摘要：

经典AOD-Net(All in One Dehazing Network)去雾后的图像存在细节清晰度不足、明暗反差过大和画面昏暗等问题。为了解决这些图像去雾问题，提出一种在AOD-Net基础上改进的多尺度算法。改进的网络结构采用深度可分离卷积替换传统卷积方式，减少了冗余参数量，加快了计算速度并有效地减少了模型的内存占用量，从而提高了算法去雾效率；同时采用多尺度结构在不同尺度上对雾图进行分析和处理，更好地捕捉图像的细节信息，提升了网络对图像细节的处理能力，解决了原算法去雾时存在的细节模糊问题；最后在网络结构中加入金字塔池化模块，用于整合图像不同区域的上下文信息，扩展了网络的感知范围，从而提高网络模型获取有雾图像全局信息的能力，进而改善图像色调失真、细节丢失等问题。此外，引入一个低照度增强模块，通过明确预测噪声实现去噪的目标，从而恢复曝光不足的图像。在低光去雾图像中，峰值信噪比(PSNR)和结构相似性(SSIM)指标均有显著提升，处理后的图片具有更高的整体自然度。实验结果表明：与经典AOD-Net去雾的结果相比，改进算法能够更好地恢复图像的细节和结构，使得去雾后的图像更自然，饱和度和对比度也更加平衡；在RESIDE的SOTS数据集中的室外和室内场景，相较于经典AOD-Net，改进算法的PSNR分别提升了4.559 3 dB和4.065 6 dB，SSIM分别提升了0.047 6和0.087 4。

关键词: 多尺度网络结构, 深度可分离卷积, 金字塔池化模块, 低照度增强模块, 图像去雾

Abstract:

This study addresses the problems of detail blurring, excessive contrast, and darkness in dehazed images generated using the classical All in One Dehazing Network (AOD-Net). The study proposes a novel multiscale image dehazing algorithm, which builds upon the improvements made to AOD-Net. In the enhanced network architecture, traditional convolutions are replaced with depth-wise separable convolutions to reduce redundant parameters for analyzing and processing foggy images at different scales, thereby better capturing image details to accelerate the computation speed, effectively reduce the memory footprint of the model, and enhance the algorithm′s dehazing efficiency. In addition, this study employs a multiscale structure to improve the network′s capability to handle image details and mitigate the blurring of details in dehazed images. Furthermore, a pyramid pooling module is incorporated into the network architecture to aggregate contextual information from different image regions, thereby enhancing the network′s ability to capture global information in hazy images and mitigate problems such as color tone distortion and detail loss. Additionally, a low-light enhancement module selectively enhances the predicted noise and improves the contrast stretch, effectively restoring noisy regions. Consequently, moderate improvements are observed in terms of the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) metrics for low-light dehazed images, concurrently achieving elevated levels of overall naturalness. Experimental results demonstrate that the proposed algorithm yields satisfactory dehazing outcomes. The processed images exhibit greater naturalness and improved balance in saturation and contrast compared with the classical AOD-Net. Finally, on the SOTS subset of the RESIDE dataset, which includes both outdoor and indoor sets, the proposed algorithm achieves improvements of 4.559 3 dB and 4.065 6 dB, respectively, in terms of PSNR over the classical AOD-Net. Furthermore, it achieves improvements of 0.047 6 and 0.087 4 in terms of SSIM.

Key words: multi-scale network structure, depth-wise separable convolution, pyramid pooling module, low-light enhancement module, image dehazing