Clothing Retrieval Algorithm Based On Deep Metric Learning

Visual content-based clothing retrieval has become increasingly important for e-commerce. Due to the complexity of its structural features, the accuracy of clothing retrieval has always been in need of improvement. A large number of researchers have applied deep metric learning methods to this field of research. Addressing the issue of repeated learning on a minority fixed set of samples for conventional hard triplet, this paper proposes a clothing retrieval algorithm based on hard negative sample generation. The algorithm utilizes a multi-scale feature fusion architecture to construct the back-bone network and dynamically constructs hard sample tri-plets in the embedding space, thus addressing the issue of the unreasonable mining of hard sample triplet in traditional approaches. The proposed scheme is evaluated on clothing da-tasets. Experimental evaluations show the proposed algorithm achieves an average retrieval accuracy of 87.86
^%
, which is higher than that of the traditional hard sample triplet algorithm, indicating its promising potential for efficient ap-plication in clothing image retrieval.