High Resolution Remote Sensing Image Ship Small Target Detection Model Improvement And Algorithm Optimization

The characteristics of unique imaging angle,small target size and complex background make the average accuracy of remote sensing image ship small target detection algorithm still have room for improvement.In order to effectively improve the performance,this paper further improves the network structure of YOLOv5 algorithm and optimizes the remote sensing image ship small target detection algorithm,focusing on data set selection and preprocessing,YOLOv5 algorithm backbone network optimization,attention mechanism and bounding box optimization.1.This paper is to improve the ship small target detection algorithm in high-resolution remote sensing images,and the existing public data set does not have a separate ship data set.Therefore,this paper selects the data set containing the ship target from the existing public data set,enriches the data set through geometric transformation,median filtering and bilateral filtering to reduce image noise,and finally obtains 2100 images as the data set of this paper.It can effectively avoid over-fitting in training and improve the stability and generalization ability of the model.2.The bidirectional feature pyramid network BiFPN and adaptive spatial fusion ASF are introduced to optimize the backbone network of YOLOv5 algorithm.The PANet network in the original YOLOv5 algorithm is replaced by BiFPN,and different weights are assigned to different scale features to improve the feature fusion efficiency of the model.The ASF module is added to the improved spatial pyramid pooling SPPF to make the model have stronger feature expression ability.3.The YOLOv5 algorithm is prone to missed detection and false detection in small target detection.Therefore,this paper introduces attention mechanism and weighted box fusion method in neck network and head network.Based on the convolution block attention module CBAM,an efficient convolution attention module ECBAM is proposed by using the ECA module and weight calculation method in ECANet,which makes the model pay more attention to the important features in ship images and reduce the interference of background information.Compared with the traditional NMS and Soft-NMS,the weighted box fusion method can fuse the information of multiple bounding boxes and make the final bounding box more accurate.Through target detection experiments and analysis on DOTA and DIOR datasets,the average accuracy of YOLOv5 algorithm is 88.1%and 87.7%respectively,and the average accuracy of this algorithm is 92.8%and 92.5%respectively.The average accuracy of the latter increases by 4.7%and 4.8%respectively,and the detection speed decreases by 2.7fps.At the same time,the average accuracy of the proposed algorithm on other datasets AID and LEVIR is 90.6%and 90.7%respectively,which is 3.8%and 3.5%higher than that of YOLOv5 algorithm.In summary,the target detection algorithm based on the improved network structure of YOLOv5 algorithm proposed in this paper can better complete the task of ship small target detection,and the model has better generalization ability.