Search

The color image denoising based on deep learning usually uses convolution on each channel, and then merges multi-channel data into single channel data. This method does not fully consider the spectral correlation between color channels, which may casuse distortion of the denoising results. Quaternion convolution can solve this problem by treating a color pixel as a whole. However, a single quaternion convolutional network can not restore the image details well. To solve the problem, a dual-channel quaternion convolutional network (DQNet) for color random impulse noise removal is proposed. Firstly, according to the strategy of structure channel and color channel fusion, a structure detail restoration network based on dilated convolution is proposed to obtain structure and edge features, and quaternion convolution network is used to extract cross-channel color information. Secondly, aiming at the problem that convolution operation will cause partial global information loss, the long line connection is used to fuse the input noise image with the convolution results, and then, a feature enhancement module based on attention mechanism is designed to guide the network to extract potential noise features from complex background. Finally, the residual learning is used to achieve the restoration of color random impulse noise. Experimental results show that the proposed algorithm has better denoising performance, especially in moderate noise level or high noise level.

The denoising method based on deep learning can achieve better denoising effect than the traditional method, but the existing deep learning denoising methods often have the problem of excessive computational complexity caused by too deep network. To solve this problem, a progressive residual fusion dense network (DnRFD) is proposed to remove Gaussian noise. Firstly, dense blocks are used to learn the noise distribution in the image, and the network parameters are greatly reduced while the local features of the image are fully extracted. Then, a progressive strategy is used to connect the shallow convolution features with the deep features to form a residual fusion network to extract more global features for noise. Finally, the output characteristic images of each dense block are fused and input to the reconstructed output layer to get the final output result. Experimental results show that, when the Gaussian white noise level is 25 and 50, the network can achieve higher mean PSNR and mean structural similarity, and the average time of denoising is half of the DnCNN method and one third of the FFDNet method. In general, the overall denoising performance of the network is better than that of the correlative comparison algorithms, and it can effectively remove the white Gaussian noise and natural noise in the image, and can restore the edge and texture details of the image better.

The traditional rail surface area extraction method has some preconditions, for instance, the width of rail surface should be given in advance, the rail surface should be assumed to be in the center of the rail image and the boundary should be selected manually. Besides, it also has some problems, such as poor self-adaptability, light sensitivity, and the inability to extract the rail surface completely when there are noises such as dust and mud at the rounded corners of the rail head. Aiming at the above problems, a method of extracting rail surface area based on YUV space greedy algorithm selection and slope detection expansion is proposed. Firstly, the RGB rail image is converted to YUV space, and its V component is extracted to reduce the interference of ambient light and noise. Secondly, the grayscale projection inversion curve of the V component is drawn, and the gray mean and median of the curve are used to divide the candidate orbital intervals. Then, greedy algorithm is used to calculate the interval of the maximum suborder sum in the divided curve for rough extraction of rail surface. At last, the slope detection expansion method is used to accurately extract the rail surface, the slope detection at a certain distance is carried out on both sides of the coarse-extracted boundary, and the rail surface boundary is updated with the middle position where the deflection angle is greater than the set threshold. Experimental results show that the proposed method can accurately and rapidly extract the rail surface area, with an average precision of 0.9296, an average accuracy of 96.67%, and an average time of 25.96 ms, which is of certain practical value.

Volumetric medical image registration has important clinical significance. Traditional registration methods may be time-consuming when processing large volumetric data due to their iterative optimizations. In contrast, existing deep learning-based networks can obtain the registration quickly. However, most of them require independent rigid alignment before deformable registration; these two steps are often performed separately and cannot be end-to-end. Moreover, registration ground-truth is difficult to obtain for supervised learning methods. To tackle the above issues, we propose an unsupervised 3D end-to-end deformable registration network. The proposed network cascades two subnetworks; the first one is for obtaining affine alignment, and the second one is a deformable subnetwork for achieving the non-rigid registration. The parameters of the two subnetworks are shared. The global and local similarity measures are used as loss functions for the two subnetworks, respectively. The trained network can perform end-to-end deformable registration. We conducted experiments on brain MRI datasets (LPBA40, Mindboggle101, and IXI) and experimental results demonstrate the efficacy of the proposed registration network.