基于改进扩散模型的图像去雨方法

Research on image de-raining methodbased on improved diffusion model

针对图像去雨过度、泛化性差的问题,提出运用改进扩散模型进行单幅图像去雨的方法。通过前向过程添加高斯噪声使数据变为高斯分布,设计残差模块双输入信息通道、添加ECA(efficient channel attention)通道注意力机制模块以构建噪声估计网络,实现全局平均池化而不降低维数,从而捕获局部跨通道交互信息;利用模型网络进行反向采样,预测并剔除雨痕噪声,实现图像去雨。最后通过模拟雨滴数据集和Rain100数据集对改进的扩散模型与其他4种算法进行对比实验测试,实验结果表明改进的扩散模型能够有效去除雨痕,其中雨滴和雨线的峰值信噪比分别为30.328 5和34.896 5,结构相似性分别为0.927 1和0.962 0;自制真实雨图数据集,使用YOLOv7算法对去雨后的图像进行车辆检测,结果表明采用改进的扩散模型去雨能够有效提高车辆检测置信度,进一步验证了所提方法具有良好的去雨效果和泛化能力。

To address the excessive rain removal and poor generalization of images,this paper proposes a single-image de-raining method by improving diffusion model.The data becomes Gaussian distribution by adding Gaussian noise to the forward process.The dual input information channels of the residual module are designed and the ECA(Efficient Channel Attention)channel attention mechanism module is added to build a noise estimation network.Thus,a global average pooling without reducing the dimension is achieved and the local cross-channel interaction information is captured.The model network is employed to reverse sampling,predict the noise as a rain mark and remove it,and thus achieve image de-raining.By employing simulated raindrop datasets and the Rain100 dataset,comparative experiments are conducted to compare our improved diffusion model with other four algorithms.The experimental results demonstrate our improved diffusion model effectively removes rain streaks,with peak signal-to-noise ratios of 30.3285 for raindrops and 34.8965 for rain lines,and structural similarities of 0.9271 and 0.9620 respectively.A real rain image dataset is built,and the YOLOv7 algorithm is employed to perform vehicle detection on the rain-removed images.Our results show the improved diffusion model for rain removal effectively enhances the confidence of vehicle detection,further confirming it has outstanding de-raining performance and generalization capability.