Research On Criminal Scene Investigation Image Retrieval Based On Deep Learning

With the rapid development of information technology in China,a large number of criminal scene investigation images have been collected for use in modern case investigation.The images not only contain many valuable clues,but also provide strong evidence for case investigation.At present,retrieval methods based on key words and traditional shallow features are used for criminal scene image retrieval of the public security investigation application.However,the accuracy and efficiency of retrieval cannot meet the needs of modern case investigation.For increasing the accuracy and efficiency of criminal scene investigation image retrieval,on summarize of the research and techniques in the field of image retrieval,this paper adopts a retrieval method based on deep feature to increase the accuracy and efficiency of image retrieval.The main work includes:In the first section of this paper,the author uses fine-tuned VGGNet and ResNet to extract deep feature for image retrieval.Experimental shows that these models have the following problems: first,cannot adapt to the variant of image scale,and second,being unbalance in retrieval accuracy for unbalanced sample classes.Thus proposed the following optimizations: first,introduced spatial pyramid pooling to improve the robustness of the model for scale variant of objects in images,second,adopt data augmentation on samples to make it balance and retrain the model,thus improve the retrieval accuracy for classes with less samples.In addition,query expansion was introduced to enhance the expressive ability of image feature.By means of these optimizations the retrieval accuracy increased by 5.7%.In the second section of this paper,the author proposed a retrieval method with multi-level index to increase the efficiency of retrieval on massive image samples,in which the expression of image was abstracted into three levels through convolutional neural network.The first level was semantic descriptor,being used to build inverted index,the second level was hash code,being used to execute approximate nearest neighbor search,and the third level was deep feature,being used to make precise match.For utility of conditional query in multiple levels,this retrieval method avoids sequential scanning in image feature database.Moreover,the author uses retrieval re-ranking and query expansion to reduce the accuracy loss of approximate nearest neighbor search.Experimental results show that the efficiency increased by 87% while the accuracy remains at the same level.In the third section of this paper,according to the requirement of image management and retrieval for public security department,the author encapsulated the retrieval model and algorithm into web services,and based on these services,designed the Yunying Image Management System for Criminal Scene Investigation.