Dual-stream infrared object detection network combined with saliency images
Peng Qin, Yunfeng Liu, Gaofan Zhou, Chuanming Tang, Jianlin Zhang · 2022 IEEE Asia-Pacific Conference on Image Processing, Electronics and Computers (IPEC) · 2022
Compared with visible spectrum images, thermal imaging has lower requirements for lighting conditions, but it contains many problems (e.g. blurred edges and low contrast). Aims to these defects, a dual-stream infrared object detection network (named DINet) is proposed. In order to bring into play the saliency information of the infrared object, the pseudo-color image is generated from the infrared image. Saliency information is employed as the attention mechanism of our detector. This framework design a dual-stream network with a cross-shaped receptive field aggregation module. It combined with convolutional neural network (CNN) with Transformer to create a local-global complementary dual-stream backbone, which separately extracts the feature of infrared and pseudo-color images. The prediction head fuses the adjacent layers to generate a more refined grid map, which contains high-definition positions and rich semantic information. The experimental results demonstrate that DINet achieves significant performance on infrared object detection benchmark FLIR. The quantitive analysis shows that the mean average precision can reach an excellent level.