%0 Journal Article
%T 基于多尺度特征提取与融合的单幅图像去雾算法<br>A Single Image Dehazing Algorithm Based on Multi-Scale Feature Extraction and Fusion
%A 李金函
%A 魏伟波
%A 王博
%J Journal of Image and Signal Processing
%P 117-129
%@ 2325-6745
%D 2024
%I Hans Publishing
%R 10.12677/jisp.2024.132011
%X 为解决随着CNN网络层数加深而导致的学习成本过高或过拟合问题，提出了一种基于多尺度特征提取与融合的单幅图像去雾算法。该算法结合U-Net思想，对输入图像进行物理分割和下采样得到多个尺度的特征图，采用残差连接的方式进行多维度融合，可以更好的适配大尺度数据集。同时，在网络中加入了深度监督模块，引入额外的监督信号有助于梯度传播，加快收敛速度，保证了训练的稳定性，这种多任务的学习形式提高了网络对不同输入的适应性，可以增强去雾效果。此外，使用自带多维度天气系统渲染的3D游戏引擎，自建了一份大尺度全高清数据集，模型训练的鲁棒性和泛化能力得到显著提升。实验结果表明，所提算法在训练速度和模型大小控制上具有一定优势，在主观评价上，远景去雾效果明显，峰值信噪比(Peak Signal-to-Noise Ratio, PSNR)和结构相似性(Structure Similarity, SSIM)两个客观评价指标分别为26.75 dB和0.907，相较于对比算法中性能第二的模型分别提高了3.5和5.9个百分点，加入自建数据集进行组合训练后进一步提升了模型的去雾性能。<br />
To solve the problem of high learning cost or overfitting caused by the deepening of CNN network layers, a single image dehazing algorithm based on multi-scale feature extraction and fusion is proposed. This algorithm combines the U-Net idea to physically segment and down-sampling the input image to obtain multi-scale feature maps. It uses residual connections for multi-dimensional fusion, which can better adapt to large-scale datasets. At the same time, a deep supervision module has been added to the network, introducing additional supervision signals to facilitate gradient propagation, accelerate convergence speed, and ensure training stability. This multi-task learning form improves the network’s adaptability to different inputs and can enhance the dehazing effect. In addition, a 3D game engine with a built-in multi-dimensional weather system rendering was used, and a large-scale high-definition dataset was built. The robustness and generalization ability of the model training was significantly improved. The experimental results show that the proposed algorithm has certain advantages in training speed and model size control. In terms of subjective evaluation, the long-range dehazing effect is obvious. The objective evaluation indicators of Peak Signal-to-Noise Ratio (PSNR) and Structure Similarity (SSIM) are 26.75 dB and 0.907, respectively, which are 3.5 and 5.9 percentage points higher than the second-best-performing model in the comparison algorithm. The addition of a self-built dataset for combined training further improves the model’s dehazing performance.
%K 单幅图像去雾，多尺度特征融合，U形网络，深度监督，自建数据集<br>Single Image Dehazing
%K Multi-Scale Feature Fusion
%K U-Net
%K Deep Supervision
%K Self-Built Dataset
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=84103