MCPNet: Multi-space color correction and features prior fusion for single-image dehazing in Non-homogeneous haze Scenarios
Zhiyu Lyu,Yan Chen,Yimin Hou
DOI: https://doi.org/10.1016/j.patcog.2024.110290
IF: 8
2024-01-30
Pattern Recognition
Abstract:When capturing images in non-homogeneous haze conditions, the degree of haze impact varies across the scene, and the information in certain parts of the scene are lost, making the scene nearly invisible. However, most existing dehazing convolutional neural networks (CNNs) are designed for homogeneous haze and do not consider the challenge of feature extraction arising from non-homogeneous haze distribution. Moreover, these networks primarily rely on the RGB color space for feature extraction, often fail to extract color and detail feature effectively, resulting in color distortion and loss of details in the output. To tackle this problem, we introduce image priors in the YCbCr and HSV color spaces, proposing a novel Multiple Color Space Prior Network (MCPNet) to enhance the dehazing performance specifically for non-homogeneous hazy images, while simultaneously correcting the color, preserving the visual quality of the output. Leveraging image priors, we designed two parallel sub-networks to extract color and detail features from the YCbCr and HSV color spaces. Moreover, to capitalize on these features and incorporate them effectively into the dehazed image, we introduce a Comprehensive Fusion Module (CFM). This module judiciously takes into account both the fusion of multiscale features and the interrelation among channels to optimize feature fusion. By employing a dual network architecture coupled with the CFM, our model proficiently amalgamates and exploits the mined features, accurately restoring the color and detail information of the image, especially for images containing non-homogeneous haze. Extensive experimental highlight the effectiveness of our model in addressing homogeneous and non-homogeneous hazy images, concurrently preserving the visual appeal of the dehazed outcomes. When compared with other SOTA models, our MCPNet demonstrates superior results in dehazing.
computer science, artificial intelligence,engineering, electrical & electronic