Dark-Channel Soft-Constrained and Object-Perception-Enhanced Deep Dehazing Networks Used for Road Inspection Images

Abstract

Haze seriously affects the visual quality of road inspection images and contaminates the discrimination of key road objects, which thus hinders the execution of road inspection work. The basic assumptions of the classical dark-channel prior are not suitable for road images containing light-colored lane lines and vehicles, while typical deep dehazing networks lack physical model interpretability, and they focus on global dehazing effects, neglecting the preservation of object features. For this reason, this paper proposes a Dark-Channel Soft-Constrained and Object-Perception-Enhanced Deep Dehazing Network (DCSC-OPE-Net) for the information recovery of road inspection images. The network is divided into two modules: a dark-channel soft-constrained dehazing module and a near-view object-perception-enhanced module. Unlike the traditional dark-channel algorithms that impose strong constraints on dark pixels, a dark-channel soft-constrained loss function is constructed to ensure that the features of light-colored vehicles and lane lines are effectively maintained. To avoid resolution loss due to patch-based dark-channel processing for image dehazing, a resolution enhancement module is used to strengthen the contrast of the dehazed image. To autonomously perceive and enhance key road features to support road inspection, edge enhancement loss combined with a transmission map is embedded into the network to autonomously discover near-view objects and enhance their key features. The experiments utilize public datasets and real road inspection datasets to validate the performance of the proposed DCSC-OPE-Net compared with typical networks using dehazing evaluation metrics and road object recognition metrics. The experimental results demonstrate that the proposed DCSC-OPE-Net can obtain the best dehazing performance, with an NIQE score of 4.5 and a BRISQUE score of 18.67, and obtain the best road object recognition results (i.e., 83.67%) among the comparison methods.

Efficient Haze Removal from a Single Image Using a DCP-Based Lightweight U-Net Neural Network Model

“…AOD-Net [ 13 ] is characterized by the fact that it is based on a self-reconstructed atmospheric scattering model, unlike most dehazing models [ 14 , 15 ] that focus on the process of estimating transmission maps and atmospheric light based on physical models such as DCP and CAP. DCSC-OPE-Net [ 16 ] is a haze removal model aiming to improve visibility and object identification in complex road environments. The neural network is based on MSBDN [ 17 ] as the backbone, and it goes through a sub-neural network that performs coarse dehazing by adding color attention and a sub-neural network that extracts detailed features and then uses a DCP transmission map to supplement the details to output the final dehazing image.…”