Design And Implementation Of Object Detection Algorithm Based On Feature Fusion And Adversary Occlusion Networks

Object detection is an important research content of computer vision technology,and has been widely used in many fields such as intelligent security,unmanned driving,scene recognition,and medical diagnosis.However,due to the influence of many factors such as the scale changes,occlusion levels,and external environmental changes in real-life application scenarios,the current object detection algorithms have low accuracy for small-scale objects and partially occluded objects.Therefore,designing a detection algorithm that can cope with real scene changes and accurately detect objects is the focus of current object detection algorithms.Based on Faster R-CNN,this paper proposes an object detection algorithm based on feature fusion and adversary occlusion networks.By introducing feature fusion network and adversary occlusion network,this algorithm improves the detection performance of small-scale objects and partially occluded objects.Feature fusion network uses deconvolution operation to fuse high-level feature maps with low-level feature maps to increase the extraction capability of low-level feature maps in the network,and finally generate a single high-level feature map with high resolution and high semantic information and make predictions on it,so that small objects in the image can be detected more effectively.Adversary occlusion network creates occlusion on a deep feature map of the object,and generates an adversary training sample that is difficult for the detector to discriminate.At the same time,the detector classifies the generated occluded adversary samples accurately by self-learning.The two compete with and learn from each other to further improve the performance of the algorithm.In addition,this paper also uses an improved non-maximum suppression algorithm to filter the prediction boxes to improve the detection result when the objects overlap or the object is partially blocked by adjacent objects.The proposed algorithm was trained on the PASCAL VOC,MS COCO and KITTI data sets.A number of quantitative and qualitative experiments show that the proposed algorithm achieves a state-of-the-art detection accuracy.