Multi-Scale Image Dehazing Algorithm with Attention Mechanism
Xiaoxue Li, Yue Hou · 2024
This paper introduces a novel large convolutional kernel parallel defogging network, meticulously crafted to overcome the limitations of current image defogging algorithms, which often struggle with image artifacts, blurring of fine details, and color distortion. The proposed network is engineered to significantly elevate the performance of existing methodologies by incorporating larger convolutional kernels. The architecture is anchored by two pivotal components: the first is the Dynamic Convolution Kernel Layer (DCKL) module, which harnesses the power of three distinct convolutional kernels of sizes 3x3, 5x5, and 7x7. These kernels operate in parallel to capture feature information across various scales, adeptly merging global and local features to extract comprehensive spatial information from the input image. The second component integrates a multilayer perceptron with a sophisticated multidimensional dual-channel attention mechanism, designed to refine and amplify local features, alongside channel significance, spatial information, and the nonlinear interrelations among features. Through extensive experimentation on the RESIDE dataset, our model demonstrates a marked superiority over traditional defogging techniques, achieving a notable reduction in computational complexity and parameters, while delivering images with enhanced sharpness and color accuracy, suitable for both indoor and outdoor environments.