Research On Enhancement Technology Of YOLO Object Detection Algorithm

Object detection is one of the fundamental tasks in the field of computer vision.With the progress of society,the demand for real-world applications has been increasing,leading to higher requirements for object detection algorithms.Current object detection algorithms need to possess high detection performance,adaptability to diverse scenarios,real-time detection capabilities,and ease of use.Among them,YOLOv5 object detection algorithm is one of the most popular object detection algorithms,known for its excellent detection performance and real-time capabilities.However,realistic scenes are complex and variable,and there are many situations such as large-scale change,object occlusion,extremely small object and miscellaneous background information.These reasons will lead to poor detection effect with the YOLOv5 object detection algorithm.Therefore,a more in-depth study on the YOLOv5 object detection algorithm has great significant academic and application value.Accordingly,this thesis is based on the YOLOv5 object detection algorithm and aiming to address the issues of missed detections and false alarms in detecting small-scale objects.Multiple-stage improvements are made to the YOLOv5 object detection algorithm.Aiming at the improvement of multi-scale feature fusion,this thesis replaces the feature fusion structure of the original algorithm to strengthen the algorithm’s network multi-scale feature fusion capability,using the cross-layer connectivity and weight control features of the weighted bidirectional feature pyramid network.A multi-branch spatial feature pyramid module is constructed to provide the algorithm network with feature maps containing more informative data.Aiming at the improvement of the extraction of important features,a maxpool module is constructed so that the improved algorithmic network can obtain a feature map with a larger receptive field and more significant features.This thesis also reconstructs the attention mechanism module and adds it at different positions of the original algorithm network,enabling the algorithm network to focus more on important features and enhance its detection capabilities.In this thesis,the effectiveness of the improved algorithms is verified by comparative experiments and visualization on the VOC datasets.The proposed improvements have achieved significant performance enhancements compared to the YOLOv5 object detection algorithm.In this thesis,the improved algorithms and the fusion algorithms that combines these improvements are verified by comparative experiments and visualization in the application scenarios of face mask recognition and car recognition.The experiments demonstrated that the improved algorithms and the fusion algorithms proposed in this thesis outperformed the YOLOv5 object detection algorithm in both application scenarios.The best fusion algorithms in both application scenarios exhibited improved detection performance compared to the YOLOv5 object detection algorithm,with an increase in m AP values of 1.6% and 1.4%,and an increase in F1 values of 1.6% and 2.2%,respectively.Moreover,the detection speed of the best fusion algorithms exceeded 80 FPS in both application scenarios,meeting the requirements for real-time detection.Through real-world application experiments,the effectiveness of the proposed improvements scheme in this thesis based on the YOLOv5 object detection algorithm was demonstrated,and improves the detection performance of the YOLOv5 object detection algorithm.