Arbitrary-Oriented Object Detection on High Resolution Images Based on Differentiable Architecture Search
Lianyu Cao, Xiaolu Zhang, Zhaoshun Wang · Canadian Journal of Remote Sensing · 2021
High-resolution images exhibit wide field of vision, high background complexity, special angle of view, rotation, and small objects, thus making automatic object detection a challenging problem. Recently, this problem has been studied by many researchers through the application of deep learning methods, and good detection results have been achieved. However, most of the current networks for object detection on remote sensing images are designed manually, which is not necessarily the optimal structure. Therefore, DARTS-FPN, a network based on differentiable architecture search, has been constructed to improve the accuracy of object detection on high resolution images. The DARTS algorithm is first used to search the remote sensing images data set. The neural architecture search technique of NAS-FPN is then integrated into this network and merged with RetinaNet, a single-stage rotating network for object detection. Experiments are conducted on the DOTA data set to evaluate the performance of DARTS-FPN. Compared with existing classical networks, DARTS-FPN achieves 55.86%, 45.06%, 5.78%, and 2.19% higher mean average precision than the SSD, YOLOv2, R2CNN, and DM algorithms, respectively.