Nonsalient Object Detection Algorithm Based on Infrared Image

In view of the current situation of low visibility at night and increasing poaching activities in wildlife reserves, the use of unmanned aerial vehicle (UAV) monitoring equipped with thermal infrared (TIR) cameras has become a trend. To overcome the difficulties of small objects and low resolution in infrared images, this letter proposes a nonsalient object detection algorithm in infrared images based on deep learning. Based on Grid R-CNN, we designed a feature extraction network, improved the region proposal network by guided anchoring (GA-RPN), and introduced the slice inference mechanism. Our method makes the network selectively fuse multiscale features to generate high-quality proposals and improve detection accuracy. Experimental results show our method is superior to other object detection algorithms for nonsalient elephant images on the BIRDSAI dataset.