397 results about "Inpainting" patented technology

The invention provides an image inpainting method and system based on antagonistic generation neural network, which comprises the following steps: firstly, constructing a self-encoder convolutional neural network (including encoder and encoding discriminator), a decoder (generator) convolutional neural network, a discriminator convolutional neural network, a global discriminator, and a local discriminator; Then, different loss functions are constructed for the five networks, and the whole network is trained by step-by-step training. Finally, when the network training is completed, the defect image is put into the network for repairing, and the result graph generated by the decoder (generator) is the final repairing result graph. The invention has the advantages that: the potential constraint of the image is kept and the image is thinned; an end-to-end image restoration network is implemented. The dependence of the repair network on the missing position mask information of the image iseliminated. The robustness in practical application is improved.

The invention discloses a depth-based cavity repairing method in viewpoint synthesis. The method comprises steps as follows: (1) performing depth edge one-way expansion preprocessing on an input reference depth map to eliminate inter-infiltration of a foreground pixel and a background pixel in a synthesized virtual viewpoint view; (2) performing pixel mapping transformation by using the preprocessed reference depth map and an original reference view to obtain a virtual viewpoint view with a cavity and a virtual viewpoint depth map; (3) repairing the cavity in the virtual depth map with a pixel-based image repairing method in combination with the relative position relation between the foreground and the background around the cavity in the virtual viewpoint depth map; (4) repairing the cavity in the virtual viewpoint view with an improved sample-based image repairing method in combination with the repaired virtual viewpoint depth map. With the adoption of the method, the virtual viewpoint view with better quality can be obtained, the continuity of image edges can be better kept, and the inter-infiltration of the foreground and background pixels can be eliminated.

The invention belongs to the technical field of digital image intelligent processing, and particularly relates to a multi-level image restoration method based on a partial-to-overall attention mechanism. Image restoration refers to replacing and generating lost or defective image data by using an algorithm. The method comprises the steps that a multi-level deep convolutional neural network generator structure is provided; and an attention mechanism convolution layer from is partially and wholly integrated into a generator and a discriminator of the network. An image block discriminator and four loss functions of reconstruction loss, perception loss, style loss and adversarial loss are introduced in the training process of the network to assist a generator to learn an image restoration task. Experimental results show that the restored image with vivid details and a reasonable overall structure can be generated, and the image restoration problem is effectively solved.

A hyperspectral image inpainting method based on E-3DTV regularity is provided. The method comprises the steps of expanding the original three-dimensional hyperspectral data with noise into matrix along the spectral dimension, and initializing the matrix representation of the noise term and the hyperspectral data to be repaired, and other model variables and parameters under the ADMM framework; performing differential operation on the hyperspectral data to be repaired along horizontal, vertical and spectral dimensions to obtain three gradient maps in different directions, which are expanded into matrices along the spectral dimensions; decomposing the gradient graph matrices in three directions by low rank UV, and constraining the basis matrices of the gradient graph by sparsity to obtain E-3DTV regularity; adding the E-3DTV regularity to the data to be repaired, writing out the optimization model, using the ADMM framework to solve iteratively; obtaining the restoration image and noisewhen the iteration is stable. The invention performs denoising and compression reconstruction on the hyperspectral image data, and improves the enhancement of the traditional 3DTV so as to give consideration to the structural correlation and sparsity of the gradient image, thereby overcoming the defect that the traditional 3DTV can only depict the sparsity of the gradient image and ignores the correlation.

The invention belongs to the technical field of digital image intelligent processing, and particularly relates to a multi-modal feature fusion text-guided image restoration method. The method sequentially comprises the following steps that: a network takes a defect image and corresponding text description as input, and is divided into two stages: a rough repair stage and a fine repair stage; in the rough restoration stage, the network maps the text features and the image features to a unified feature space for fusion, and generates a reasonable rough restoration result by using prior knowledgeof the text features; in the fine repair stage, the network generates more fine-grained textures for the rough repair result; reconstruction loss, adversarial loss and text-guided attention loss areintroduced into network training to assist the network in generating more detailed and natural results. Experimental results show that the semantic information of the object in the missing area can bebetter predicted, the fine-grained texture is generated, and the image restoration effect is effectively improved.

The invention provides a visual tracking and positioning method based on dense point cloud and a composite view, and the method comprises the steps: carrying out the three-dimensional scanning of a real scene, obtaining a color key frame image and a corresponding depth image, carrying out the image restoration, and carrying out the image coding of the key frame image; performing image coding on acurrent frame image acquired by the camera in real time, and selecting a composite image closest to the current frame image as a reference frame image of the current image; obtaining stable matching feature point sets on the two images, and performing processing to obtain six-degree-of-freedom pose information of the current frame camera relative to a three-dimensional scanning point cloud coordinate system; and performing judging by using an optical flow algorithm, if the requirement cannot be met, updating the current frame image to be the next frame image acquired by the camera, and performing re-matching. According to the invention, the association problem of the three-dimensional point cloud obtained by the laser radar and the heterogeneous visual image can be solved, and the effect of realizing visual navigation rapid initialization positioning is achieved.

The invention discloses an image restoration model PRGAN. The image restoration model PRGAN is composed of two independent generative adversarial network modules. An image restoration network module PConv-GAN is formed by combining part of a convolutional network and an adversarial network and is used for restoring an irregular mask, and meanwhile, making the overall texture structure and color ofan image closer to those of an original image according to feedback of a discriminator. In order to solve the problems of local chromatic aberration and mild boundary artifacts of an image caused bythe deficiency of a restoration network module, the invention designs an image optimization network module. The image optimization network module Res-GAN combines a deep residual network with an adversarial network, and the image optimization network module is trained by combining the adversarial loss, the perception loss and the content loss, so that the information of the non-missing area in theimage is reserved, the consistency of the texture structure of the image in the non-missing area is kept, and the purposes of eliminating the local chromatic aberration phenomenon and solving the pseudo boundary are achieved.

The invention relates to an image restoration method based on a wavelet transform attention model, and is used for providing a high-quality image restoration method. The method is based on an image restoration network, and the network is based on a GAN and comprises a generator network and a discriminator network. The generator network part firstly utilizes DWT to decompose known region features in a damaged image into multi-frequency sub-bands, then further extracts deep information of a known region through an attention mechanism, and finally generates a restored image through IDWT. A full convolution discriminator network structure of a PatchGAN is adopted in a discriminator network part; the restored image and the ground-true image are jointly used as input, the generator network and the discriminator network are alternately optimized, back propagation is carried out by minimizing a global loss function, generator network parameters are iteratively adjusted, and finally, the optimized generator network can generate a high-quality image restoration result.

The invention relates to a face image restoration method based on a multi-scale local self-attention generative adversarial network. The method comprises the following steps: acquiring a face missing image and a corresponding mask, and performing preprocessing; constructing a multi-scale-based local self-attention generative adversarial network, and performing training modeling on the multi-scale-based local self-attention generative adversarial network by using the defective face image data set to obtain a face repair model; and through the multi-scale local self-attention generative adversarial network model, repairing the defect face image to be detected. According to the invention, the multi-scale structure and the dual-channel local self-attention module are added into the generative network, so that the technical problems of unstable training, low repair precision and efficiency, lack of symmetry and mode collapse of the generative adversarial network in the face repair problem are effectively solved, and an efficient, accurate and stable repair method is provided for face repair.

The invention relates to a three-dimensional target detection method and system based on image restoration. The method comprises the following steps: obtaining an RGB image and radar point cloud of athree-dimensional target; generating a two-dimensional target detection box on the RGB image according to a two-dimensional target detection algorithm; for the picture with the shielded target, carrying out instance segmentation by adopting an instance segmentation algorithm to obtain a mask at the shielded position of the target, and then calculating a complete mask of the target according to morphological closed operation; converting the radar point cloud into a depth map through a camera matrix, performing image restoration on the shielded part of the target on the depth map, and extractingdepth information of the target in a depth map form according to a complete mask of the target after the restoration is completed; converting the depth information in the depth map form of the targetinto restored point cloud; and inputting the repaired point cloud into a three-dimensional target detection network for three-dimensional target detection. Compared with the prior art, the three-dimensional target detection method plays a role in reducing offset and improving precision in three-dimensional target detection.

The invention discloses an image restoring method and device. The method comprises the following steps of: step A, obtaining a first layer region to be restored of an image damaged region from inside to outside, and reading structural information and a color of the known image region which is close to the first layer region to be restored; step B, carrying out gradient analysis and color estimation on the first layer region to be restored according to the structural information and the color of the known image region to obtain gradient information and a color of the first layer region to be restored; step C, taking the gradient information and the color of the first layer region to be restored as initial values to be extended to the first layer region to be restored; step D, carrying out texture synthesis restoration on the first layer region to be restored according to the initial values to obtain a first restoring result; and step E, taking the first restoring result as the known image region to circularly execute the steps A to D until the image damaged region is restored. With the adoption of the image restoring method and device, disclosed by the invention, the structural region can be accurately positioned, the block matching accuracy is improved, and the introduction and the reproduction of error information are reduced.

The invention discloses an image restoration method based on a Criminisi algorithm. The core idea of the Criminisi algorithm is to calculate the priority of a repair block, determine a repair area according to the priority, find an optimal matching block according to an SSD criterion and fill the optimal matching block, and finally update the confidence of the repair block after repair until the repair of a to-be-repaired area is finished. According to the image restoration method based on the Criminisi algorithm, optimization is carried out on the basis of an existing Criminisi algorithm, andthe phenomena that in the prior art, after image restoration, an image is broken, and the image structure is disordered are solved.

The invention belongs to the technical field of construction site monitoring and protection, and discloses a construction site worker dressing detection method, system and device, and a storage medium. The method comprises the steps: receiving a real-time scene image collected at a construction site, carrying out the quality evaluation of the collected real-time scene image, carrying out the imagerestoration and enhancement of the real-time scene image with unqualified quality evaluation or at night, and obtaining a real-time scene image with qualified quality evaluation; performing real-timedetection of constructors and dresses of the constructors on the collected real-time scene images of the construction site after quality restoration, and marking the positions of the constructors onthe real-time scene images; transmitting the detected real-time scene image to mobile equipment, and carrying out alarm processing. The high safety management level of construction personnel is improved; once a constructor is found, dressing of the constructor is automatically recognized, safety threats faced by the constructor are analyzed, real-time intelligent analysis is carried out, WeChat ispushed to an operation and maintenance manager, and safety accidents can be greatly reduced.

The invention discloses a human body shape and posture estimation method for an object occlusion scene, and the method comprises the steps: converting a weak perspective projection parameter obtainedthrough calculation into a camera coordinate, and obtaining a UV image containing human body shape information under the condition of no occlusion; adding a random object picture to the human body two-dimensional image for shielding, and obtaining a human body mask under the shielding condition; training a UV map restoration network of an encoding-decoding structure by using the obtained virtual occlusion data; inputting a human body color image shielded by a real object, and constructing a saliency detection network of an encoding-decoding structure by taking the mask image as a true value; supervising human body coding network training by using the hidden space features obtained by coding; inputting the shielded human body color image to obtain a complete UV image; and recovering the human body three-dimensional model under the shielding condition by using the vertex corresponding relationship between the UV image and the human body three-dimensional model. According to the method, the shape estimation of the shielded human body is converted into the image restoration problem of the two-dimensional UV chartlet, so that the real-time and dynamic reconstruction of the human body inthe shielded scene is realized.

The invention relates to a face image restoration method introducing an attention mechanism, and the method comprises the steps: (1) obtaining an original data set, carrying out the image preprocessing, obtaining a needed face image data set, and reasonably dividing and arranging the face image data set into a test set and a data set; (2) inputting the training data set into an image restoration model introduced into a context attention layer for training, wherein two parallel encoder networks are introduced into a generator network, one encoder network is used for performing convolution operation to extract advanced feature images, and the other encoder is used for introducing a context attention layer network to realize long-range association between a foreground region and a backgroundregion; and (3) inputting the test data set into the trained face restoration model, and testing the restoration capability of the trained restoration model for the defective face image. According tothe method, after the context attention layer is introduced, the problem that the background region information cannot be fully utilized by the restoration model due to the limited receptive field size of the convolutional neural network is solved, the long-range association of the background information and the foreground region is realized, and the background region information is fully utilizedto fill the foreground region. After the context attention layer is introduced, the restoration model obtains a better restoration effect on some detail textures, and the restoration effect of the face image is also improved on the whole.

The invention discloses a face image restoration method based on a dense expansion convolution self-coding adversarial network. The method comprises the following steps: firstly, preprocessing face public data to obtain a face data set; secondly, constructing a dense expansion convolution self-coding adversarial network; then pre-training a dense expansion convolutional self-encoding generation network by utilizing reconstruction loss, and alternately carrying out the following training steps: (1) training a double-discrimination network by utilizing adversarial loss; (2) training a pre-trained generative network by using joint loss; and then obtaining a trained dense expansion convolution self-encoding generation network, finally inputting the image to be restored into the generation network, and fusing the generation image and the defect image to obtain a final restored image. According to the method, the face image restoration problem of serious semantic information loss and large-area random region loss is solved.