Vehicle Small Object Detection In Remote Sensing Images

Target detection in satellite remote sensing images is of great significance in both military and civilian applications,with small target detection being a difficult and hot research topic.This article focuses on satellite remote sensing images with spatial resolutions of 0.5m and 1m,and conducts research on vehicle small target detection based on the YOLOv5 s algorithm.The main work and innovative points of this article can be summarized as follows:(1)In order to better study vehicle target detection in satellite remote sensing images with different spatial resolutions,a satellite remote sensing image dataset with unified spatial resolution and annotated with a large number of vehicle target instances was built.100 remote sensing images with a spatial resolution of 0.15 m were collected,and over 75000 vehicle target instances were labeled in horizontal boxes through labeling as a high-resolution remote sensing dataset.Vehicle small target data sets with spatial resolution of 0.5m and 1m are obtained by Gaussian blur and resampling.(2)In order to improve the detection performance of small vehicle targets in remote sensing images with a spatial resolution of 0.5m,the SA-YOLOv5 network is proposed.This network is based on YOLOv5 and designs a feature extraction network with fewer downsampling operations in the Backbone section to preserve the texture and geometric information of small targets;Introducing the Bi FPN structure in the Neck section to enhance the fusion effect of small target features by adjusting the weight of network branches;In the Head section,the CBAM attention mechanism module is utilized to suppress the noise impact of background and pseudo targets,in order to improve the signal-to-noise ratio of small targets.Compared with the prototype YOLOv5 s on the test set,SA-YOLOv5 s improved the Precision,Recall,and m AP metrics by 7.7%,13.6%,and 15.4%,respectively.(3)In order to improve the detection ability of YOLOv5 for vehicle small targets in remote sensing images with a spatial resolution of 1m,a weak target detection algorithm based on super-resolution reconstruction,CWT-YOLOv5,is proposed.Firstly,based on wavelet transform and residual network ideas,a CWT module was designed to extract three high-frequency components of the image,and the original low resolution image was used as the low-frequency component to achieve super-resolution reconstruction through discrete wavelet inverse transform.Finally,super-resolution reconstruction is combined with YOLOv5 s to achieve end-to-end detection of small and weak targets.The experiment showed that 91%,85.7%,and 92.9% performance was achieved on Precision,Recall,and m AP indicators.