Clothing Classification Algorithm Based On Geometric Deep Learning

With the rapid development of clothing e-commerce platforms,computer vision based clothing classification methods are widely used in various fields such as clothing recognition,clothing retrieval and recommendation,and clothing fashion prediction.In recent years,with the development of deep learning technology,the performance of clothing classification has been significantly improved.However,traditional deep learning methods often view image data as vectors in Euclidean space,failing to fully utilize the potential low dimensional nonlinear geometric structure information in high-dimensional clothing image data.Therefore,this article explores and mines the inherent geometric structure information of clothing image data from the perspective of non Euclidean geometric deep learning methods,in order to achieve better clothing classification performance.Firstly,this article designs and implements a manifold structured clothing classification network based on second-order convolution,which utilizes second-order statistics of clothing features for image classification.The network structure includes covariance pooling module,manifold structure neural network module,prediction module,etc.Firstly,the input clothing image features are extracted through convolutional neural networks.After passing through the covariance pooling module,the second-order statistical covariance is obtained and converted into an SPD manifold to represent the feature information of the clothing image set(video).Then,a complete manifold structured neural network is constructed,enhancing the model’s ability to represent the geometric internal structure of the clothing image set and effectively improving classification accuracy.The experimental results on the multi view clothing image dataset MVC show that this method has good effectiveness,robustness,and accuracy.Secondly,in response to the problem that second-order convolutional methods can only handle SPD manifolds and have a large number of parameters,this paper designs and implements a manifold value convolutional clothing classification network based on the Frechet mean.The network structure includes manifold value convolution module based on Frechet mean,activation function module based on Frechet mean,manifold value full connection module based on Frechet mean,prediction module,etc.Specifically,we convert each clothing feature into an SPD matrix,perform manifold value convolution operations,and aggregate it onto a global feature vector.Finally,we use a classifier to classify the vector.And comparative experiments were conducted with neural networks using traditional Euclidean mean convolution.The results indicate that the method using FM has significantly improved classification accuracy and average accuracy.Through experiments and analysis on the public clothing image dataset MVC,it has been proven that the proposed method in this paper has higher accuracy and precision compared to traditional convolutional neural network clothing classification methods when the perspective of clothing images changes.Compared with the manifold valued method,the second-order convolution method has more flexibility in network structure design,while the manifold valued convolution classification method runs faster and has lower parameter quantities.Overall,the two proposed methods each have advantages and disadvantages and are more suitable for application scenarios,but both can effectively capture local details and global feature information of clothing products,providing more efficient,accurate,and intelligent classification methods for the clothing industry.