Research On Improved Algorithm Of Aircraft Small Target Detection In Color Remote Sensing Images

Target detection in color remote sensing images is of great significance.Because remote sensing images have the characteristics of long imaging distance,multiple and complexity targets,the detection of small targets such as airplanes needs to be focused on.This paper takes airplane as the representative to complete the research of small target detection in remote sensing images.The high efficiency and high accuracy of detection algorithm is the key to related research.In recent years,the target detection algorithm based on deep learning has performed well in practical applications,so it is considered to be applied to the task of detecting small aircraft targets in color remote sensing images.Aiming at the problems of complex background information,small size and different scales of aircraft targets,the YOLOv3 algorithm was selected to detect small aircraft targets,and the algorithm was improved to further improve the detection accuracy.Finally,an improved YOLOv3 target detection algorithm based on spatial pyramid pooling and attention mechanism is proposed.The main research content of this article has the following four parts.1.The open data sets of color remote sensing images are different in format and contain a variety of targets,so they are not suitable for the research of aircraft target detection directly.Therefore,I process the public data set by means of rotation,cropping,scaling and other methods,and made a color remote sensing image data set,and finally obtained 1200 images.It effectively avoids the over-fitting phenomenon in model training,and improves the stability and generalization ability of the model.2.In order to match the anchor frame of the YOLOv3 algorithm with the target to be detected in this research,the K-means clustering algorithm is used to generate a new anchor frame on the self-made data set.At the same time,the original frame loss function is optimized,and the generalized intersection is used to replace the original IOU than GIOU,so as to more accurately reflect the coincidence of the anchor frame and the bounding frame,and enhance the accuracy of the target detection result.The optimized YOLOv3 algorithm has an average accuracy of 92%,which is 1.8%higher than before optimization;the detection rate is 0.0434s,and each image is reduced by 0.013s on average.3.The spatial pyramid pooling and CBAM attention mechanism are introduced on the basis of the YOLOv3 algorithm optimized by the frame loss.The Spatial Pyramid Pooling(SPP)module uses three different pooling cores to perform a fixed block size pooling operation to achieve the fusion of local features and global features.The channel and spatial attention modules in the CBAM attention mechanism refine the intermediate feature maps,adjust feature weights,reduce the weights of background features,and increase the attention to effective features.By testing and analysing,the algorithm introduced the new module has an average accuracy of 93.7%,a recall rate of 97.6%,and an average detection rate of 0.0472s per picture.4.Adjust the FPN structure in the network,use multiple levels of features for learning,and integrate different levels of features to enrich the semantic information of high-dimensional features,and enhance the entire network’s ability to detect small targets.An improved YOLOv3 target detection algorithm based on spatial pyramid pooling and attention mechanism is proposed.The average accuracy of the final improved algorithm is 4.3%higher than that of the previous algorithm,and the recall rate is increased by 10.92%.Compared with the Fast R-CNN algorithm,the detection rate is increased by 0.1337s when the average accuracy is only 1%lower.The improved algorithm is tested on the public data set.The average accuracy is 91.6%,the recall rate is 92.4%,and the detection rate is 0.0625s.Experimental studies have proved that the improved YOLOv3 target detection algorithm based on spatial pyramid pooling and attention mechanism proposed in this paper can well complete the task of detecting small aircraft targets in remote sensing images,and has good robustness and generalization capabilities.