Deep Learning Based Commodity Object Retrieval

It can greatly improve the shopping experience of users by accurately retrieving the commodity objects.In this paper,we propose two kinds of commodity object retrieval algorithms based on deep leaning,which are the saliency detection based unsupervised commodity object retrieval algorithm and the multi-attention based cross-domain beauty product retrieval algorithm.In the first algorithm,since most commodity objects are conspicuous and not complicated in commodity images.The proposed scheme utilizes the saliency box predicted by saliency detection to filter the proposals extracted by Selective Search.The reserved proposals have a big overlapping ratio with saliency box to a large extent.This work composes both the saliency box and the reserved proposals as saliency proposals.Further,we propose a channel weighting generalized mean pooling(CWGMP)feature to represent saliency proposals.On the one hand,the reduction of proposals' number after filtering significantly improves the computational efficiency;on the other hand,the new feature more accurately represents the objects to be retrieved which results in higher retrieval precision.In addition,we built and manually annotated a commodity dataset named PRODUCT.Extensive experiments demonstrate the superior performance of our scheme compared to other state-of-art methods.In the second algorithm,we extract a subset of the images in the Perfect-500 K dataset to build a “few-shot” beauty product dataset with clear label information.Considering that the beauty product objects are conspicuous and not complicated in images and the text regions in the images have strong discrimination,we propose an end-to-end classification network based on the saliency and text attention mechanism.The network focuses on accurately learning the feature representation of the beauty product by using attention mechanisms,and suppresses interference such as complex image background and scale changes,etc.Further,in order to more accurately retrieve the instance beauty products,we propose a saliency based regional maximum activation of convolutions feature,which takes into account the fine-grained features of beauty products by aggregating the local features of the salient regions.The retrieval performance of this algorithm on the Perfect-500 K dataset outperforms state-of-the-art methods.