Research on Vehicle Detection Accuracy Improvement of Convolutional Neural Networks in Low Contrast Remote Sensing Images

In order to solve the problem that low-contrast images in UAV remote sensing, especially the shadow coverage of trees and buildings, affect the detection accuracy of vehicle targets, this paper takes chroma enhancement as an example to conduct spectral enhancement experiments based on SSD and YOLO v3s. Since the object to be detected and the background in the low-contrast image are easily confused, the detection accuracy can be effectively improved by using the method of chromaticity enhancement to improve the degree of discrimination between the object to be detected and the background. The experimental results show that the detection accuracy of the original SSD and YOLO v3 for vehicles in low-contrast images is 93.4% and 80.0% respectively, and the detection accuracy of the two algorithms after chroma enhancement is increased by 3.1% and 9.5% respectively. The original algorithm is basically flat. Therefore, the method can improve the detection of vehicle objects in low-contrast images without affecting the detection speed.

I. Introduction

Most of the current vehicle detection technologies use ground shooting or electronic eyes to obtain images or videos. The images obtained in this way have limited coverage and more limitations, and are insufficient to provide information for vehicle detection, vehicle tracking, traffic flow monitoring, and road condition analysis. The remote sensing images of UAVs have a long range, a wide coverage area, and high execution efficiency, which can effectively solve the above problems^[1]-[3]. However, the traditional target detection algorithm does not have high target detection accuracy for remote sensing images, and UAV imaging can only obtain the top features of vehicles. Therefore, how to use the top features of vehicles for target detection is a scientific problem to be solved.