Research On Image Retrieval Method Based On Deep Learning

Content-based image retrieval methods have attracted more attentions of researchers in the field of computer vision,which extracts the global semantic features of natural images and then searches semantically similar images in the database.Image feature extraction plays a significant role in the retrieval tasks.However,the hand-crafted feature can only extract the local texture and color of the image and the global semantic features expressed by the image are ignored which overcomes the development of retrieval tasks.In recent years,the convolutional neural networks have made breakthroughs in the field of computer vision.The convolutional layer and the full connection layer all can be used to image retrieval tasks.Based on the convolutional network,the main work of this paper is as follows:(1)Based on the convolutional layer,this paper proposes a Spatial Weighted Feature(SWF)to map the three dimension feature maps to one dimension vector by assigning the corresponding weight to each pixel.We visualized the feature maps of the SWF method and the traditional linear additive aggregation,which demonstrated that the feature aggregation method proposed in this paper can highlight the contour of the region of interest and improve the retrieval performance.(2)Based on the full connection layer,traditional methods design an end-to-end framework to map the feature vector into low dimension by a nonlinear function.However,the gradient maybe explosion by Sigmoid function.Therefore,we add a classification layer in the first fully connection layer and apply constraint items in the encoding layer proposed new method Deep Constraint Binary Code.We analyze the distribution of encoding layer activations in different conditions,which show that we proposed DCBC methods can improve the discrimination of extracted feature and improve the retrieval performance.(3)In complex scenes,image retrieval tasks should extract the object which wants to retrieval.However,traditional image feature extraction methods extract the features from the global image rather the region of interest.We developed a new method which utilized the attention mechanism to extract the region of interest and then extract the feature to search for similar images.The experimental analysis shows that the image retrieval based on attention mechanism can better overcome the influence of interference pixels on the retrieval task and improve the retrieval performance.