The research of case retrieval system based on deep learning

Abstract

A large number of criminal investigation images are collected for modern case detection, and these images not only contain many valuable clues, but also can provide powerful evidence. At present, most of the current criminal investigation image retrieval methods for public security investigation applications use text-based or traditional shallow feature-based image retrieval methods, and the accuracy and efficiency of retrieval can hardly meet the needs of modern criminal investigation cases. In this paper, we summarize and analyze the current research results and technologies in the field of image retrieval, and adopt the image retrieval method based on depth features to improve the accuracy and efficiency of the criminal investigation present investigation image retrieval. In the depth feature-based image retrieval, VGGNet and Res Net are fine-tuned using the criminal investigation image dataset, and then the image depth features are extracted for retrieval experiments. The experimental results show that the retrieval model has the following two shortcomings: the model cannot adapt to the target scale change; the retrieval accuracy is lower than the average in the category with fewer samples. Two optimizations are proposed to address these problems: introducing pyramid pooling to improve the robustness of the model to target scale changes; retraining the network after enhancing the data samples to make the retrieval accuracy of the model more balanced for different categories of samples. In addition, query expansion is introduced to enhance the image feature representation during retrieval. After using the above optimization methods, the retrieval accuracy is improved by 5.7%.