39results about How to "Good image restoration" patented technology

The invention discloses a two-channel convolutional neural network-based single image super-resolution calculation method. The method comprises the steps of (1) performing fuzzy degradation processing on a known high-resolution image to obtain a low-resolution image with the same size; (2) decomposing the obtained low-resolution image after the fuzzy processing in the step (1) into a texture part and a smooth structure part of the low-resolution image, and obtaining a texture part and a smooth structure part of the high-resolution image; (3) combining the low-resolution texture part obtained in the step (2) and an original low-resolution image to obtain a two-channel input, and obtaining an output of the high-resolution texture part; (4) combining the obtained output of the high-resolution texture part in the step (3) and the original low-resolution image to obtain a final image super-resolution reconstruction result, thereby finishing super-resolution reconstruction; and (5) calculating a difference value between the high-resolution texture parts obtained in the steps (4) and (2) to obtain texture part loss, and minimizing a sum of the texture loss and image loss to optimize network structure parameters.

The invention belongs to the field of video image enhancement, and particularly relates to a video image sea fog removal and clearing method which integrates frame difference method background estimation with the rapid single-frame sea fog removal algorithm based on edge detection and is used for an offshore aircraft rapid video image sea fog removal and clearing system. The video image sea fog removal and clearing method includes the steps of obtaining a sea fog video image, conducting sea fog removal and clearing on a single-frame sea fog image, and conducting sea fog removal and clearing on the video image. The video image sea fog removal and clearing method is suitable for all offshore aircrafts, and performance of visual systems of the offshore aircrafts in sea fog can be greatly improved. The operating speed is high, and sea fog removal and clearing can be conducted on the video image in real time in the sea surface scene. Compared with other algorithms, the rapid single-frame sea fog removal algorithm has a good edge keeping effect. The method has the advantages of being remarkable in fog removal effect and good in image restoration effect. The detecting performance, the tracking performance and the recognizing performance of targets in the later period can be effectively improved with sea fog removal and clearing as earlier stage processing on the visual systems.

The invention discloses an image restoration method based on an adaptive residual neural network. The method comprises the steps that an adaptive residual neural network model is built, wherein the adaptive residual neural network comprises a number of adaptive residual units connected in series; training sets for image denoising, image super-resolution and image deblocking effects are respectively selected, and corresponding training parameters are respectively set; according to the adaptive residual neural network model and the training parameters for image denoising, image super-resolution and image deblocking effects, the corresponding target neural network model is trained respectively with the goal of minimizing a loss function; and according to the trained target neural network model for the problem of image denoising, image super-resolution and image deblocking effects, an image to be processed is input to the corresponding target neural network model, and a corresponding high-quality image is output. According to the invention, the PSNR, SSIM and visual effect of the image can be remarkably improved, and the method has the advantages of good recovery effect, high speed and strong robustness.

The invention relates to a blind convolutional motion fuzzy image restoration method. The method comprises steps that S1, an observation image is acquired, derivative filtering of the observation image is carried out through utilizing a differential filter, a high frequency degeneration image is generated; S2, an intelligible image is updated; S3, a formula described as specifications is solved through utilizing an iteration least squares, and a point spread function is updated and solved; S4, if a dimension of the point spread function is greater than a set value, the dimension of the point diffusion function is enhanced, the process returns to the step S2, otherwise, iteration stops; and S5, according to the optimal point spread function acquired in the step S4 and the observation image, so a restoration image is solved through an original non-blind convolutional method. The method is advantaged in that the point spread function acquired through solution is stable, influence of noise on the restoration image is inhibited, and substantial fuzzy recovery effect, high automation degree, simple operation and rapid performing speed are realized.

The invention discloses a panoramic image vignetting phenomenon compensation method. The method comprises the steps: obtaining a panoramic image, obtaining the external contour of the panoramic image, obtaining the center and radius of the circle of the external contour, carrying out the iterative estimation of a circumferential region of the obtained panoramic image, obtaining an estimation parameter corresponding to a minimum image entropy, correcting an existing image according to the parameter obtained through iterative estimation, and obtaining a finally compensated panoramic image. The invention specifically discloses methods for iterative estimation, minimum image entropy obtaining and parameter correcting. The method does not need to build a special scene for obtaining a compensation parameter. The method autonomously optimizes the compensation parameter through employing an autonomous compensation method, estimates a compensation coefficient at an initial stage for a specific region, carries the correction of the subsequently obtained image through the compensation coefficient, and can obtains a better image restoration effect quickly and accurately.

The invention discloses a method for generating an image de-marking of an antagonistic neural network, which comprises the following steps: S10 constructing a training data set and a test data set; S20 constructs a generator network for generating an image without a stage mark according to the input stage mark image and the stage mark mask image; S30, a discriminator network is constructed and connected with the output end of the generator network, and the discriminator network is used for judging the true image without bench mark and the image without bench mark output by the generator network; S40 trains the generator network and the discriminator network according to the training data set; S50 uses the trained generator network to de-label the test data set. The effect of the generatornetwork is obviously better than the traditional algorithm. In most scenes, the residual traces of the marker can not be seen. The image restoration degree is good, and the image distortion is not easy to occur in the restoration area.

The invention discloses an image enhancement method based on contrast and structural similarity. The method comprises steps: an original single haze image is inputted; robust estimation on atmospheric light is realized through iterative quadtree segmentation; content-based superpixel segmentation technology is adopted for segmenting the image into local area blocks; a cost function is built, an atmospheric propagation chart of each local area block is estimated, and the optimal transmission parameter for each block is obtained; and an atmospheric degradation model is used for image restoration. The de-haze method provided by the invention can effectively improve visibility of a recovery scene image, has the advantages of fast operation speed and good image recovery effects, and can be applied to a system with high real-time performance requirements.

The invention provides a method for removing motion blur of a single image, which comprises the following steps: S01, calculating the direction of a blur kernel, and calculating the direction of the blur kernel by using a threshold method on the basis of using Radon transform according to the characteristics of the image in a frequency domain, so that the method can accurately obtain the directionof the blur kernel; and S02, calculating the length of the blurred kernel, calculating the length of the blurred kernel according to the direction of the blurred kernel and the information of the input image, calculating the blurred kernel according to the direction of the blurred kernel and the length of the blurred kernel, and restoring the picture according to the blurred kernel and the original picture information. Through the threshold method, the direction with interference is eliminated, so that the wrong direction is prevented from being found, and the accuracy of the obtained direction is ensured.

The invention discloses an impulse noise inhibition method based on recursion Gauss maximum likelihood estimation of confidence similarity. The method includes the steps that firstly, assuming that, pixels with the gray value of 0 and 255 are pixels polluted by noise, a mask image is obtained, and noise density is calculated; secondly, the restored value of each pixel is determined in a cyclic mode, if a noise point exists, a weighed estimated value is assigned to a target restored image, or else, a current pixel value is assigned to the target restored image, a current pixel is calculated to be a window weight matrix, the estimated value of the current pixel is calculated through the Gauss maximum likelihood estimation, the gray values of pixels which are not polluted by noise in the image are calculated again in each time of iteration, the peak signal to noise ratio of the gray values of the pixels to the gray values of pixels at corresponding positions in a noise image is calculated, and if the peak signal to noise ratio is not increased any more, iteration is stopped. According to the impulse noise inhibition method, impulse noise is effectively inhibited, meanwhile, local details are stored so that a local structure has the better contrast ratio, and the better image restoring effect is achieved.

The invention belongs to the field of image processing, and discloses an underwater image data set construction method and system, and the method comprises the steps: constructing underwater fuzzy image data sets of various types, and building a background light data set based on a manual marking method; meanwhile, evaluzting the underwater image quality based on four-dimensional area evaluation indexes of image chromaticity, contrast ratio, gradient and sharpness, so that constructing an underwater high-definition image data set. Based on human subjective discrimination, the accuracy of output underwater background light can be ensured; the constructed underwater blurred image data set and the background light data set thereof can be used as a data source and background light estimation reference in underwater image restoration research; the underwater high-definition image data set construction method is simple and rapid, the constructed underwater image data set can provide trainingsamples for an underwater image quality enhancement algorithm based on deep learning, and the underwater image quality evaluation method can be used for underwater image quality evaluation.

The invention provides a ciphertext domain image repair method which is used for carrying out repair on a damaged image generated by suffering from damage of a mask image in a ciphertext domain and ischaracterized by comprising the following steps of: S1, acquiring a to-be-repaired encryption image EP, an auxiliary encryption image EJ and a mask image theta; S2, carrying out marking on the to-be-repaired encryption image EP by the mask image theta to obtain a damaged region omega and an undamaged region phi; S3, calculating priorities of to-be-repaired pixels c in the damaged region omega bytaking a block as the unit so as to obtain a target block (with the reference to the specification) with the highest priority; S4, under the assistance of the auxiliary encryption image EJ, carrying out block matching on the target block (with the reference to the specification) and all candidate blocks BS in the undamaged region phi to obtain an optimal matched block (with the reference to the specification); S5, carrying out repair on the target block (with the reference to the specification) by the optimal matched block (with the reference to the specification) to select an initial repairedencryption image EP1; S6, respectively carrying out updating on the auxiliary encryption image EJ, the damaged omega and the undamaged region phi; and S7, repeating the steps S2 to S6, until the damaged region omega is completely repaired to obtain a repaired encryption image E'P.

The invention discloses an image enhancement and fusion method applied to a CMOS image sensor. Particularly, for image data output by the CMOS image sensor, firstly, an image is subjected to segmentation processing to obtain two images, the two new images are subjected to image quality enhancement processing to enable image contrast to be modulated, and image denoising processing is performed forreducing noises caused by previous image processing to meet the demands of subsequent processing on the image quality; secondly, the two images are subjected to wavelet decomposition; and finally, theimages subjected to the wavelet processing are subjected to fusion and invert wavelet transform processing to obtain a final output image after restoration. According to the image enhancement and fusion method applied to the CMOS image sensor, the problem of image distortion caused by the noises is solved; the image quality and the image transparency are enhanced; the image edge details are kept;and real-time processing is realized in finite platform resources.

The invention discloses a small optical lens. The optical lens comprises a lens cone. A first lens, a second lens, a third lens, a fourth lens, a fifth lens and an optical filter are sequentially arranged in a cavity of the lens cone from object space to image space, wherein the first lens is a crescent-shaped spherical glass lens with negative power; the second lens is a crescent-shaped aspherical plastic lens with negative power; the third lens is a crescent-shaped aspherical plastic lens with positive power, and the concave surface of the third lens faces towards the image space; the fourth lens is a drum-shaped aspherical plastic lens with positive power; the fifth lens is a crescent-shaped aspherical plastic plus lens, and the concave surface of the fifth lens faces towards the object space. The small optical lens is ultra-wide in field and small in size.

The embodiment of the invention discloses an image restoration method and device, electronic equipment, a computer program and a storage medium. The method comprises: constructing a prior model according to the characteristic map of the current image to be processed, wherein the prior model is used for learning the information of the current image to be processed, the current image to be processedincludes at least one image degradation phenomenon; and performing image restoration processing for the image degradation phenomenon of the current image to be processed based on the prior model so as to obtain the target image. The great image restoration result can be acquired.

The invention belongs to the filed of digital image processing and particularly relates to an HR (restrict restriction)-LLE (locally linear embedding) weight constraint based face image super-resolution restoration method. The restoration method includes: calculating a great many of HR face samples and average restoration weight constraint, relative to neighboring samples, of residual HR faces; during restoration, subjecting traditional LLE based face super-resolution restoration weight calculation method to weight constraint. The restoration method includes overall restoration and local detail compensation. The overall restoration aims to restore basic characteristics of a standard face as required, and the local detail compensation aims to restore the face image so as to enable the face to be provided with personality characteristics different from other faces. The HR-LLE weight constraint is added in the method in estimating the LLE based restoration weight of the target HR image, so that the weight is closer to the true restoration weight of the HR image in 12 norms. By the method, better image restoration results can be acquired.

The invention discloses an image shadow removal method based on a bidirectional mapping network, and the method comprises the steps: 1, inputting a shadow image to be processed, constructing a color extraction network, and obtaining the color invariance information of the shadow image as the guide information of shadow removal; 2, constructing a bidirectional mapping module to realize feature extraction and feature mapping of a bidirectional mapping network; and 3, inputting the shadow image and related color guidance information into the constructed bidirectional mapping network to obtain a reconstructed shadow-free image. According to the method, the auxiliary supervision effect of shadow generation on shadow removal is fully considered, a bidirectional mapping network is constructed to remove the shadow in the image, and meanwhile, color guide information is introduced during shadow removal, so that the problem of color deviation possibly occurring during image restoration is reduced, and a better shadow removal effect can be achieved.

The invention provides a method for eliminating defocus noise of random encryption optical scanning holography, belongs to the field of optical encryption and holographic reconstruction, and mainly solves the problem of defocus noise in random encryption optical scanning holography technology. The invention mainly uses convex optimization reconstruction and high-order amplification to eliminate the defocus noise of random encryption optical scanning holography, effectively eliminates the influence of defocus noise of the optical scanning holographic system on holographic reconstruction, and improves the resolution of holographic images.

The invention discloses a method for compensating the vignetting phenomenon of a panoramic image. Including: obtaining the panoramic image, obtaining the outer contour of the panoramic image, and obtaining the center and radius of the outer contour, iteratively estimating the circumferential area of ​​the obtained panoramic image, obtaining the estimated parameters corresponding to the minimum image entropy, and correcting the parameters according to the iterative estimation According to the existing image, a final compensated panoramic image is obtained; and methods such as iterative estimation, obtaining minimum image entropy, parameter correction and the like are specifically disclosed. The present invention does not need to build a special scene in order to obtain the compensation parameters. Through the method of the present invention, an autonomous compensation method is adopted to optimize the compensation parameters autonomously, and the corresponding compensation coefficients are estimated for specific areas through the initial stage, and the subsequent acquisition The images in the image are corrected sequentially through compensation coefficients, which can quickly and accurately obtain better image restoration effects.

The invention proposes a remote sensing image cloud removal residual neural network system, method, equipment and storage medium based on multi-scale convolution and attention, belonging to the field of remote sensing image processing, in order to solve the problem that traditional algorithms have poor robustness and recovery effects do not conform to remote sensing The problem of visual characteristics of images. The deep neural network method achieves a balance between the speed of the high-resolution remote sensing image cloud removal task and the restoration effect; the use of multi-scale context convolution with a larger range of convolution kernel size reduces the memory required by the model and the algorithmic complexity. processing time; and before multi-scale convolution, fine-grained convolution with channel attention module is spliced ​​in the form of residual connection to increase the feature extraction capability of the network; the present invention is more realistic and more in line with the actual scene dedicated to high resolution For the dataset of remote sensing image cloud removal task, no matter what kind of network model, the network weights trained on this dataset have higher adaptability and stronger robustness.

The invention relates to a new fusion method of infrared images and low-light images. The process of the method is: first, perform multi-scale and multi-directional NSCT transformation decomposition on the infrared images and low-light images, and obtain the low-frequency subbands and low-frequency subbands of the infrared images respectively. High-frequency subbands and low-frequency subbands and high-frequency subbands of low-light images; in the low-frequency subbands, a smoothing linear filter is first used to preprocess the low-frequency coefficients. The smoothing linear filter is located at the weight ratio of the mask center. The weights of the four sides and corners are large, and the weights located at the four corners are the smallest. The output response of this smooth linear filter is the weighted average within the template, and is fused using a method based on a combination of local energy ratio and local energy weighting; in high-frequency sub-areas On the belt, first use the Laplacian operator based on the second-order differential to preprocess the high-frequency coefficients. The Laplacian filter is an isotropic filter. Finally, NSCT multi-resolution reconstruction is performed on the fusion coefficients. Get the fused image.

The invention belongs to the filed of digital image processing and particularly relates to an HR (restrict restriction)-LLE (locally linear embedding) weight constraint based face image super-resolution restoration method. The restoration method includes: calculating a great many of HR face samples and average restoration weight constraint, relative to neighboring samples, of residual HR faces; during restoration, subjecting traditional LLE based face super-resolution restoration weight calculation method to weight constraint. The restoration method includes overall restoration and local detail compensation. The overall restoration aims to restore basic characteristics of a standard face as required, and the local detail compensation aims to restore the face image so as to enable the face to be provided with personality characteristics different from other faces. The HR-LLE weight constraint is added in the method in estimating the LLE based restoration weight of the target HR image, so that the weight is closer to the true restoration weight of the HR image in 12 norms. By the method, better image restoration results can be acquired.

The present invention provides an image restoration method in the ciphertext domain, which is used to restore the damaged image generated by the damage of the mask image in the ciphertext domain. Repair encrypted image E P , auxiliary encrypted image E J And the mask image θ; step 2, through the mask image θ to be repaired the encrypted image E P Mark the damaged area Ω and the intact area Φ; step 3, calculate the priority of the pixel c to be repaired in the damaged area Ω in block units to obtain the target block with the highest priority Step 4, in the auxiliary encrypted image E J with the assistance of the target block and all candidate blocks B of the intact region Φ S Perform block matching to get the best matching block Step 5, use the best matching block to repair the target block and select the initial repaired encrypted image E P1 ; Step 6, respectively to the auxiliary encrypted image E J , the damaged area Ω and the intact area Φ are updated; step 7, repeat steps 2 to 6 until all damaged areas Ω are repaired to obtain the repaired encrypted image E' P .

The invention relates to an incremental image restoration method based on a wireframe and an edge structure. The method comprises the following steps: acquiring a scene data picture; constructing a masking layer adapted to a downstream task to carry out model training; constructing a structure recovery model and training the structure recovery model; constructing an up-sampling network of the wireframe structure and training the up-sampling network; if the resolution of the masked image is greater than 256 * 256, performing up-sampling on a 256 * 256 repair wireframe and an edge structure by using a structure up-sampling network until the resolution of the repair wireframe and the edge structure is the same as that of the masked image; inputting the repaired wireframe and edge information into a structural feature encoder to obtain structural features; obtaining a covering position code according to the covering layer; constructing an image inpainting network and training the image inpainting network; and after the model training is finished, image restoration is carried out. Compared with the prior art, the method has the advantages of good image restoration effect, high adaptability and the like.

The present invention is a blind convolution motion blur image restoration method. The method includes the following steps: S1: Collect the observation image, use a differential filter to perform derivative filtering on the observation image, and generate a high-frequency degraded image; S2: Update the clear image; S3: Use the iterative least squares method to solve the following equation to update and solve the point spread Function; S4: If the size of the point spread function is less than the set value, increase the size of the point spread function and return to S2. Otherwise, abort the iteration. S5: According to the optimal point spread function and the observation image obtained in S4, the restored image is obtained by solving the existing non-blind convolution method. The point spread function solved by the present invention is stable, thereby suppressing the impact of noise on the restored image, making the algorithm's fuzzy restoration effect remarkable, highly automated, simple to operate, and fast to execute.

The invention discloses a method for removing a station logo from an image generated against a neural network, comprising: S10 constructing a training data set and a test data set; S20 constructing a generator network, and the generator network is used for inputting an image with a station logo and a station logo The mask image generates an image without a station logo; S30 builds a discriminator network, which is connected to the output of the generator network, and the discriminator network is used to compare the real image without a station logo and the image without a station logo output by the generator network. True or false judgment; S40 trains the generator network and the discriminator network according to the training data set; S50 uses the trained generator network to perform delabeling operation on the test data set. The obtained generator network has a significantly better effect of removing the logo than the traditional algorithm. In most scenes, there are basically no traces of the logo. The image repair degree is good, and the image deformation is not easy to appear in the repaired area.

The invention discloses an ultrasonic super-resolution imaging device based on compressed sensing. According to the device, a signal controller and an ultrasonic transmitting array driving circuit areused for controlling an ultrasonic transmitting array, parallel simultaneous transmitting of ultrasonic waves is achieved according to a binary switch matrix, multiple times of spatial modulation areconducted on an ultrasonic sound field, and the modulated sound field penetrates through a target object to be measured; the single ultrasonic receiver receives a measurement signal and uses a high-speed acquisition module to acquire the measurement signal; a computer reconstructs an image by combining a sound field theory with a compressed sensing algorithm, high-resolution imaging can be achieved at a low sampling rate, the resolution limit of a traditional imaging method is broken through, and the resolution of a restored image can far exceed the resolution of an ultrasonic emission arrayfor emission.

The invention provides a sparse photoacoustic image reconstruction method combined with target detection. The method comprises the following steps: constructing a first training data set; constructing a target detection network; training a target detection network by using the first training data set; processing the original photoacoustic data set by using a target detection network to obtain a cut image group data set; constructing a second training data set; constructing an image reconstruction network; training the image reconstruction network by using the second training data set; processing the sparse photoacoustic image by using a target detection network to obtain a coordinate vector group data set of the region of interest; and reconstructing and outputting a final photoacoustic image by using the image reconstruction network. According to the method, the target detection network based on the Twood-stage network and the Transformers model and the image reconstruction network based on the GAN network and the patchGan network are combined to process the sparse photoacoustic image sampled at intervals, so that the problem of poor image restoration effect caused by sparse photoacoustic image information and excessive background invalid information is solved.

The invention discloses a method for preparing a spar image by using polymethyl methacrylate, relates to the field of plastic high polymer materials, and solves the problems of low color trueness, poor image detail reduction degree and no aesthetic property of a finished product caused by the existing method for making an image into a transparent product. Comprising the following steps: mixing methyl methacrylate and polymethyl methacrylate to obtain viscous slurry M, then putting the viscous slurry M into an autoclave for polymerization to obtain spar A, printing an image on the spar A, casting the viscous slurry M again, and putting the viscous slurry M into the autoclave for polymerization to obtain a spar finished product B; the spar finished product B formed by heating and polymerizing methyl methacrylate and polymethyl methacrylate has better high transparency and is easy to dye, so that the manufactured spar is high in image color trueness and good in image reduction degree, and when the spar is polymerized in the autoclave, the spar A and the viscous slurry M can be fused together, so that a seamless effect can be achieved.