Research on Single Object Detection Technology Based on Infrared Multi-spectrum Fusion

Infrared multi-spectrum fusion technology is investigated to detect single object in this paper. Considering the characteristics of multi-spectral image, the preprocessing algorithms are implemented on the infrared multi-spectral simulation system to enhance short-wave image and filter noise for medium-wave image and long-wave image. After comparison and design, Nonsubsampled Contourlet Transform is selected as the fusion strategy. For infrared object detection scenario, CenterNet-E network based on CenterNet is designed to achieve better performance in detection speed and accuracy. Its backbone feature extraction network is resnet50 and integrates attention mechanism, feature enhancement module, and multi-head prediction module. To evaluate the detection performance, Infrared Target Dataset (ITD) is built by means of optical modeling. The experiment results show that the mAP value of CenterNet-E on ITD is 3% higher than CenterNet and the frame rate is 49% higher than YOLOv4.