Deep soft threshold feature separation network for infrared handprint identity recognition and time estimation

Abstract

With the development of hardware devices, infrared technology has become an important detection method in criminal investigation, military and other fields. Infrared technology can capture the thermal infrared information of targets, achieve efficient and rapid individual identification, provide important technological support for crime investigation, improve case solving rate and public safety level. However, infrared technology still faces some difficulties. Due to factors such as environmental temperature changes and heat transfer, the infrared images captured over time gradually blur, resulting in highly aliased image features and difficulty in extracting effective information. This article addresses this issue by combining the specific tasks of infrared fingerprint identity recognition and time estimation. By analyzing the captured thermal traces to predict the target identity and departure time, a deep soft threshold feature separation (DSTFS) network is proposed. This network effectively improves the accuracy of identity recognition and time estimation by separating the characteristics of identity information and time information. Specifically, this article is based on the ResNet backbone network, extracting common features through shallow residual convolution blocks, extracting corresponding features for different tasks using branch residual convolution blocks, and introducing spatial and channel attention mechanisms to solve the problems of deep blur and feature separation. In order to enhance the model's expressive power, the soft threshold activation function SPRelu was adopted. By dynamically balancing the separation of features to learn weights, the weight balancing problem is solved, and periodic validation is used to evaluate task performance. According to the task requirements, we collected infrared fingerprint images of multiple testers and constructed an infrared hand heat trace dataset with labels such as departure time, gender, hand posture, and identity category as experimental data. Through the analysis of experimental results, our proposed deep feature separation network has better identity prediction results compared to other deep learning models in infrared fingerprint identity recognition and time estimation tasks. At the same time, it performs well in time estimation tasks, with lower error rates and higher stability.