Research On Small Object Detection Method In Aerial Photography Based On Deep Learning

With the rapid development of deep learning theory and the upgrading of computer hardware level,the research of general object detection methods based on deep learning has made significant progress,and many representative algorithms have emerged: YOLO,Faster RCNN and Center Net,etc.,with a wide range of applications.However,when the general object detection algorithm is applied to the field of small object detection,there are phenomena such as network structure redundancy,low detection accuracy of small objects,and missed detection of dense small objects.Considering the requirements of accuracy and real-time performance in actual object detection scenarios,this paper studies the classical single-stage object detection algorithm YOLOv4 and improves it for the above problems.And make a small target data set,and experiment on it.First,in view of the network redundancy problem of YOLOv4 in the small object detection task,the network is simplified,and the feature extraction,feature fusion and prediction branch of the large target in the network are pruned,and a pruned version of the YOLOv4 network structure is proposed.The number of network layers of the improved network model is reduced,and the amount of parameters is about a quarter of the original.Experiments have proved that the pruned network structure can achieve detection performance similar to the YOLOv4 network structure on the small object data set.Then,aiming at the low detection accuracy of small targets and the missed detection of dense small targets,the feature fusion module of the pruned network was improved,and the feature information extracted from the shallow network was introduced into the feature fusion module to strengthen the application of the feature information of small targets,and the prediction branch of corresponding scale was added.Experiments have proved that the improved multi-scale feature extraction network can improve the detection accuracy of small objects,and the missed detection phenomenon is improved.Finally,the loss function is optimized for the imbalance of information learned by the convolutional neural network in the unbalanced data set.The category weight factor is introduced into the loss function,and different degrees of penalty intensity are adopted for different categories of objects to reduce the impact of unbalanced data distribution on the model.Experiments have proved that the use of an optimized loss function based on category weights can improve the detection accuracy of target categories with a small amount of data,and the overall detection performance of the algorithm is improved.