一种融合多尺度与局部特征增强的实时图像去雾算法
A Real-time Image Dehazing Network Fusing Multi-scale and Local Feature Enhancement
-
摘要: 图像去雾是计算机视觉领域一个经典且具有挑战性的任务, 常应用于自动驾驶、户外监控等易受外界环境干扰的场景, 作为一种预处理手段保证下游视觉系统功效的正常发挥. 为了实现对雾天图像的实时处理, 提出一种融合多尺度与局部特征增强的图像去雾算法. 首先设计了一个轻量化特征提取模块, 在提取图像局部特征的同时减少网络计算量, 提升模型运行效率; 此外考虑到不同尺度特征包含的语义信息可以帮助模型获取更具辨别力的特征表示, 进而准确地定位雾区范围并对其有针对性地进行处理, 提出一种多尺度特征融合模块; 最后引入增强解码策略, 将上一层解码特征与多尺度特征进行融合, 使网络能够自适应地关注雾霾区域, 增强模型的去雾性能. 在合成和真实数据集上的实验结果表明, 与现有主流的图像去雾方法相比, 所提算法在去雾性能和推理速度上均表现优异, 在SOTS室内和室外数据集上PSNR和SSIM指标分别为32.32 dB/0.979和33.39 dB/0.982.Abstract: Image dehazing is a classic and challenging task in the field of computer vision, which is often applied in automatic driving, outdoor monitoring, and other scenarios susceptible to environmental interference, and is often used as a pre-processing manner to ensure the normal play of the efficacy of the downstream vision systems. To realize real-time processing of hazy images, an image dehazing algorithm fusing multi-scale and local feature enhancement is proposed in this work. Firstly, a lightweight feature extraction module is designed to reduce the network computation while extracting local features to improve the model operation efficiency; Additionally, considering the semantic information contained in different scale features, which can help the model to obtain more discriminative feature representations, and then accurately locate the extent of the haze area and process it in a targeted way, a multi-scale fusion module is proposed; Finally, an enhanced decoding strategy is introduced to fuse the upper layer decoding features with the multi-scale features so that the network can adaptively focus on the haze region and boost the dehazing performance of the model. Experiments on synthetic and real-world datasets demonstrate that the proposed algorithm excels in both dehazing performance and inference speed compared to existing mainstream image dehazing methods, with PSNR and SSIM metrics of 32.32 dB/0.979 and 33.39 dB/0.982 on SOTS indoor and outdoor datasets, respectively.