Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (17): 223-231.DOI: 10.3778/j.issn.1002-8331.2204-0527

• Graphics and Image Processing • Previous Articles     Next Articles

Low Illumination Traffic Image Enhancement Model Based on Multi-Scale Dense Blocks

WANG Weihao, WANG Xiali, WU Lizhan, ZHANG Qian, LI Chao   

  1. School of Information Engineering, Chang’an University, Xi’an 710064, China
  • Online:2023-09-01 Published:2023-09-01

融合多尺度密集块的低照度交通图像增强模型

王炜昊,王夏黎,武历展,张倩,李超   

  1. 长安大学 信息工程学院,西安 710064

Abstract: Vehicle detection and tracking is an important content in the field of intelligent transportation. The complex illumination intensity and changeable traffic scene make the details of expressway images blurred, the contrast is low, and the image information extraction is highly disturbing. A generation countermeasure network based on attention mechanism and multi-scale residual dense blocks is proposed for low illumination traffic image enhancement. Different illumination images are synthesized by gamma correction, camera response function and manual adjustment as data sets, covering a wider range of brightness curves and simulating the real night scene. The attention mechanism is introduced to represent the correlation between different channels and high-frequency information. At the same time, maximum pooling and average pooling are used to capture the dependency between texture information and background information to enhance the overall integrity of the image. It builds a multi-scale fusion residual dense connection network. Deep extraction of complex image features uses parallel branches to fuse different ranks and gradations of information to improve the overall perception of details of the network and retain the consistency of picture information. Bilinear convolution structure is used to replace deconvolution layer to eliminate artifacts. The experimental results show that compared with the mainstream methods, the enhancement effect evaluation indexes PSNR and SSIM of the network in this paper are improved by 26.37% and 14.14% respectively. The image enhances the detailed texture and improves the natural visual effect of the image, which provides technical support for the visual task in the field of transportation.

Key words: low illumination image enhancement, traffic images, attention mechanism, residual dense blocks, multi-scale fusion, generative adversarial network

摘要: 车辆检测与跟踪是智能交通领域的重要内容,复杂的光照强度与多变的交通场景使高速公路拍摄图像细节模糊、对比度低,图片信息提取干扰性强。提出一种基于注意力机制融合多尺度残差稠密块的生成对抗网络用于低照度交通图像增强。通过伽马校正、相机响应函数和手工调节方法合成不同照度图像作为数据集,涵盖更广泛亮度曲线,模拟真实夜晚场景;引入注意力机制通过表征不同通道与高频信息间关联性,同时采用最大池化和平均池化捕获纹理信息和背景信息间依赖关系增强图像整体完整性;搭建多尺度融合的残差稠密连接网络,深度提取图像复杂特征利用并行支路融合不同级别和层次的信息,提升网络对细节的整体感知力,保留图片信息一致性;采用双线性加卷积结构代替反卷积层消除伪影现象。实验结果表明与主流方法相比,该网络的增强效果评价指标PSNR和SSIM分别提升26.37%、14.14%,图片增强使细节纹理清晰、图像自然视觉效果提高,为交通领域的视觉任务提供技术支持。

关键词: 低照度图像增强, 交通图像, 注意力机制, 残差稠密块, 多尺度融合, 生成对抗网络