An Improved Adaptive Angle Weakly Supervised Learning Object Detection

Remote sensing target detection can provide timely and accurate information, assist decision-making and management, and promote sustainable development of society, economy, and environment. The existing remote sensing images have characteristics such as variable target directions, large dataset size and difficulty in annotation, a large number of small targets, and complex surrounding environments.To solve these problems, this paper proposes a framework, which includes two stages. In the first stage, the algorithm is based on weakly supervised learning. By filtering and annotating the target project, large-scale annotation data is obtained. A weakly supervised pre training model based on rotation angle was used to learn a large number of feature representations.In the second stage, the high-level feature enhancement model of FPN is added to the backbone network to enhance the recognition ability of small targets. Generate rough oriented bounding boxes and classification scores using anchor free box object detection. On the basis of the original detector structure, a rotation feature extraction module (RRoI) is added, and additional angle parameters are added to adapt to the target detection of the rotation box. Adopting feature alignment and improved Fast R-CNN detection head for refinement processing, further improving detection accuracy.The experimental results show that the detection accuracy achieved $77.65 \%$ in the large-scale target detection data DOTA V1 dataset of remote sensing images. Compared to the benchmark network, mAP has increased by $1.5 \%$. The above results indicate that compared with traditional object detection methods, our proposed Fast ResNet ARM strategy uses an improved object detection model to achieve optimal overall performance while addressing large-scale data annotation.