基于改进CycleGAN模型的图像去雾算法

Image Dehazing Algorithm Based on Improved CycleGAN Model

图像去雾一直以来都是一个极具挑战的课题。本文基于循环生成对抗网络(cycle-consistent generative adversarial network, CycleGAN)进行改进。CycleGAN的生成器模块将引入基于泰勒展开的线性Transformer网络(MB-TaylorFormer),有效解决注意力机制的二次计算复杂度,同时使用不同大小卷积以及深度可分离卷积和可变形卷积结合进行多尺度标记,提高去雾能力。实验表明,算法改进后图像去雾任务性能有效提高。Image dehazing has always been a highly challenging subject. This paper presents improvements to the cycle-consistent generative adversarial network (CycleGAN). The generator module of CycleGAN incorporates an MB-Taylor Former network, which is a linear Transformer based on Taylor expansion, effectively addressing the quadratic computational complexity of the attention mechanism. Additionally, it uses convolutions of varying sizes in combination with depthwise separable convolutions and deformable convolutions to perform multi-scale marking, thereby enhancing dehazing capabilities. Experiments demonstrate that the performance of image dehazing tasks is effectively improved following the enhancement of the algorithm.