Research On Deep Local Feature Extraction Model Based In Image Retrieval

With the rapid development of information technology in various fields,content-based image retrieval technology has bec ome one of the research hotspots in this field,and feature extraction model is one of the most critical.With the rapid rise of deep learning,the researchers found that the traditional feature extraction model based on the underlying features of the imag e,such as color,texture and shape,has some limitations.The main performance is to make the semantic regions strong and the search task meaningless areas equal,forming a large number of redundant characteristics.The development of deep learning and vi sual attention mechanism provides a new way of solving this problem,the depth features extracted through deep learning are often more semantic than the traditional features,and the application of visual attention,that is,by imitating the human eye,automatically concentrates attention on the more valuable areas of the task,strengthens the characteristic weight useful to the task,and neglects the meaningless feature.Such as the integration of visual attention mechanism of the DELF(DEep Local Feature)model,it is better to make up for the limitations of traditional image retrieval.The main work of this paper is to study and optimize the DELF model,and the research work starts from the following two aspects.First,based on the optimization of chan nel domain attention mechanism.In recent years,many scholars have studied visual attention,dividing visual attention according to different scopes into channel domain attention,spatial domain attention,and so on,often the model effect of combining at tention in different scopes is better than that of a single attention mechanism.Therefore,this paper holds that there is some optimization space in the DELF model,which only integrates the attention of the spatial domain.The DELF model is optimized by introducing channel domain attention and combining channel domain attention with spatial domain attention.Second,optimize the network structure.The continuous development of convolutional network leads to the increasing complexity of network model,how to optimize the network structure and reduce the complexity of the model has become another research hotspot.By consulting the relevant literature of network structure optimization,this paper finds that most networks have certain redundant parameters,that is,the utilization rate of parameters is not high,based on this idea,this paper introduces the idea of group convolution to improve parameter utilization and optimize the feature extraction model under the condition of ensuring the complexity of the model.Based on these two optimization strategies,experimental comparisons are made with the original DELF model in different data sets.Experiments show that the attention of two different domains is combined at the same time,and the feature extraction model of the network infrastructure optimization is introduced by the group convolution thought,which has better performance in image retrieval tasks.