Research on Unmanned Infrared Night Vision System Based on Improved Faster R-CNN Target Detection Algorithm

Abstract

An improved Faster R-CNN infrared target detection algorithm is proposed for the problems of large feature information loss, low accuracy, and weak performance in the detection process of infrared night vision images in unmanned technology. Firstly, the FPN structure of the spatial adaptive (ASF) module is utilized to reduce the loss of feature fusion information, secondly, the fully connected layer and the fully convolutional layer head structure are used to better represent the target regression and localization effects, and finally, the attention mechanism is introduced on the basis of the network model to improve the model performance. The experimental results show that the improved algorithm compares with the Faster R-CNN target detection algorithm, its mAP, mAPs, and mAPl rise by 1%, 4.1%, and 6.7%, respectively, and the effectiveness and feasibility of the improved Faster R-CNN algorithm are verified by multiple sets of experimental data.