Guided-Pix2Pix+: End-to-end spatial and color refinement network for image dehazing

Abstract

Inevitable hazy contamination degrades the visibility of images, and the resulting haze removal is one of the essential prerequisites for image processing and computer vision tasks. We proposed an end-to-end dehazing network, referred to as Guided-Pix2Pix, to estimate spatially refined transmission maps and dehaze via the physical scattering equation. The remaining enhancement of color contrast, ill-posed adversarial training, and redundant backbone, however, should be thoroughly investigated, as required in the prospect of Guided-Pix2Pix. In this paper, we inherit the end-to-end structure of Guided-Pix2Pix and accordingly propose Guided-Pix2Pix+ as an update, which concatenates transmission estimation and refinement, physical scattering equation-based dehazing, together with color refinement, to achieve haze removal in a one-stage way. Specifically, we make use of the pretrained instance of EfficientNetB0 to estimate coarse-grained transmission maps and concatenate a guided filter layer to perform spatial refinement for the incoming transmission maps. Restored by the physical scattering equation, color refinement of dehazed proposals is finally performed via the standardization and clipping of pixel intensities. All the operations are differentiable, making it possible to achieve end-to-end, tight training. Furthermore, adversarial and perceptual losses are employed to regulate the performance of our model, giving rise to structurally similar but photo-realistic dehazed proposals. Extensive experiments confirm that our Guided-Pix2Pix+ yields dehazed proposals with fine-grained spatial refinement and relatively effective color contrast, compared to our previous Guided-Pix2Pix, the baseline, and advanced dehazing methods. The source code is currently available at https://github.com/92xianshen/guided-pix2pixplus.