Compound Multi-branch Feature Fusion for Real Image Restoration

Abstract

Image restoration is a challenging and ill-posed problem which also has been a long-standing issue. However, most of learning based restoration methods are proposed to target one degradation type which means they are lack of generalization. In this paper, we proposed a multi-branch restoration model inspired from the Human Visual System (i.e., Retinal Ganglion Cells) which can achieve multiple restoration tasks in a general framework. The experiments show that the proposed multi-branch architecture, called CMFNet, has competitive performance results on four datasets, including image dehazing, deraindrop, and deblurring, which are very common applications for autonomous cars. The source code and pretrained models of three restoration tasks are available at https: //github.com/FanChiMao/CMFNet.

Deep Learning-Based Technique for Remote Sensing Image Enhancement Using Multiscale Feature Fusion

“…Currently, the mean squared error ( MSE ) loss function [ 13 ] and the mean absolute error (MAE) loss function [ 28 ] are the predominant loss functions that measure the error between corresponding pixels in the field of computer vision. The MSE is susceptible to outliers, resulting in over-constraint [ 29 ], whereas MAE lacks gradient constraints [ 30 ], resulting in weak model convergence. Some studies have proposed a structural similarity index ( SSIM ) loss function based on human visual perception [ 31 ], which optimizes the model according to visual sensory direction; however, the illuminance and color restoration results are unsatisfactory.…”

“…In order to constrain the training process, accelerate the convergence speed of the model, and improve the visual quality of the enhanced image, we take into account the characteristics of the aforementioned loss functions. The weighted part of the proposed loss function (Fan et al 2022) [32] is removed, and that which consists of the quotient of PSNR and SSIM to predict the error between low-light images normal images is used, according to the following equation:…”

A multi-stage feature fusion defogging network based on the attention mechanism

“…From the subjective visual point of view, because of the limitations of the traditional algorithm of DCP based on the dark channel theory, there is a vast deviation in removing the dense fog from the image. It can be seen that the image (shown in Figure 7(b)) after the de-fogging by the DCP algorithm has a severe problem [11] 13.84 0.425 AOD [13] 13.14 0.414 GCANet [15] 14.25 0.497 CMFNet [37] 14.46 0.533 Ours 15.01 0.531 Even in outdoor scenarios with dense fog, the method proposed in this paper offers superior defogging capabilities. There is the least amount of fog compared to other advanced algorithms, the authentic image is visible, and the subjective visual advantage is evident.…”

“…The mean value of SSIM increases from the best of 0.971dB to 0.976dB; the mean value of PSNR increases from the best of 31.38dB to 32.96dB, and the relative increase of PSNR value is 1.58dB. [11] 20.64 0.800 AOD [13] 19.16 0.850 GCANet [15] 30.23 0.975 CMFNet [37] 31.38 0.971 Ours 32.96 0.976…”