光学 精密工程, 2019, 27 (12): 2702, 网络出版: 2020-05-12   

混合残差学习与导向滤波算法在图像去雾中的应用

Application of hybrid residual learning and guided filtering algorithm in image defogging
作者单位
西安建筑科技大学 理学院, 陕西 西安 710055
摘要
为解决雾天场景图像恢复过程中图像清晰度和对比度下降的问题, 提出了一种结合残差学习和导向滤波的单幅图像去雾算法。使用雾天图像与对应的清晰图像构建残差网络; 采用多尺度卷积提取更多细节的雾霾特征; 利用导向滤波各向异性的优点, 对残差网络去雾后的图像进行滤波以保持图像边缘特性, 得到更加清晰的无雾图像。通过与DCP算法、CAP算法、SRCNN算法、DehazeNet算法和MSCNN算法相比, 在合成雾天图像上, 峰值信噪比值最高达到31.951 8 dB, 结构相似度值最高达到0.979 6, 在自然雾天图像上的运行时间最低达到了0.4 s, 主观评价和客观评价均优于其它对比算法。实验结果表明, 所提去雾算法不仅去雾效果较优, 而且速度较快, 具有较强的实用价值。
Abstract
To solve the problem of image clarity and contrast degradation in fog scene image restoration, a single image defogging algorithm based on residual learning and guided filtering was proposed. A residual network was first constructed by using foggy images and corresponding clear images. Multi-scale convolution was then used to extract more detailed haze features. Taking advantage of the anisotropy of the guided filter, the algorithm then obtained a clearer fog-free image after the residual network was filtered to maintain image edge characteristics. Experiments produced the following results as compared with the dark channel prior, CAP, super-resolution convolutionalneuralnetwork, DehazeNet, and multi-scale convolutional neural network algorithms.On synthetic foggy images, the peak signal-to-noise ratio reacheda maximum of 27.840 3 dB, the structural similarity index measurereacheda maximum of 0.979 6, and the runtime on natural foggy images was as low as 0.4 s.In addition, the subjective and objective evaluations proved to be better than those of the other comparison algorithms.Thus, the proposed defogging algorithm not only produces a better defogging effectbut is also faster, there by offering a greater practical valuefor defogging applications than the other algorithms.

陈清江, 张雪. 混合残差学习与导向滤波算法在图像去雾中的应用[J]. 光学 精密工程, 2019, 27(12): 2702. CHEN Qing-jiang, ZHANG Xue. Application of hybrid residual learning and guided filtering algorithm in image defogging[J]. Optics and Precision Engineering, 2019, 27(12): 2702.

本文已被 10 篇论文引用
被引统计数据来源于中国光学期刊网
引用该论文: TXT   |   EndNote

相关论文

加载中...

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!