Vision-based edge end high-speed anomaly detection method

The invention provides a vision-based edge end high-speed anomaly detection method, and belongs to the technical field of video processing. The method is an anomaly detection method based on image reconstruction, and a model is only trained on a normal image; comprising the following steps: inputting a single-frame image during detection, firstly extracting image features through an encoder, then carrying out feature fusion through spatial pyramid pooling, then reconstructing the input image through a decoder, and finally comparing the input image with the reconstructed image through a detection head to obtain an abnormal score graph. Setting an abnormal score threshold value, comparing the abnormal score with the score threshold value, and if the score is higher than the threshold value, indicating that the corresponding position on the image is abnormal, and obtaining an abnormal area mask image; and the model can be deployed to an edge end and reasoned in real time by adopting quantitative processing. When