Research on image detection methods based on YOLO algorithm and deep learning

This research provides a comprehensive improvement of the YOLO algorithm's application in deep learning image detection. Key innovations include optimising the grid structure and improving the loss function, and innovation in training strategies. The introduction of attentive and multi-scale attribute blending techniques enhances the model's ability to detect targets at different scales. An improved multi-task loss filter is offered to balance bounding box regression, target classification and confidence prediction. A progressive learning method is adopted, combining curriculum learning and adaptive learning rate adjustments to enhance training effectiveness. Experimental results show that the enhanced YOLO model achieves 0.89 mAP and a recognition speed of 60 FPS on the COCO set, surpassing several mainstream algorithms. The study also explores model deployment and acceleration technologies, providing new insights for the promotion of YOLO algorithms in practical applications.