Research Of Multi-label Clothing Image Classification Based On Deep Learning

In fashion analysis domain,combining deep learning with clothing images classification is one of the research hotspots.However,at present,clothing image classification is mainly divided into a single label to process separately.In real life,with the emergence of new fashion business schemas such as online shopping and the urgent need for complex decision-making,single-label clothing image classification can no longer solve these problem,multi-label clothing image classification has become an important learning problem,showing great application value.The goal of multi-label clothing image classification is to predict a set of clothing attribute labels for each clothing image.This paper deeply studies the methods of deep learning in the field of multi-label clothing image classification,focusing on solving the three typical problems of garment attribute recognition:discriminative feature learning,handling the correlations among clothing attribute labels,and imbalanced training data.Specifically,the main contributions of this paper are as follows:(1)A dual-stream feature fusion module is proposed to extract more discriminant features from the input clothing image and make full use of the global prior information and local example information in the input clothing image.In an image with multiple labels,there may be more than one object in an image,with different positions,scales and postures,a label may be associated with multiple objects.Moreover,some labels come from global information of clothing image instead of local information.Therefore,both the global and local information are important for classification.The module consists of two paths.A spatial pyramid transformation layer is introduced in the first path to learn the local characteristics of the multi-scale label-related instance;the second is to capture global priors from the input image as global features.(2)In order to deal with the correlation between clothing attribute labels and better improve the performance of the multi-label clothing image classification task,a graph convolutional neural network was constructed to capture the label correlation,and the nodes in the proposed GCN were mapped to a set of interdependent clothing attribute label classifiers.This set of classifiers is applied to the representation of clothing image features learned by the dual-stream feature fusion module,and the whole network can be trained end-to-end.In addition.an effective reweighting scheme is proposed by optimizing the correlation matrix of GCN.(3)The threshold tuning strategy is adopted to reduce the influence of data imbalance on the performance of multi-label clothing image classification model performance.In this paper,the same fixed threshold is not adopted for each label,instead,the threshold optimizer is used to automatically search for the optimal threshold of each label,and the iterative greedy strategy is used to update the threshold list column by column with the optimal threshold.Finally the updated threshold list is used to intercept the output probability to obtain the final evaluation result.In this paper,the multi-label clothing image classification algorithm based on deep learning is compared with other classical single-network models on FashionML,the basic clothing attribute dataset constructed in this paper.The objective result data show that this paper proposed algorithm has higher mAP value,C-F1 value and O-F1 value than the comparison algorithm;The subjective result display that this algorithm's multi-label output results are more comprehensive and accurate.