81results about How to "Preserve edge information" patented technology

The invention discloses an image denoising method based on multi-scale parallel CNNs, comprising five steps: 1, building a multi-scale parallel convolution neural network model, wherein only that convolution layer and the activation lay are included, and residual learning is added at the same time; 2, setting training parameter of a multi-scale parallel convolution neural network model; 3, selecting a training set and cutting and flipping the selected training image to enhance the number of the training sets; 4, selecting the mean square error as a loss function and train a multi-scale parallel convolution neural network model with a minimization loss function to obtain an image denoising model; 5, inputting the noise image of arbitrary size to the image denoising model, and outputting thedenoised clean image. The invention can preserve the edge information and the detail information of the image as much as possible while denoising, can improve the structural similarity of the image,and can obtain a high-quality denoised image.

The invention discloses a construction method and application of a three-dimensional MRI image denoising model based on a Wasserstein generative adversarial network. According to the invention, a Wasserstein generative adversarial network is used as a basic model to process an MRI noise image; large-scale data is used for training, a model can automatically learn potential association between a noise image and a non-noise image from the data, confrontation loss, perception loss and MSE loss are introduced into construction of a training model loss function, and the constructed model has a gooddenoising effect on a three-dimensional MRI image.

The invention discloses a DR image denoising method based on wavelet transformation. The method comprises the following steps: (1) reading image data to be processed; (2) determining wavelet and a number of layers of wavelet decomposition, and calculating a wavelet decomposition coefficient containing a noise signal; (3) choosing a threshold for each layer of coefficient, using an improved soft threshold to process an image; (4) carrying out wavelet reconstruction to obtain a first layer of A1, H1, V1 and D1; (5) carrying out classic hard threshold filtering on the first layer of H1, V1 and D1; and (6) using a coefficient after filtering to carry out reconstruction again to obtain a final image. A proposed method by the invention is more flexible, shortcomings of a hard threshold function and a soft threshold function can be overcome through a smoothing function and hard threshold filtering, noise can be removed, edge information of the image can be retained well, a signal-to-noise ratio of image is improved, and the method is a good processing scheme.

The invention discloses a weighted sparse-based mixed denoising method. According to the method, a calculus of variations is added on the basis of a weighted sparse representation non-local training dictionary to match non-local similar blocks, and a dual method is used to solve a mixed denoised image, so that edge information of the image can be better stored. The method has a denoising effect superior to that of the existing algorithms, has a high peak signal to noise ratio and a high image feature similarity, has a good inhibiting effect for mixed noises, particularly can well store the image information of the image, and has certain improvement for the preservation of image features.

The invention discloses an image defogging method based on dark channel and bright channel priori. The method comprises the following steps of establishing an atmospheric light scattering model basedon the physical model of an degraded image; taking an atmospheric light value as a global variable; performing prior estimation on an atmospheric light scattering model in combination with a dark channel and a bright channel to obtain an atmospheric light value; obtaining coarse transmissivity according to the atmospheric light scattering model and the obtained atmospheric light value to acquire more accurate data; then, using the gray level image of the original image as a guide image to carry out guide filtering to refine the transmissivity, thereby reserving the edge information of the depth of field, while reducing the time complexity of the algorithm, finally, acquiring the transmissivity through self-adaptive adjustment according to the self-adaptive transmissivity compensation function, sky region segmentation is not needed, the color distortion problem of a bright area is avoided, and the self-adaptability of the algorithm is improved. The problem of distortion of the color ofthe bright area during foggy image recovery is effectively solved, the defogging effect is natural, and the contrast ratio of the image is remarkably improved.

The invention belongs to the field of light field full-focus image fusion. According to a traditional image fusion algorithm based on guided filtering, only spatial information of an image is utilized; the method is applied to light field full-focus image fusion. The reasonable utilization of the angle information can effectively improve the image fusion precision; based on this, the invention provides a light field camera full-focus image fusion algorithm of spatial information guide angle information. According to the algorithm, a preliminary fusion weight graph is obtained based on angle information of a 4D light field; the preliminary fusion weight graph is taken as a guided image; spatial information obtained by coordinate transformation and integration of the 4D light field is used as a guide image to complete guide filtering; experiments demonstrate the effectiveness of the algorithm provided by the invention: a quantitative evaluation result shows that on the premise of not sacrificing image information richness and perception definition, a fused image obtained by the algorithm provided by the invention is higher in quality under performance evaluation indexes based on features and structural similarity.

The invention discloses an image denoising method based on a biological vision receptive field mechanism. Specifically, the method comprises the steps of noise judging, noise processing and result image brightness correcting. The method in the invention first makes a vision mechanism introduced into the image denoising research field. The noise position and size can be determined after a noise image passes through a receptive field template. The noise is suppressed in an adaptive way through a corresponding receptive field model according to determined noise. The brightness of a result graph is corrected according to the brightness of the noise image. The image acquired after the result graph brightness correction serves as a noise image which will undergo iteration processing until an optimum denoising effect is achieved. The image denoising method which can perform effective denoising can further make the image structure information well kept. A higher signal to noise ratio and great subjective effects can be achieved.

The invention provides an infrared and visible light image registration method for power equipment. The method comprises the following steps: respectively carrying out side window box filtering on aninfrared image and a visible light image of the power equipment; respectively extracting the edges of the two images by using an RCF network; extracting feature points of the edges of the two images respectively by adopting a KAZE algorithm, generating feature descriptors of the feature points of the edges of the two images respectively by adopting a directional filter based on an MPEG7 standard and a weight response matrix, and finally screening matching points through an Euclidean distance between the feature descriptors of the edges of the two images; and obtaining an optimal homography matrix H between the edge of the infrared image and the edge of the visible light image, and completing image registration. According to the method, the noise of the infrared images is reduced, meanwhile, the number of matching points between the two images is not obviously influenced, the edge detection of the two images is more accurate, the edge feature extraction mode is more stable, and the registration accuracy between the images is effectively ensured.

The invention relates to a hyperspectral ground object classification identification method. The method includes the following steps: carrying out atmospheric correction and geometric correction on remote-sensing images of known classification conditions in sequence to obtain remote-sensing images after correction; denoising the remote-sensing images after correction to obtain remote-sensing images after denoising; carrying out recombination and batch dividing on the remote-sensing images after denoising to obtain remote-sensing images after recombination; using a constructed multi-layer deepnetwork to carry out training on the remote-sensing images after recombination to extract deep-layer features of the remote-sensing images after recombination; and using an image, on which classification identification is to be carried out, to carry out training through the constructed multi-layer deep network to obtain a deep-layer feature of the image on which classification identification is tobe carried out, and comparing the deep-layer feature of the image, on which classification identification is to be carried out, with the deep-layer features, which are obtained by training, to obtaina final classification result of the image on which classification identification is to be carried out. According to the ground object classification identification method of the invention, data of the hyperspectral remote-sensing image can be quickly preprocessed, and ground object classification can be highly precisely carried out on the hyperspectral remote-sensing image.