2002 results about "Image area" patented technology

The invention relates to the technical field of medical image detection, and discloses a medical image quality detection method based on image processing. The method comprises the following steps: acquiring medical image data to be detected, wherein the medical image data comprises a multi-modal scanning image sequence and corresponding acquisition parameters; the medical image data are preprocessed, standardized image data are generated, and the standardized image data comprise unified parameters of spatial resolution, gray scale range and noise level; extracting structural features of the standardized image data, wherein the structural features comprise tissue boundary gradient distribution, texture consistency and local contrast information; constructing a quality evaluation model according to the structural features, wherein the quality evaluation model analyzes a mapping relationship between the structural features and preset quality indexes through a dynamic convolutional network; and outputting a quality defect detection result based on the quality evaluation model, wherein the quality defect detection result marks an image region with artifacts, fuzziness or distortion.

The invention discloses an intelligent glasses image adjusting system based on eye movement tracking and gesture fusion, and relates to the technical field of intelligent equipment. A multi-modal sensing module is arranged to construct a multi-modal sensing layer to capture eyeball movement tracks and gesture actions; a fixation point prediction module is set to process a dynamic scene through a space-time attention mechanism to obtain a fixation point prediction area, a gesture semantic understanding module is set to process gesture actions based on a Transform architecture, and the gesture actions of a user are converted into image adjustment instructions. An image enhancement strategy setting module designs a multi-stage image enhancement strategy according to the fixation point prediction area and the image adjustment instruction, and sets a dynamic adjustment intensity control module to perform adaptive adjustment to obtain a dynamic adjustment intensity control result; an eye movement-gesture cooperative control module is arranged to provide an eye movement-gesture cooperative control mechanism to realize image area selection and parameter adjustment, and accurate image area selection and parameter adjustment are realized.

The invention relates to the technical field of data processing, and provides a target detection method and device, a model training method and device, electronic equipment and a medium. The target detection method comprises the steps that a to-be-recognized image and a query text are acquired, and the query text is used for querying a target object corresponding to the query text in the to-be-recognized image; performing image recognition on the to-be-recognized image to obtain image description features and region detection visual features; performing regional multi-modal fusion processing on the image description features and the regional detection visual features to obtain regional multi-modal fusion features; performing feature fusion processing on text features obtained based on the query text and the regional multi-modal fusion features to obtain text regional fusion features corresponding to the query text; and a target detection result is obtained based on the text features and the text region fusion features, so that the fusion degree of text semantics and image region features is improved, and the accuracy and robustness of target detection in a complex scene are improved.

The invention relates to the technical field of display screens, in particular to a liquid crystal display screen backlight local dimming method and device, a terminal and a medium. Comprising the steps of obtaining brightness and color distribution data of a display picture; calculating a target brightness value of each backlight block through the dimming algorithm module based on a dynamic partition control strategy and a preset brightness mapping model; wherein the dynamic partition control strategy is used for dividing adjacent backlight blocks into dynamic dimming groups according to brightness characteristics of a picture area on the premise that the number of the backlight blocks is not increased; the backlight control device drives an LED backlight block of the backlight module according to the target brightness value; and monitoring a display effect in real time, and adjusting parameters of the brightness mapping model based on feedback. According to the invention, the problems of cost increase and related derivation caused by increase of the number of partitions in the prior art can be solved.

The invention provides a safety production behavior monitoring method and system based on AI video analysis. The method comprises the steps of collecting a real-time video data stream of a production area; inputting each frame of video image in the real-time video data stream into a target detection model for target detection to obtain a personnel target output by the target detection model and a target position coordinate of the personnel target in each frame of video image; cutting out a local image area of the personnel target in each frame of video image based on the target position coordinate, and performing feature recognition based on the local image area to obtain personnel features; performing comparison on the basis of the personnel characteristics and the personnel standard behavior characteristics to obtain personnel behavior states, and performing track association on the basis of the personnel behavior states corresponding to the continuous multi-frame video images to obtain personnel behavior tracks; and performing safety production behavior monitoring based on the personnel behavior state and the personnel behavior track, and generating an abnormal behavior early warning signal. According to the method and the device, the real-time performance and the accuracy of safety monitoring in a production scene are improved.

The invention provides a digital evaluation method and system for fair-faced concrete based on Mockup, and belongs to the field of building construction. According to the technical scheme, the method comprises the steps of collecting a surface image of a bare concrete Mock sample plate, extracting a reference Lab color value and a reference texture feature from the collected sample plate image, and constructing an evaluation reference model; collecting a surface image of the to-be-evaluated bare concrete member, and generating image data; performing defect area identification on the image data, extracting an effective image area, and extracting a Lab color value to be evaluated and a texture feature to be evaluated from the effective image area; generating a comparison result based on the evaluation reference model; and based on the comparison result, generating evaluation output information. The method has the beneficial effects that the color and texture double-feature model is constructed, a defect identification and shielding mechanism is introduced, and a structured comparison algorithm and an output system are adopted, so that the automation of the whole process from data acquisition, feature extraction, defect avoidance to intelligent comparison and evaluation output is realized.

The invention relates to the technical field of target detection, in particular to a search and rescue task-oriented unmanned aerial vehicle thermal imaging visual target detection method, which comprises the following steps of: acquiring multiple frames of thermal imaging images of an unmanned aerial vehicle, extracting regional thermal difference characteristics according to a window, marking a non-background region to generate a candidate set, and fitting and reconstructing a suspected thermal target contour. And analyzing the track and the thermal change rate, screening background interference, identifying jump abnormity, positioning the gravity center, and generating a target repositioning signal. According to the method, the heat value range and variance index sequence in the image area is constructed, the thermal anomaly area is judged in combination with the temperature baseline difference, background disturbance comparison is executed in combination with the direction vector of the coordinate trajectory in the multi-frame image and the thermal change parameter, the false detection probability caused by background noise is reduced, and the detection accuracy is improved. The target jump identification is carried out according to the inter-frame heat value and area change rate in linkage with the thermal isoline closure degree, the target discrimination accuracy in a shielding scene is improved, and the robustness of thermal target extraction in a complex search and rescue environment is integrally improved.

The invention relates to the field of machine learning, and discloses a printing quality detection method and device based on machine learning, and the method comprises the steps: obtaining the original image data of a printed matter under the conditions of multiple batches, heterogeneous illumination and multi-angle imaging, and constructing a printing image training set through combining an image region equalization strategy and a light interference shielding mechanism; performing image block-level fine-grained division on the printing image training set, performing multi-channel feature extraction on each image block by adopting a deep convolution model based on a residual attention mechanism, and generating a composite feature map through a feature coupling structure; based on the composite characteristic spectrum, constructing a defect positioning network, and identifying a potential defect response area; and in combination with the initial defect candidate set, introducing an image background disturbance simulation mechanism and a local artifact inversion model, establishing a defect error maintenance positive frame, and eliminating an interference area caused by non-printing errors. The method has the advantage of improving the accuracy of defect identification.

The invention discloses a sludge treatment automatic monitoring method based on computer vision, and relates to the technical field of sludge treatment automatic monitoring, and the method comprises the following steps: collecting sludge image information through computer vision, extracting an edge integrity index and a brightness distribution index, and constructing a fusion judgment function based on the time sequence change of the two indexes, determining whether a thin water film exists on the sludge surface or not according to an output result of the fusion judgment function; under the condition of determining that a thin water film exists on the sludge surface, acquiring a pixel highlight gradient polymerization rate, a texture missing fluctuation frequency and an image saturation nonlinear deviation value of a corresponding image region, and establishing a specular reflection image feature recognition matrix; by constructing a multi-dimensional image feature recognition and dynamic regulation and control mechanism, the problem of image misjudgment caused by thin water film mirror reflection in the sludge dewatering process is solved, and accurate recognition and automatic closed-loop control of the real water-containing state of the sludge are achieved.

The invention provides a textile industry broken yarn identification method and system based on a neural network, and the method comprises the steps: collecting yarn multi-modal data, including a surface image, a fracture sound wave signal and tension change data; preprocessing the data, and extracting an image ROI region, sound wave spectrum features and a tension mutation sequence; performing space-time alignment and feature extraction on the extracted content to obtain a joint feature vector; inputting the broken yarn into a trained broken yarn identification model, wherein the model can identify broken yarn features; judging whether broken yarns exist or not according to the output result and outputting an identification result; and updating the model through online incremental learning. According to the method, multiple types of data are combined, the adaptive preprocessing and feature extraction technology is used, environmental noise is inhibited, key features are focused, and the problem of false alarm of a traditional sensor is solved; multi-modal features are fused through space-time alignment and a self-attention mechanism, and bidirectional LSTM modeling is combined, so that a broken yarn dynamic rule is accurately captured, and the problems of missing detection and delay of manual inspection are avoided.

The invention discloses a data labeling method and system based on user behaviors and attention tracking, and the method comprises the steps: synchronously collecting multi-source behavior signals of a mouse, a keyboard, eye movement and the like of a doctor in real time, combining identity and interface metadata, and carrying out the standardized normalization, abnormality elimination and short time sequence behavior unit division. And extracting individual behavior micro-modes by using unsupervised clustering, and constructing a behavior portrait library. Through multi-modal time sequence modeling and a self-adaptive space-time attention mechanism, behavior characteristics, an interface area and a report text are deeply fused, a multi-level correlation probability is output, and high-precision automatic tagging of content and an image area is realized.

The invention discloses a seal area character recognition method and device and a nonvolatile storage medium. The method comprises the following steps: determining a candidate seal image area in an image according to color information of pixel points in the image; point-by-point sliding convolution processing is carried out on the candidate seal image area through a multi-scale annular convolution kernel group, so that a target seal image area is determined in the candidate seal image area, and the multi-scale annular convolution kernel group comprises a plurality of convolution kernels which are of concentric ring structures and have different radius lengths; mapping the target seal image area into a rectangular expanded image, and identifying and extracting a character image to be identified in the rectangular expanded image; and performing identification processing on the character image to be identified to obtain a seal text corresponding to the target seal image area. The technical problem that the text image processing efficiency is low due to the fact that the text information of the seal area cannot be effectively recognized in the related technology is solved.

The invention discloses a method for identifying small target features in a ray detection image, and relates to the field of image processing, and the method comprises the steps: carrying out the size unification, pixel standardization and normalization preprocessing of an original ray image; constructing a training sample set through image region cutting, and introducing a dynamic sampling strategy to realize positive and negative sample proportion adaptive control; enhancing the number and diversity of small target samples in a training set by using point-shaped and linear artificial defect generation strategies; constructing an image segmentation model and introducing feature jump connection to fuse shallow space and deep semantic information; combining Dice loss and Focal loss to form a composite loss function, and guiding the model to pay attention to a target region with a small area and weak gray level; and finally, a segmentation result is optimized through morphological processing and connected domain analysis, and structured target detection information is output. According to the invention, the recognition accuracy and integrity of the tiny target in the ray image can be effectively improved, and the adaptive capacity of the detection method to the change of the imaging quality is enhanced.

The invention discloses a multi-modal remote sensing visual positioning method and device based on scene knowledge enhancement and a medium, and relates to the technical field of remote sensing visual positioning. The method comprises the following steps: firstly, preprocessing a plurality of remote sensing images, generating scene knowledge enhanced text description, forming a visual positioning data set, and dividing the visual positioning data set into a training set, a verification set and a test set; constructing a visual positioning model, and obtaining an optimal model through training, verification and testing; inputting a to-be-queried text to obtain remote sensing image coordinates. According to the method, cross-modal fusion of knowledge enhancement is realized, scene knowledge is embedded into a visual positioning framework, and the problem of inference of implicit semantics in a remote sensing scene is solved; multi-scale image features and scene knowledge are fused layer by layer through multi-round cross-modal attention iteration, and semantic understanding from coarse granularity to fine granularity is achieved; a similarity threshold is introduced to screen a high-correlation image region, and background interference is reduced in combination with loss constraints. The LLaMA2 is subjected to efficient fine tuning in combination with the LoRA technology, end-to-end coordinate generation is supported, and both performance and calculation efficiency are considered.

The invention discloses an image region analysis method based on entropy driving feature enhancement, and relates to the technical field of image analysis and feature enhancement. According to the method, the local channel information entropy is used as a core feature statistical magnitude, and adaptive weighting and strengthening of different importance region features are realized through explicit quantification of image feature information amount; weight distribution is dynamically adjusted according to the characteristic values, the response of a high-information dense area is remarkably enhanced, and meanwhile low-information and noise interference areas are effectively restrained. On the basis, a cross-layer attention mechanism based on entropy prior is designed, feature statistical information is embedded into a gating and weight generation process, and attention distribution with feature significance as guidance is achieved. According to the method, through an entropy-driven adaptive feature enhancement mechanism, the perception capability, the feature discrimination capability and the analysis precision of the model on the salient region of the image are effectively improved.

The invention relates to the technical field of computers, in particular to an exponential segmentation method based on visual fine-grained semantic-driven cross-modal collaboration, which follows a basic normal form of research in the field of exponential segmentation, and designs a method for carrying out image fine-grained visual enhancement, then carrying out semantic-driven cross-modal collaboration and finally carrying out semantic-driven cross-modal collaboration. And finally, segmenting the mask prediction model. According to an input image, fine-grained visual understanding is enhanced, understanding on a complex spatial position is enhanced, semantic-driven cross-modal collaborative decoding is carried out on the image after fine-grained understanding is enhanced in combination with a text, and then the decoded image is used for final segmentation mask prediction. The method can solve the problems that the correlation between an image area and related language description cannot be fully mined through an existing representative segmentation method, so that fine-grained alignment is insufficient, and a model cannot understand fine-grained correlation such as space and position.

The invention provides a small target detection method and device based on a high-resolution fusion feature map. The method comprises the following steps: dividing an image to be detected into a plurality of overlapped sub-regions and recording position information of the overlapped sub-regions; performing feature extraction and multi-scale fusion on each sub-region to generate a high-resolution fusion feature map; generating geometric parameters and initial confidence of candidate frames based on each spatial position on the image, and performing preliminary screening to obtain a candidate frame set of the sub-regions; mapping candidate frames in all the candidate frame sets to a global coordinate system according to the position information, and combining repeated candidate frames pointing to the same small target to form a global candidate frame set; for each candidate frame in the global candidate frame set, cutting a local image area in the to-be-detected image to carry out fine judgment to obtain a fine trimming confidence coefficient; and fusing the initial confidence coefficient and the refined confidence coefficient to obtain a comprehensive confidence coefficient, and outputting a final small target detection result according to a comparison result of the comprehensive confidence coefficient and an adaptive threshold value. According to the invention, the tiny target in the image can be accurately captured.

The embodiment of the invention relates to the technical field of image processing and positioning, in particular to a processing and positioning method and system for adaptive image recognition, and the method comprises the steps: obtaining a multi-source image data set of a target processing scene; performing image state adaptability analysis processing on the original image sequence to obtain an image recognition feature set of each image region in the original image sequence; performing positioning state mapping processing on the image recognition feature set based on the equipment positioning state data set, and generating a spatial positioning driving feature set corresponding to each image region; and generating a self-adaptive positioning control instruction according to the spatial positioning driving feature set, transmitting the self-adaptive positioning control instruction to an execution end of the processing equipment, and indicating the processing equipment to execute self-adaptive pose positioning control operation for the current image area. Therefore, the flexibility of processing equipment to cope with complex processing scenes is greatly improved, and it is ensured that target processing operation is accurately executed under various observation poses and environmental conditions.

The invention relates to the technical field of ecological environment protection, in particular to a mangrove forest carbon sink monitoring and metering method based on an unmanned aerial vehicle, a radar and an AI technology, and the method comprises the steps: 1, obtaining the laser radar data of a mangrove forest through a laser radar carried by the unmanned aerial vehicle; step 2, acquiring elevation data in a mangrove forest vegetation layer area, and obtaining point cloud data after topographic error correction; 3, obtaining a multispectral image of the mangrove forest, and obtaining a multispectral data matrix; 4, identifying forest growth data features, constructing a mangrove forest carbon sink prediction model, and predicting the mangrove forest carbon sink amount; step 5, marking the image region with the NDVI value higher than a preset NDVI threshold value as a blade over-dense region; and according to the area of the overdense leaf region and the multispectral data matrix, calculating a light depression factor by using a photosynthetic depression factor formula, determining the carbon sink deviation of the overdense leaf region by using a regional carbon sink deviation formula, and obtaining a real carbon sink value of the mangrove forest according to a carbon sink calculated value obtained by prediction.

The invention discloses an open vocabulary target detection method and device, equipment and a storage medium. Comprising the following steps: inputting image data into a visual feature extraction module to obtain multi-scale visual features; obtaining a text embedding vector generated by precoding, wherein the text embedding vector is generated by a text encoder based on offset fine tuning; inputting the multi-scale visual features and the text embedding vectors into a sparse hybrid expert detection head module, and outputting sparse hybrid features; and obtaining a target detection result based on the sparse mixed features. In the reasoning stage, all types of texts can be pre-coded into static embedded vectors, the static embedded vectors are directly input to the detection process, complete text branches do not need to be operated, and the reasoning speed and the resource efficiency are greatly improved. And a sparse hybrid expert detection head is provided, which can dynamically select and activate the most adaptive expert sub-network according to the characteristics of the image region, thereby improving the spatial adaptability and modeling ability of the model, and maintaining the high efficiency and low delay of reasoning.

An encoder includes circuitry and memory coupled to the circuitry. In operation, the circuitry encodes subpicture information in which a horizontal position and a vertical position of a region of a subpicture are represented in a unit of a coding tree unit (CTU). The subpicture is a rectangular region in a picture. The horizontal position is represented by a first position of a first CTU in the subpicture, the first position being relative to a left end of the picture. The vertical position is represented by a second position of a second CTU in the subpicture, the second position being relative to a top end of the picture.

An image processing apparatus includes: an image obtaining unit configured to read a print product in which a print image is printed on a pre-printed sheet and generate an inspection target image of the print product; a first alignment unit configured to generate an alignment image by aligning the inspection target image as a whole with a reference image by means of projection transformation, the reference image indicating a correct image of the inspection target image; and a second alignment unit configured to perform alignment by means of non-rigid registration in each of local regions of the alignment image. The second alignment unit uses at least one of a print image region surrounding the print image and a pre-printed image region surrounding a pre-printed image printed on the pre-printed sheet before printing of the print image in the alignment image as the local regions.

The application discloses a kind of detection and defense method of counterpatch, respectively based on abnormal positioning and edge detection.The detection method of counterpatch based on abnormal positioning utilizes clean image to train the generator-adversary network of encoder-decoder structure;The output image is obtained by inputting the image to be detected into the generator-adversary network, and the absolute error is obtained by subtracting and taking absolute value;The image region whose absolute error is greater than error threshold is the region where counterpatch is located;The detection method of counterpatch based on edge detection converts the image to be detected into gray scale image and carries out edge detection, and obtains edge image;The edge lines in edge image are connected into a closed area one by one, and the closed area whose area is less than the area of counterpatch region is the region where counterpatch is located.The application can detect counterpatch based on the two schemes of abnormal positioning and edge detection respectively, and blacken the region or use image restoration algorithm to restore the region to defend counterpatch.

The invention discloses a plastic product surface defect detection method based on reinforcement learning. The method comprises the following steps: obtaining surface image preprocessing and dividing local image areas; inputting the image into an image encoder, extracting feature representation, and performing normalization processing; inputting a potential state modeling module, and introducing a variational inference mechanism to optimize a modeling process; multi-step forward prediction is carried out, an uncertainty evaluation module is input to calculate an information entropy value, and average processing is carried out; inputting a behavior strategy network, and selecting and outputting a target area judgment action through a heuristic exploration mechanism; carrying out defect identification and recording a defect judgment result; comparing actual defect labels, calculating a prediction error value and updating parameters of the model and the network; and repeatedly executing the steps until an optimal behavior strategy network is obtained, and outputting a final defect judgment result set. According to the method, reinforcement learning and potential state modeling are fused, intelligent detection of plastic surface defects is realized, and the method has high precision, self-adaption and strong generalization ability.

The invention discloses a film processing quality evaluation method and system, and the method comprises the steps: obtaining and preprocessing a processed film collection image, and obtaining a film gray image; determining a gray value of each pixel point in the film gray image, and performing clustering analysis on the pixel points based on the gray value to obtain a plurality of abnormal image areas; determining multi-dimensional image features in each abnormal image region, and analyzing abnormal conditions of each abnormal image region based on the multi-dimensional image features to obtain an abnormal coefficient; determining the number of the abnormal image areas, and determining a comprehensive quality evaluation value of film processing based on the number of the abnormal image areas and the abnormal coefficient; and determining a quality evaluation grade of thin film processing based on the comprehensive quality evaluation value, and evaluating the processing quality of the thin film according to the quality evaluation grade. According to the invention, comprehensive quantitative evaluation of the processing quality of the film can be realized, the production efficiency of the film is improved, the quality risk is reduced, and the production of high-quality film products meeting standards and customer requirements is ensured.

The invention discloses a self-supervised monocular depth estimation method based on collaborative alignment, which relates to the technical field of image processing, and comprises the following steps: extracting global semantic features and local geometric features of an input image in parallel, and fusing the global semantic features and the local geometric features into a multi-scale visual feature map; based on the depth gradient prior of the structured scene, sampling a preset number of ordered image areas on the multi-scale visual feature map along the depth change direction; processing the depth sequencing text prompts through a text encoder with a fixed depth text, and generating corresponding text depth embedding; forcibly constraining that the self-similarity of near-region features is higher than that of far-region features in a visual feature space, and deeply embedding the visual features of the aligned regions and matched texts; after the fixed depth text is replaced with the learnable depth mark, dynamic depth embedding is generated and aligned with the dynamic depth embedding; and predicting a pixel-level depth map corresponding to the input image through the optimized model. According to the method, labeling dependence is avoided, and the problem of repeated labeling is solved.

The invention relates to the technical field of film coating dynamic scheduling, in particular to a photovoltaic cleaning film coating dynamic scheduling system based on digital twinning. The system comprises a coating control scheduling module, a coating regional processing module, a database storage module and a coating state acquisition module. The film coating regional processing module carries out data twinborn simulation in combination with state information, carries out comparison in combination with stored defect data, simulates and positions the current film coating state of a photovoltaic panel, generates control information, and feeds back the control information to the film coating control scheduling module to regulate and control photovoltaic processing equipment to carry out secondary scheduling control. Regional defect identification is carried out on the collected photovoltaic panel coating simulation image by utilizing regional division and unit region evaluation, targeted defect processing is carried out on the identified unit region in cooperation with processing data stored in the database storage module, the defect type and position are accurately positioned, and refined defect identification and processing are realized.

The invention discloses a multi-modal document data processing method and system for large language model training, and the method comprises the steps: receiving a plurality of original documents in various formats, extracting the structure information of each original document, and recognizing a text region and an image region of each original document based on the structure information; performing optical character recognition on the text region and the image region by adopting a parallel OCR (Optical Character Recognition) engine based on GPU (Graphics Processing Unit) acceleration and heterogeneous calculation to generate recognition text data of the corresponding original document; performing multi-dimensional quality evaluation and cleaning on the recognition text data of each original document, and outputting normalized text data; and storing the standardized text data into a distributed knowledge base according to a predefined structure, and performing copyright and compliance test on the standardized text data. By adopting a parallel OCR recognition engine based on GPU acceleration and heterogeneous calculation, efficient and high-precision batch processing of multi-modal documents is realized, and the processing speed, the recognition accuracy and the data quality are improved.

The method is mainly applied to the technical field of artificial intelligence processing. The invention discloses a face and license plate privacy protection video processing method, device, equipment and medium, and the method comprises the steps: extracting input video data to obtain a video frame sequence which comprises a plurality of video frames; the video frame sequence is input to a pre-trained face and license plate detection model for target detection, so that a target video frame containing a target image area is determined, position information of the target image area in the target video frame is obtained, and the target image area is a face area or a license plate area; based on the position information of the target image area, performing fuzzy processing on the target image area in each target video frame; and generating a video after fuzzy processing based on the video frame after fuzzy processing. According to the method and the device, the privacy protection and the target detection are collaboratively optimized, so that the efficiency and the accuracy of target detection are improved while the privacy security is guaranteed.

The invention relates to the technical field of image recognition analysis, in particular to a logistics image intelligent recognition method and device, equipment and a storage medium. The logistics image intelligent identification method comprises the following steps of obtaining an intelligent judgment result and a judgment result subjected to manual correction, and constructing an error correction data set; optimizing a pre-constructed optical character recognition module, a semantic rule base and an NLP model according to the error correction data set; extracting image features of the logistics image through a pre-trained visual model; the optimized optical character recognition module is adopted to extract the text information of the logistics image, and a text image region mapping table is established according to the image features and the text information; and analyzing the text information by adopting the optimized semantic rule base and the NLP model in combination with the text image region mapping table to obtain a classification result. According to the method, the texts can be accurately classified, the ambiguity problem of general OCR is avoided through semantic rule and model fusion, and file audit of an arbitration scene can be efficiently processed.