Keypoint-based defense method for object detection

Object detection is one of the very important tasks in computer vision, which is not only widely used in industrial fields such as industrial control and aerospace, but also everywhere in life fields such as human flow monitoring and service robots. In recent years, with the continuous development of deep convolutional neural networks, the accuracy of target detection models based on deep learning is getting higher and higher, but due to the vulnerability of neural networks, the reliability and security of target detection models face new challenges. In order to improve the ability of the target detection model to defend against adversarial attacks, this paper proposes a multi-model defense method based on keypoints. The method uses the SIFT algorithm to extract keypoints in the image, which are insensitive to adversarial attacks. These keypoints are used to combine into a rectangular anchor frame, which is trained to obtain the results of target detection, which is combined with the results of another target detection model to improve the detection accuracy of the model.