Research On Small Object Detection Method For Unmanned Aerial Vehicle Images

UAV aerial photography has the advantages of flexible maneuverability,high efficiency,low operating costs,and wide applicability,making it an attractive tool for object detection in UAV images with great commercial potential and extensive application areas.Deep learning-based object detection methods have developed rapidly in recent years,with superior performance,convenient design,easy operation,and strong generalization ability,which can meet the requirements of small object detection tasks in UAV images.Objects in images captured by UAVs are small in scale,have low pixel density,and limited features,making them more susceptible to lighting,occlusion,and other effects.In addition,small objects in UAV images are often densely distributed,and images of the same object taken by the UAV at different heights,angles,and motion states may have certain deformations.Currently,there is no large-scale standardized UAV image dataset,and class imbalance still exists in existing UAV datasets,affecting the training results of detection algorithms.These are the challenges that researchers need to face.In this thesis,we analyze and study the YOLO series models and improve them based on the YOLOv5 network.The main contributions of this thesis are as follows:(1)We propose the Only FPN YOLO(OF-YOLO)network model framework,which optimizes the feature fusion structure of the neck to closely integrate the backbone and neck features.The prediction part can more fully utilize the detail information of the shallow network to effectively extract features of small objects.At the same time,a compressed excitation attention module is introduced in the main part of the model,which can adaptively learn the correlation between input feature map channels and adjust the weights of different channels.(2)We propose a cross-stage connection module,which adopts cross-level connections when stacking convolutional blocks,optimizing the gradient flow branches of the module.This effectively aggregates features and extends the shortest gradient path of the network model,enabling the model to obtain richer gradient flow information while ensuring lightweight.As a result,it effectively improves the small object detection performance for drone images.(3)We propose an improved intersection over union(Io U)loss function,which introduces a coefficient on the basis of complete Io U and powers it.Practical experience has shown that this has a significant effect on targets with high Io U values,and more reasonable weight coefficients are set when calculating the loss function.On this basis,we also use an enhanced K-means anchor box clustering algorithm,which optimizes the selection of initial values compared to the K-means algorithm,making the clustering results more accurate and obtaining more accurate anchor boxes.The OF-YOLO model proposed in this thesis provides multiple network models with different widths and depths,which can be selected according to task requirements and deployment environments.Experimental results show that compared with the benchmark model with the same width and depth settings,the OF-YOLO network not only compresses the model volume but also effectively improves the detection accuracy of small objects,reducing missed detection and false detection.Taking the optimized OF-YOLOM algorithm as an example,its parameter volume is 16123815,which is 23% less than YOLOv5 m,and its Recall,m AP,and F1 score are higher than the benchmark model.The m AP of the optimized OF-YOLOM model reaches 42.4%,which is 6 percentage points higher than YOLOv5 m,an improvement of 16.5%,and 3.9 percentage points higher than YOLOv6.