methods such as DE and EDL,demonstrating that it can significantly improve the reliability and robustness of the texture recognition models.