A object detection method based on attention mechanism and reinforcement learning

Object detection technology is now widely used in areas such as unmanned vehicles, public safety, and intelligent robotics. However, the complexity and variability of object detection contexts and the lack of ability of deep learning-based object detection techniques to store sequences of information and make decisions result in performance that is not adequate for real-world scenarios. As the main components of the two-stage algorithm, feature extraction and region selection play a key role in the classification and location of target detection. Due to the superposition of network layers in deep learning, the receptive field is increased, while the correlation between feature maps and the decline of network gradient over a long distance is ignored. The different sizes of feature maps make the network objects generated by regions less dependent. In this paper, a new target detection model is proposed, and the existing problems are studied and solved from the residual feature extraction network, the tree-like deep reinforcement learning area generation network, and the fine-tuning network, respectively. Finally, the experimental and visual results verify the superiority of the overall performance of the algorithm model.