Detection of rotating small targets in remote sensing images based on improved yolov5s

Abstract

Aiming at the characteristics of clustering, multi-scale changes, low contrast, and complex background in remote sensing images, this paper proposes an improved rotation frame algorithm based on yolov5s network, which reduces the traditional horizontal detection frame in remote sensing images by rotating frames. Missing detection of small targets due to aggregation and rotation problems; by using the attention mechanism module CBAMC3 that combines space and channels in the Backbone part to replace the original C3 module, the feature representation ability of targets of different scales in remote sensing images is improved; In the loss function, the Focal Loss function is used to reduce the complex background of the remote sensing image and the certain influence of the imbalance of positive and negative samples, and strengthen the training weight of the positive samples; at the same time, the Copy Paste data enhancement is used to enrich the number of samples on the data set. Enhance the generalization ability of the model.The result on the DOTA dataset reaches map86.4%, which is 3.1% higher than the 83.3% of the original yolov5s, which proves that the model has high detection accuracy in remote sensing detection and has certain reference value for detecting images under remote sensing.