563 results about "Digital image processing" patented technology

The invention provides an image restoration and super-resolution reconstruction system and method based on deep learning, and belongs to the technical field of digital image processing. The invention aims to solve the problems of high calculation complexity and resource consumption, limitation of long sequence processing, high training difficulty and texture scene deficiency when a multi-scale residual network based on a Transform architecture is used for image resolution conversion. The reconstruction system comprises: an image preprocessing module performing window division and video memory optimization on an input low-resolution image; the multi-layer fusion network dynamically adjusts the characteristics of the low-resolution image, captures channel information in different scenes, performs interactive fusion, performs comparison supervision, establishes an information communication channel, dynamically adjusts and optimizes parameters through negative feedback, and obtains a super-resolution image. And the loss function module maximizes the similarity of the super-resolution image and the high-resolution image in the segmentation feature space to obtain a final super-resolution image.

The invention discloses a rock mass fracture multi-field coupling two-phase flow analysis method and a test system, and relates to the technical field of rock fracture temperature-seepage-stress coupling action mechanism research. The analysis method comprises the following steps: saturating a crack sample with a first dyeing fluid under a target temperature and pressure condition, and obtaining a single-phase saturated flow image of the crack sample; on the basis of the single-phase saturated flow image, through optical iteration inversion calculation, obtaining fracture opening two-dimensional distribution characteristics under a target condition; under the condition of the same temperature and pressure, displacing the first dyeing fluid with the second dyeing fluid or co-flowing with the first dyeing fluid to obtain a two-phase flow image of the crack sample; performing digital image processing on the two-phase flow image and the single-phase saturated flow image, and extracting phase distribution information of the second dyeing fluid; and performing fusion analysis on the phase distribution information of the second dyeing fluid and the fracture opening two-dimensional distribution characteristics, and calculating a two-phase flow structure and saturation distribution characteristics. According to the method, rock mass fracture two-phase flow process quantitative analysis can be realized under different temperature and pressure conditions.

The invention relates to the field of digital image processing and information security, and discloses a document tampering detection method and system based on image data processing, and the method comprises the following steps: firstly, extracting the noise residual error and microscopic penetration characteristics of a document image, and constructing a physical potential energy field and a virtual viscous resistance field; performing dynamic evolution by using a Darcy law variant model to generate a virtual flow velocity vector field to simulate the slippage behavior of the fluid in the heterogeneous medium; and then a heterogeneous graph is constructed based on the flow field divergence singular points and the streamline trajectory, deep reasoning is carried out by using graph neural network aggregation node dynamic features, and finally a tampering localization mask is generated. According to the method, the fluid mechanics field theory is innovatively introduced, hidden static texture differences are converted into remarkable dynamic flow field anomalies, the problem that microscopic tampering traces are difficult to capture in the prior art is solved, and the detection precision and generalization ability in a complex document scene are remarkably improved.

An apparatus and method for training and using a computing operation for digital image processing are provided. The apparatus and method may be used for 3-dimensional medical images. An exemplary method for digital image processing comprises: receiving an image displaying at least one detectable structure, determining the detectable structure; segmenting the image to obtain a segmentation mask that is associated with a geometric shape and comprises at least one quantifiable visual feature; generating a mesh based on the quantifiable visual feature; computing at least on quantifiable visual parameter based on the mesh; extracting quantifiable visual data from the image based on the quantifiable visual parameter; training the computing operation with the quantifiable visual data. The method for digital image processing further comprises: receiving another image; segmenting, generating a mesh, computing quantifiable visual parameters, and extracting quantifiable visual data; and classifying the extracted quantifiable visual data with the trained computing operation.

The embodiment of the invention provides an edge detection method and device based on gradient weighted fusion and an adaptive threshold, and is applied to the field of computer vision and digital image processing. The method comprises the steps of firstly preprocessing an input image, then extracting two groups of gradient magnitude diagrams and directional diagrams through an adaptive morphological operator and a traditional difference operator, and constructing a weighted fusion function according to the gradient direction consistency of each pixel point to obtain a fused gradient magnitude diagram; and adaptively determining a high threshold and a low threshold based on an information entropy maximization principle, executing an improved Canny process by combining the fused gradient magnitude diagram and the second gradient directional diagram to obtain an initial edge diagram, and outputting a final edge detection result after dynamic structure element optimization. In this way, the defects that in a traditional edge detection method, gradient information extraction is not precise, threshold selection lacks adaptability, and an edge result is fractured can be overcome, more robust and more accurate image edge detection is achieved, and the reliability of an edge detection algorithm in a complex image scene is improved.

The invention relates to the technical field of digital image processing and computer vision, and discloses a remote multi-view video stitching and monitoring method and system based on digital twinning, and the method comprises the steps: constructing a digital twinning model containing geometric, material and semantic information, and calibrating and initially registering a multi-view camera; based on the model and camera parameters, collecting a preprocessed video stream, and refining the camera pose in combination with the model; predicting expected visual content by using the model and the refined pose, and performing non-rigid alignment on the preprocessed video frame; based on the aligned video frames, a model semantic optimization splicing suture line is utilized to generate a reconstructed video, and information loss is repaired in combination with a model context; and finally, performing integrated monitoring interaction on the repaired reconstructed video and the digital twin model, comparing and detecting abnormality, and updating the model by using the reconstructed video. According to the invention, the problems of traditional remote monitoring in the aspects of scene comprehensive perception, intelligent anomaly detection and synchronism of environment digital representation and physical reality are solved.

The invention relates to a computer vision and digital image processing technology, in particular to an image defogging method of a multi-scale convolutional neural network based on dark channel prior. Estimating the initial transmissivity and atmospheric light of the foggy image based on a dark channel prior method; the foggy image is converted into a gray level image, a gray level threshold value is selected to be used for distinguishing a sky area and other areas, and the initial transmissivity after screening is obtained; optimizing the initial transmissivity through a multi-scale convolutional neural network by adopting a coarse optimization stage and a fine optimization stage in sequence to obtain refined transmissivity; and reconstructing a fogless image according to the refined transmissivity by adopting an atmospheric scattering model. According to the method, the advantages of a physical model and deep learning are fused, the problems of supersaturation, edge artifacts, color distortion and the like in a sky region in a traditional method are solved, the PSNR (Peak Signal to Noise Ratio) of a defogged image is remarkably improved (up to 23.30), the SSIM (Subscriber Identity Module) (up to 0.978) and the like, and a robust visual enhancement scheme is provided for scenes such as automatic driving and traffic monitoring.

The invention discloses an event-image dual-mode fusion video turbulence correction method, medium and system, and belongs to the field of digital image processing, and the method comprises the steps: synchronously obtaining an event voxel of a target and a to-be-corrected image sequence; inputting the event voxels into a pre-selected coding and decoding structure to obtain features, projecting the features along the optical flow direction, and performing three-dimensional reconstruction to obtain target motion features; extracting background scene representation; fusing the target motion feature and the background scene representation in a channel dimension to obtain an edge guide feature; inputting the image sequence and the image sequence into a pre-trained video restoration network to obtain a turbulence-corrected video; the training loss of the coding and decoding structure comprises an optical flow field estimated according to an image sequence without turbulence disturbance and a target motion field obtained by projecting the feature along the optical flow direction. The method can accelerate the recovery process and improve the recovery quality.

The invention discloses a discharge channel extraction method and system based on edge detection and texture feature fusion, and belongs to the technical field of power equipment fault diagnosis and digital image processing, and the method comprises the steps: obtaining an ultraviolet weak light image of power equipment, and carrying out the preprocessing; multi-scale edge detection is executed based on the preprocessed image, edge information is fused through a multi-scale voting mechanism, and a candidate edge graph is generated by combining direction consistency connection fracture edges; extracting texture features of the pre-processed image, wherein the texture features comprise a rotation invariant local binary pattern uniformity feature and a multi-direction gray level co-occurrence matrix feature; according to a dynamic distribution weight coefficient of a discharge type, carrying out weighted fusion on the edge intensity of the candidate edge graph and the texture features, and generating a discharge channel confidence score graph; and carrying out binarization segmentation and contour optimization processing on the confidence score graph of the discharge channel, and extracting morphology quantization parameters of the discharge channel.

The invention discloses a micro-fluidic chip for basalt CO2 mineralization reaction and a preparation method, the micro-fluidic chip comprises a bottom glass plate, a middle epoxy resin layer and a top glass plate, the preparation process comprises the following steps: installing a metal injection / outpouring pipeline on the bottom glass plate drilled with corresponding bolt holes, and forming an accommodating cavity by using a sealing ring; preparing and screening basalt particles according to the particle size distribution of a target reservoir, and filling the accommodating cavity with the basalt particles; quantitatively analyzing a particle arrangement structure based on a digital image processing technology, ensuring that the geometric similarity between the particle arrangement structure and a target reservoir is greater than or equal to 90%, and if the particle arrangement structure does not reach the standard, refilling; after the verification is passed, covering a top glass plate and pressing by using a bolt; and finally, pouring epoxy resin and sealing to form the micro-fluidic chip. The chip can bear high temperature and high pressure, supports visual observation and quantitative analysis of CO2 mineralization reaction in a deep geological environment, has high-precision pore reproduction, long-period sealing performance and high preparation efficiency, and is suitable for interdisciplinary research in the fields of energy environment and geophysics.

The invention relates to the field of digital image processing, in particular to a super-resolution imaging method based on focal plane splicing and adaptive fusion. According to the method, sub-pixel offset among nine CCDs is preset through hardware, and nine frames of low-resolution image sequences with accurate displacement are obtained in push-broom. A central image is taken as a reference frame, high-precision mapping is realized based on hardware offset, motion estimation errors are avoided, effective pixels are screened by calculating robustness weight, an anisotropic Gaussian kernel function with a self-adaptive local structure is constructed so as to maintain image edge and detail features, and each frame is accumulated to a high-resolution grid in a weighting mode, so that a high-resolution image is obtained. And a sample compensation mechanism based on cumulative robustness is introduced, a fusion strategy is adaptively adjusted in an information insufficient area, and finally a high-resolution image is generated through normalization. The method significantly improves the imaging quality, suppresses artifacts and noise, and is suitable for the field of satellite remote sensing.

The invention discloses an image local fuzzy region detection method and device, equipment, a medium and a product, and relates to the field of digital image processing, and the method comprises the steps: dividing an original image into a plurality of sub-images; calculating a global mean value of the high-frequency energy value of the original image according to the high-frequency energy value of each sub-image; calculating a global mean value and a global standard deviation of the texture entropy of the original image according to the texture entropy of each sub-image; calculating the energy sensitivity and entropy sensitivity of each sub-image according to the results; obtaining an energy weight coefficient and an entropy weight coefficient of each sub-image according to the energy sensitivity and the entropy sensitivity of each sub-image; according to the energy weight coefficient and the entropy weight coefficient, fusing the normalized high-frequency energy value and texture entropy of each subimage to obtain a fusion feature value of each subimage; obtaining a target sub-image of which the fusion feature value is greater than a threshold value; and marking a target sub-image in the original image, wherein the target sub-image is a local fuzzy region in the original image. According to the invention, the reliability of local fuzzy region detection can be improved.

The invention provides an image caching method and circuit and a computer readable storage medium, and relates to the technical field of digital image processing. The method comprises the steps that n line caches are provided, the number of pixels capable of being stored in each line cache is m, and the pixels of each line in an original image of a current picture are stored in the corresponding line cache; according to the row index value i (0 < = i < n) and the column index value j (0 < = j < m) of the current processing pixel, sequentially outputting the neighborhoods of the current processing pixel according to the neighborhood size of p * q; wherein the size of the original image is C * R, C is the total column number of the original image, and R is the total row number of the original image; wherein the size of the neighborhood on which the method depends is p * q, p is the column number of the neighborhood, and q is the row number of the neighborhood. The circular processing of the data is realized by circularly utilizing a small amount of line cache, so that the delay of each level of image processing algorithm is controlled within a plurality of lines, the data processing flow length is reduced, and the processing delay is reduced.

The invention relates to a method for extracting the mass center of a laser spot image based on laser size and gray scale, and relates to the technical field of digital image processing, optical testing and Matlab programming. The method comprises the following steps: step 1, obtaining a fourth grayscale image; 2, processing the generated fourth gray level image, and solving an X coordinate value and a Y coordinate value of the mass center of each image in the test time period; and step 3, obtaining a final centroid coordinate value. According to the method for extracting the mass center of the laser spot image based on the laser size and the gray level, the influence of noise points, stray light and the like on the overall measurement result is reduced to the maximum extent, meanwhile, the numerical image processing process is clear in logic and reasonable in analysis, the actually obtained result meets the expectation, data calculation is simple, and the processing process is efficient and rapid. And adopted data processing software is also universal and common, and is easy to popularize.

The invention discloses a vision enhancement robust digital dark watermarking method based on deep learning, a storage medium and equipment, and belongs to the technical field of digital image processing and information security. According to the method, a visual enhancement module comprising a self-adaptive watermark region adjustment module and a frequency enhancement module is constructed, and a multi-region local loss training mechanism and a screen shooting simulation module are combined, so that the visual quality of a watermark image is improved, and meanwhile, the robustness of the watermark image for resisting physical attacks such as screen shooting is enhanced. The specific processing flow comprises the following steps: acquiring an original image and watermark information; embedding the watermark information into the image by using an encoder, wherein an embedded region is optimized by a self-adaptive region guide matrix generated based on the image edge and gray information; in the model training process, a noise layer simulating screen shooting physical distortion is introduced, and the area weight is dynamically adjusted according to residual error distribution in the later stage of training so as to focus and optimize the image quality; watermark information is extracted from an image which may suffer from an attack by using a decoder.

The invention discloses an ultrahigh-definition panoramic image adaptive HDR fusion method based on data driving, and relates to the technical field of digital image processing, and the method comprises the steps: carrying out the coarse alignment and high dynamic range fusion of a low-resolution line graph sequence, and generating a low-resolution HDR panoramic image; carrying out sharpening processing on the low-resolution HDR panoramic image to obtain a low-resolution sharpened HDR image, and meanwhile, obtaining a multi-scale residual spectrum generated by sharpening; performing up-sampling on the multi-scale residual spectrum, and performing small-sample element learning fine tuning on the weight prediction network in combination with the low-resolution sharpened HDR graph to obtain a fine-tuned weight prediction network; performing weight prediction and layer-by-layer fusion on the registered high-resolution tile sequence on a multi-scale pyramid based on a fine tuning weight prediction network model; according to the method, the sharpening processing is implemented on the low-resolution HDR panorama, and the multi-scale residual spectrum is extracted, so that the detail sensitivity and adaptability in a complex scene are improved.

The embodiment of the invention provides a three-dimensional space change detection method and device, equipment and a storage medium, and relates to the field of digital image processing. The method comprises the following steps: acquiring image sequences of the same area shot by the unmanned aerial vehicle in two periods; for each image sequence, performing point cloud reconstruction on the region based on the image sequence to obtain the group of point cloud data; point cloud registration is carried out according to the similarity between the feature descriptors of all the points in the two sets of point cloud data, a matching point pair set comprising a plurality of matching point pairs is obtained, and two points in each matching point pair belong to two pieces of point cloud data and are closest to each other; and for each matching point pair, geometric information of two points in the matching point pair in respective point cloud data is acquired, and if it is determined that the difference between the two pieces of geometric information meets a preset difference condition, it is determined that the corresponding position of the matching point pair in the region is changed. According to the embodiment of the invention, the problem of insufficient spatial change detection precision in a complex scene is solved.

The invention discloses a lightweight mural virtual restoration method based on an inactive generative model, and belongs to the technical field of digital image processing and computer vision. The objective of the invention is to solve the technical problem of semantic and texture recovery imbalance of an existing deep learning repair model in face of cultural heritage such as murals. A damaged image and a mask are stacked into a multi-channel tensor to be input into a pre-trained AFMNet model, the model adopts an encoder-bottleneck-decoder architecture, and an encoder uses a multi-domain partial convolution module without an activation function to synchronously extract and fuse local texture and frequency domain global structure information and synchronously update the mask; the bottleneck portion models the lowest resolution features and masks thereof through a lightweight multi-head attention module to capture remote dependencies, and a decoder upsamples in conjunction with the features of the jump connection transfer to reconstruct image details. The method is mainly used for carrying out efficient and high-fidelity digital virtual restoration on damaged visual cultural heritage such as ancient murals.

The invention relates to the technical field of digital image processing, in particular to a real-time video semantic segmentation system and optimization method fusing an edge size model, and the system comprises an edge end lightweight model belonging to terminal equipment, an edge server large model belonging to an edge server, a dynamic frame distribution module and a self-adaptive cache module. An optical flow change rate and scene complexity are analyzed through edge detection and entropy calculation by a dynamic frame distribution module, then a dynamic distribution strategy is output, the dynamic distribution strategy is formed by processing a low-change frame by a small model and processing a key frame by a large model, a high-confidence segmentation result of the large model is cached by a self-adaptive cache module, and a high-confidence segmentation result of the large model is obtained. The method can be used for reuse of small models in similar scenes so as to reduce calling frequency of large models, flexibly deploy resources of the large models and the small models during operation and solve the real-time performance-precision contradiction in edge computing scenes, and compared with a single model scheme, the method has the advantages that precision is improved, and load is reduced.

The invention discloses a bridge disease detection method based on a semantic segmentation algorithm and digital image processing, and belongs to the field of infrastructure health monitoring. The method comprises the following steps: carrying out disease detection on an image through an RT-DETR-R101 model, generating a bounding box and mask information of a disease region, reconstructing DeepLab v3 + and FCN into M-DeepNet and A-FCNet by utilizing a multi-scale feature complementation module (MFCM) and an adaptive mask optimization module (AMOM) respectively, and automatically identifying and extracting a bridge disease mask image in the image to be detected by combining an M-DeepLabNet semantic segmentation algorithm and an A-FCNet semantic segmentation algorithm. And then calculating the pixel size occupied by the disease by utilizing technologies such as digital image processing and the like, and measuring the real size of the disease by combining a pixel calibration method, thereby realizing qualitative and quantitative analysis on the bridge disease.

The invention belongs to the technical field of digital image processing, and particularly relates to a knowledge distillation-based cross-modal fusion lightweight defect detection method, which comprises the following steps of S10, cross-modal fusion distillation; through a bidirectional vision-language alignment mechanism, the frozen multi-modal knowledge of a teacher model vision-language basic model VLM is migrated to a lightweight student model, and the dual-path fusion module comprises text condition region representation injected with semantic context and region anchoring semantic embedding fused with spatial vision clues; step S20, cross-header word-region alignment is carried out; embedding the fusion visual features generated by the two-way fusion module and the enhanced text to generate cross-head prediction so as to simulate the semantic-space association capability of a teacher model; s30, knowledge distillation loss is fused; according to the method, the multi-modal basic model is fused and distilled into the lightweight single-modal detection model, the detection performance in a defect detection scene can be improved, and compared with the basic model, the reasoning speed is greatly improved, and the parameter quantity is reduced.

The invention belongs to the technical field of digital image processing, and discloses an image restoration method, system and device based on multi-modal large model driving. The method comprises the following steps: receiving a to-be-recovered degraded image, inputting the degraded image and a preset multi-task text cue word into a multi-modal large model, analyzing and processing the degraded image, and generating an operation instruction and a visual description prompt of the degraded image; encoding the operation instruction and the visual description prompt to respectively obtain a corresponding task intention vector and a content guide vector; and fusing the task intention vector and the content guide vector through an image restoration model, reconstructing the degraded image, and outputting a restored image. According to the method, the image problem can be automatically and accurately diagnosed, rich guidance information is provided from the two dimensions of operation and content, and intelligent, automatic and high-fidelity image recovery is realized.

The invention relates to the technical field of computer vision and digital image processing, and discloses a two-stage damaged face image restoration method based on a generative adversarial network, and the method comprises the following steps: inputting a damaged face image into a reconstruction network, the input feature map is processed through at least one self-adaptive harmonic purification convolution module built in the reconstruction network, and a purified feature map is generated; based on the purified feature map, the reconstruction network outputs a low-frequency basic image and a final degradation feature map; inputting the low-frequency basic image and the final degradation feature map into a detail generator to generate a high-frequency detail map; and fusing the low-frequency basic image and the high-frequency detail image to obtain a final restored image. According to the method, the adaptive harmonic purification convolution module is arranged in the reconstruction network, frequency domain analysis can be carried out on local image features in the initial stage of feature extraction, and the harmonic attention mask is adaptively generated according to the content.

The invention relates to the technical field of digital image processing, in particular to a real-time fusion method and system for an infrared image and a visible light image, and the method comprises the steps: firstly, carrying out the self-adaptive morphological target extraction of the infrared image, and generating an accurate binary mask through the dynamic adjustment of the size of a structural element, the small size of a high gray level change region, and the large size of a low change region; meanwhile, performing multi-scale decomposition on the visible light image, and adaptively enhancing texture features of each scale based on local contrast; a space attention weight map is generated based on the binary mask, and a channel attention weight map is generated based on the multi-scale features; and finally, realizing image fusion through a wavelet domain weighted fusion algorithm. The problem of edge blurring caused by inaccurate infrared target contour extraction in a traditional method is effectively solved, and the target detection accuracy in a complex environment in military reconnaissance, automatic driving and other scenes is remarkably improved.

System and method for improving the shaving experience by providing improved visibility of the skin shaving area. A digital camera is integrated with the electric shaver for close image capturing of shaving area, and displaying it on a display unit. The display unit can be integral part of the electric shaver casing, or housed in a separated device which receives the image via a communication channel. The communication channel can be wireless (using radio, audio or light) or wired, such as dedicated cabling or using powerline communication. A light source is used to better illuminate the shaving area. Video compression and digital image processing techniques are used for providing for improved shaving results. The wired communication medium can simultaneously be used also for carrying power from the electric shaver assembly to the display unit, or from the display unit to the electric shaver.

The invention relates to the technical field of image recognition processing, in particular to an automatic organoid labeling method based on machine vision and deep learning, which comprises the following steps: acquiring a to-be-processed image in a high-throughput culture environment, and preprocessing the image; inputting the preprocessed to-be-processed image into a preset digital image processing framework; the digital image processing architecture comprises an image filtering unit, a three-dimensional boundary reconstruction unit and a texture gray analysis unit; generating an independent target area mask through a three-dimensional boundary reconstruction unit; generating a high-density core mask through a texture gray analysis unit; based on the target area mask and the high-density core mask, constructing a three-dimensional geometric topological graph of the organoid; according to the method, the three-dimensional space morphological characteristics of the neighborhood nodes are obtained through processing by means of a numerical iterative aggregation algorithm, and the image feature parameters are output through analysis of the three-dimensional space morphological characteristics.

At least a method for training a target machine learning model for enhancing a digital image processing is provided. The method comprises receiving a first data set including a first plurality of digital images, training a first machine learning model using the first data set and a second data set including a second plurality of digital images, generating, by the first machine learning model that is trained, a target data set including a third plurality of digital images, the third plurality of digital images having noise represented by respective noise values that are lower than the noise represented by the respective noise values of the first plurality of digital images, and training the target machine learning model using the target data set and the first data set including the first plurality of digital images for enhancing at least one characteristic of a new digital image.

The invention relates to the technical field of digital image processing, and discloses a gray level conversion system in digital media image processing, which comprises an image input module used for receiving a color image and carrying out normalization processing on each channel pixel value of the image; the structure information extraction module is used for calculating pixel gradient information based on at least one color channel of the normalized color image and constructing a structure tensor; the saliency region extraction module is used for performing region comparison analysis on the normalized color image; the weighted tensor construction module is used for calculating and obtaining a fusion weight tensor based on the structure response information and the normalized saliency map; and the visual perception optimization module is used for readjusting the fusion weight tensor or the initial grayscale image. By adopting the technical scheme based on the combination of salient region extraction and structure perception, the technical effect of simultaneously considering the image structure and visual saliency in the grey-scale map generation process is achieved.

The invention discloses a monitoring image multi-target tracking method based on deep learning, and relates to the technical field of digital image processing, and the method comprises the following steps: S1, generating a multi-scale feature pyramid; s2, extracting causal feature vectors; s3, generating a discrete codebook index; s4, constructing a dynamic graph; s5, inputting the dynamic graph into an improved ASTGCN network, iteratively fusing space-time attention weights and Hamilton dynamics evolution characteristics between nodes through cascaded Hamilton space-time blocks, and generating a final prediction state of each historical track node; s6, solving an optimal correlation matching matrix by using a Hungary algorithm; and S7, extracting all active track information. According to the method, the limitations of identity drift caused by variable appearance characteristics and inaccurate prediction caused by lack of physical priori in a traditional multi-target tracking method are overcome, and an efficient, accurate and robust solution is provided for intelligent video monitoring.

The invention discloses a video subtitle erasing method, device and equipment and a storage medium, and relates to the field of digital image processing, and the method comprises the steps: detecting subtitles of a subtitle video to be erased, merging timestamps of the same subtitles in the subtitle video to be erased, and determining a subtitle fragment set and a subtitle-free fragment set; splitting the subtitle segment into independent shot segments by using a preset lens splitting algorithm, and analyzing video frames of the independent shot segments to obtain a first frame and a tail frame; determining a target reference frame based on the first frame and the tail frame, generating an expanded independent shot segment according to the target reference frame and the independent shot segment, and segmenting a target character mask; and generating an erased independent lens segment according to the expanded independent lens segment and the target character mask through a preset erasure algorithm, performing a preset post-processing optimization operation on the erased independent lens segment to obtain a target independent lens segment, and integrating the target independent lens segment and the subtitle-free segment set to generate a target video. According to the method and the device, the video subtitles can be accurately erased.