Figure 1 presents some sample results for real-world hazy images processed by TGL-Net. By introducing transmission guidance, the proposed TGL-Net can achieve state-of-the-art (SOTA) performance with fewer parameters and smaller network dimensions, as well as faster training and processing speed.
Lightweight photo reader manual#
Instead of producing the transmission from predefined DOF maps, we apply a filter-based DCP (F-DCP) method to estimate transmission maps from input hazy images automatically, thereby avoiding additional manual calibration or the collection of transmission information. Notably, this paper presents a feasible solution for introducing priors obtained from non-learning- based image processing techniques as guidance for training DNNs. We take advantage of the effectiveness of DNNs and introduce the transmission map as a prior information to guide the efficient training of the network. In this work, a transmission-guided lightweight network (TGL-Net) for fast natural image dehazing is proposed. However, it is difficult to derive optimal network parameters without any guidance from prior knowledge because the features extracted by a network may not always be related to the degradation caused by haze. A well-trained DNN can perform dehazing and enhancing operations on real images with higher efficiency and superior visual effects than prior-based methods. Such methods include DehazeNet, all-in-one dehazing (AOD), Cycle-Dehaze (Cycle), proximal dehazing network (PDN), and grid dehazing network (GDN). To alleviate these issues, recent learning-based methods have attempted to train deep neural networks (DNNs) from a set of examples without formulating prior knowledge. Furthermore, the differences between introduced priors and real degradation processes often have negative effects on the final outputs, including insufficient or excessive dehazing, colour distortions, halos, and artifacts. However, it is typically very time consuming to estimate prior constraints for each input image. Traditional solutions commonly add various constraints to their optimization processes by including prior information, such as colour attenuation priors (CAPs), non-local priors, dark channel priors (DCPs), and scene-depth priors. Image dehazing is an ill-posed inverse problem that consists of computing a desired haze-free image J from an observed hazy image I, as well as estimating atmospheric light A and transmission map T using (1).
Extensive experimental results demonstrate that, in terms of several reference and non-reference evaluation criteria for real-world images, the proposed method can achieve state-of-the-art performance with a much smaller network size and with significant improvements in efficiency resulting from the training guidance.
The method provides a feasible solution for introducing priors obtained from traditional non-learning-based image processing techniques as a guide for training deep neural networks. A double-error loss function that combines the errors of a transmission map with the errors of a dehazed image is used to guide network training. Unlike most current dehazing methods that produce simulated transmission maps from depth data and haze-free images, in the proposed work, guided transmission maps are computed automatically using a filter-refined dark-channel-prior (F-DCP) method from real-world hazy images as a regularizer, which facilitates network training not only on synthetic data, but also on natural images. In this paper, a transmission-guided lightweight neural network called TGL-Net is proposed for efficient image dehazing.
0 kommentar(er)
