计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (3): 237-245.DOI: 10.3778/j.issn.1002-8331.2209-0112

• 图形图像处理 • 上一篇    下一篇

自适应Transformer网络下的单幅图像去雾方法

金海波,马琳琳,田桂源   

  1. 辽宁工程技术大学 软件学院,辽宁 葫芦岛 125105
  • 出版日期:2024-02-01 发布日期:2024-02-01

Single Image Defogging Method Under Adaptive Transformer Network

JIN Haibo, MA Linlin, TIAN Guiyuan   

  1. School of Software, Liaoning Technical University, Huludao, Liaoning 125105, China
  • Online:2024-02-01 Published:2024-02-01

摘要: 针对现有图像去雾方法对不同雾浓度的雾处理不完善、颜色失真、细节恢复不佳等问题,提出一种基于自适应Transformer的单图像去雾方法。该方法包括编码器、自适应Transformer、解码器三个模块。编码器将输入的图像进行局部特征提取,得到细节特征。为更好地对不同雾浓度进行提取,提出自适应Transformer,自适应Transformer可以在提取全局特征的基础上,自适应提取不同尺度颗粒雾度特征。解码器用于图像进行恢复与重建。该方法在SOTS室外测试集进行实验,其PSNR为27.45?dB,SSIM为0.954?6,与对比方法中最高的相比分别提高0.43?dB和0.005?3,从而验证出该方法取得了较好的客观结果。此外主观方面去雾图像有效改善了颜色失真,以及不同雾浓度处理不理想的问题。

关键词: 图像去雾, Transformer, 特征提取, 注意力机制

Abstract: Aiming at the problems of imperfect haze processing, color distortion, and poor detail recovery of existing image dehazing methods with different haze concentrations, a single image dehazing method based on adaptive Transformer is proposed. The method includes three modules:encoder, adaptive Transformer, and decoder. The encoder performs local feature extraction on the input image to obtain detailed features. In order to better extract different haze concentrations, an adaptive Transformer is proposed. The adaptive Transformer can adaptively extract the haze features of different scale particles on the basis of extracting global features. The decoder is used for image restoration and reconstruction. The method is tested on the SOTS outdoor test set, and its PSNR is 27.45 dB and SSIM is 0.954 6, which are respectively 0.43?dB and 0.005?3 higher than that in the comparison method, which verifies that the method achieves good objective results. In addition, the subjective aspect of the dehazing image can effectively improve the color distortion and the problem of unsatisfactory handling of different fog concentrations.

Key words: image dehazing, Transformer, feature extraction, attention mechanism