一种基于卷积神经网络的多特征服装图像检索方法

A Clothing Image Retrieval Method Based on Convolutional Neural Network for Multi-feature

在服装图像检索问题上,由于受服装样式个性化和背景因素的影响,传统的神经网络不能准确地提取出图像细节特征,会导致检索不理想.本文在残差神经网络的基础上,利用多尺度串联图像特征,提出了一种多尺度特征的最大池化方法.该方法能输出固定维度的特征向量,同时大幅度减少了池化层的模型参数.基于这种新型池化方法再融合哈希函数的相似度信息,提出了一种新型服装图像检索算法.实验证明,这种算法能明显提升检索的准确度.

Due to the effect of the personalization of clothing style and background factor,the traditional neural network cannot extract image details accurately in the problem of image retrieval,which leads to unsatisfactory for image retrieval. To solve this problem,this paper builds a residual neural network,and proposes a new type of pooling method based on multi-scale feature. It can reduce the pooling level of mode parameters greatly. At the same time,it can output the information of feature vector for image dimension unchanged. A clothing image retrieval method based on convolutional neural network for multi-feature is proposed. It makes full use of the validity and hierarchy of deep convolutional neural network in image feature extraction,and the similarity information of hash function. The experiment results show that this method can improve the retrieval accuracy significantly.