Research On Small Object Detection Method In Remote Sensing Image Based On Improved YOLO Algorithm

With the characteristics of spatiality,multi-temporal,etc,remote sensing has made up for the problems of small field of vision and short information updating period.The remote sensing technology and computer vision are usually integrated for environmental monitoring,disaster protection and so on.At present,the image object detection technology has been mature in terms of face recognition,vehicle detection,etc.However,due to the long distance of remote sensing observation,wide range and complex image background,the existing object detection technology is not effective for remote sensing images.In this paper,an in-depth study was conducted mainly on the problem of small and dense objects in remote sensing images.To achieve the accurate detection of dense small objects in remote sensing images,a small object detection algorithmic model was studied and optimized based on YOLOV5 algorithm.The main research contents are presented as follows:(1)To solve the problem of a few pixels and information features of the small objects in remote sensing images,a method integrating the attention mechanism,local receptive field and refinement feature pyramid was proposed.In order to enhance the attention to effective features,the attention mechanism CA was introduced,so that the model could acquire more critical multi-scale local and global information from a large amount of information.To enhance the network context information,the local receptive field module was also introduced,so that the corresponding semantic information could be acquired through dilated convolution of different dilation rates in the receptive field.Furthermore,in order to enhance the feature expression ability of the network,the PANet feature fusion network in YOLOV5 was replaced by a dense feature pyramid network,and the refined feature(RFM)structure was introduced for adaptive feature fusion to eliminate the hierarchical conflict information.The experimental results on DOTA dataset reveal that the proposed algorithm can improve the detection accuracy of remote sensing small objects.(2)Aiming at the dense small objects and unbalanced sample distribution in remote sensing images,an object detection algorithm combining multiple iterations and hardeasy sample balance factor was put forward.By optimizing the confidence loss function and adding a balance factor,the contribution of easy samples to the total loss was reduced,and the penalty weight of hard samples was increased,so that more attention would be paid to the learning of hard samples in the detection process and the problem of imbalanced samples could be alleviated.Meanwhile,the multiple iteration module was adopted so that the above improved algorithm could obtain the missed small objects in the iteration process,so as to solve the detection missing problem of dense small objects.The experiment indicates that the detection accuracy of dense small objects in remote sensing images can also be improved.In this paper,the detection of dense small objects in remote sensing images in complex scenes was studied.Technologies such as attention mechanism,local receptive field and refinement feature pyramid,and an algorithm framework combining the easyhard sample balance factor and multiple iterations were proposed respectively.Relevant achievements can be widely applied to urban planning,environmental detection,disaster protection and other fields.