| This thesis focused on how to improve the detection performance of object detection algorithms in computer vision and artificial intelligence technology in the field of small object detection and achieve significant weight reduction of the algorithm.Based on deep learning and neural network theory,and based on YOLO series of object detection algorithms,a series of methods including complex long-distance dense face detection,small object search detection on the sea surface,general life scene detection small object detection and algorithm lightweight improvement strategies in complex and unknown environments,including unmanned aerial vehicle(UAV)air to ground small object detection and other scenarios.Through experimental comparison,the new algorithm formed by combining the improved strategy in this article has better detection performance and lightweight effect than the original algorithm in various small object detection scenarios.The specific research content of this article is as follows:Aiming at the existing object detection algorithms that focus on detecting full-size targets and less consider the detection optimization of small targets in special scenarios,as well as the problems of excessively large models and difficult deployment during algorithm deployment,this thesis comprehensively analyzed the model composition and detection principles of mainstream object detection algorithms,and proposed multiple methods to improve the accuracy of the algorithm based on YOLOv5.An original strategy to reduce the consumption of algorithm computing resources,for example,by reasonably tailoring the model features to extract the final feature map output of the network portion,significantly reducing the computational resources required by the model and achieving significant lightweight of the model.An improved feature fusion method(PB-FPN)for small object detection based on PANet and Bi FPN is proposed to effectively improve the small object detection ability of the model.By introducing the spatial pyramid pooling layer(SPPF)in the feature extraction network into the feature fusion network and connecting it with the model prediction head,the comprehensive detection performance of the model is effectively improved.At the same time,based on the research of lightweight small object detection algorithms,this article used object detection algorithms combined with unmanned aerial vehicle air to ground detection scenarios as an application.Aiming at the problems of common object detection algorithms in unmanned aerial vehicle air to ground detection scenarios,such as excessive algorithm models,high deployment difficulties,and small-scale object detection difficulties,a series of improvements have been made to the classic YOLOv5 algorithm.A novel dual branch CSPNet(DR-CSPNet)architecture is proposed,which effectively reduced the complexity and computational complexity of the algorithm model.A novel feature fusion path(FS-FPN)is proposed,which effectively improved the comprehensive detection accuracy of the model.By integrating a new attention mechanism(ACmix),the performance of the algorithm in UAV air-to-ground detection scenarios is effectively improved.We named the original algorithm SF-YOLOv5 and UAV-YOLOv5,which integrate the improvement theory and strategy proposed in this article,and demonstrated the advantages,effectiveness,and versatility of the improved algorithm through detailed experiments. |
