Infrared Fuzzy Video Data Cleaning Based on Improved YOLOv5

Aiming at the problem that infrared fuzzy video data affect video quality, an improved YOLO v5 deep network model is established to carry out intelligent and accurate infrared fuzzy video data cleaning. Considering the problem of missing infrared fuzzy video data, a dataset named SYLU-INFRARED VIDEO DATASET is constructed. This dataset includes 5000 aerial and ground background infrared fuzzy video frames. Considering the spatiotemporal joint features of video data, a CBAM attention mechanism with two-dimensional attention allocation strategies is applied. This attention can automatically allocate the importance of video features extracted by YOLOv5 backbone in both spatial and channel dimensions. The experimental results show that the fusion model of CBAM and YOLO v5 enhances the cleaning effectiveness. CBAM-YOLOv5 has an increase of 17.8%, 28.4%, and 26.5% in Precision, Recall, and mAP compared to YOLOv7, respectively.