Abstract: Image dehazing algorithms based on Transformer achieved good effect, but there were still problems such as large network parameters and slow dehazing speed. In order to prune the redundant parts of dehazing network directionally and shorten dehazing time without affecting dehazing quality, a Transformer image dehazing method based on component collaborative optimization pruning, CCOP-IDT (Component Collaborative Optimization Pruning Image Dehazing Transformer), was proposed. First, a 5-level Transformer was used to construct a pre-training model of dehazing network. Then, network pruning was modeled as an optimization problem. Fisher information was used to evaluate the importance of weight parameters, and Hessian matrix was used to measure the joint influence of pruning components on network output, so as to establish a collaborative optimization method for multiple pruning components. Finally, evolutionary algorithm was employed to solve the optimal pruning rate sequence, and the optimal sub-network of the pre-trained model was obtained. After pruning, the parameters of the network are controlled to 0.476MB, which reduces 28.8% compared with that before pruning, and the dehazing time is shortened by 25.0%. In the synthetic hazy dataset RESIDE-6K, the peak signal-to-noise ratio (PSNR) of the proposed method reaches 29.60dB, and the structural similarity (SSIM) reaches 0.9687, which reduces 1.63% and 0.46% compared with that before pruning, respectively. The quantitative and qualitative experimental results show that the proposed method can greatly reduce the model parameters and improve the image dehazing speed while basically maintaining the quantitative index and subjective perception.

Key words: image dehazing, Transformer, model pruning, Fisher information, Hessian matrix, evolutionary algorithm

摘要： 基于Transformer的图像去雾算法取得了较好去雾效果，但仍存在网络参数量大、去雾速度慢的问题。为实现对去雾网络冗余部分的定向修剪，从而在不影响去雾质量的前提下缩短去雾时间，提出一种基于组件协同优化剪枝的Transformer图像去雾方法——CCOP-IDT（Component Collaborative Optimization Pruning Image Dehazing Transformer）。首先，采用5级Transformer构建去雾网络预训练模型；其次，将网络剪枝建模为优化问题，使用费雪信息评估权重参数重要度，利用黑塞矩阵衡量剪枝组件对网络输出的联合影响，从而建立对多种剪枝组件的协同优化方法；最后，采用进化算法求解最优剪枝率序列，得到预训练模型的最优子网络。剪枝后网络参数量控制在0.476MB，相较于剪枝前减少了28.8%，去雾时间缩短了25.0%；在合成有雾数据集RESIDE-6K中，所提方法的峰值信噪比（PSNR）达到29.60dB，结构相似度（SSIM）达到0.9687，与剪枝前相比分别仅降低了1.63%和0.46%。定量和定性实验结果表明，所提方法能够在基本保持量化指标和主观观感的前提下，大幅减少模型参数量，提高图像去雾速度。

关键词: 图像去雾, Transformer, 模型剪枝, 费雪信息, 黑塞矩阵, 进化算法