35 results about "Gaussian convolution" patented technology

The invention discloses a normalized stream-based super-resolution industrial image anomaly detection and positioning system. The system comprises an information cascade extension module, a multi-core feature weighting module and a probability distribution-based normalized stream module, the information cascading expansion module integrates an image super-resolution technology and Gaussian convolution processing and aims to capture fine texture information and macrostructure information of an image at the same time, and the multi-core feature weighting module performs intelligent weighting on features by fusing complementary information from different visual angles, so that the image quality is improved. The method effectively enhances the distinguishing capability of the model for a normal region and an abnormal region, and can accurately predict the specific position of the abnormality through deep learning of the probability distribution of global and local features based on a probability distribution normalization flow module. According to the method, the problems of fuzzy industrial image imaging, noise pollution, fuzzy boundaries of a normal region and an abnormal region and the like are solved, and the industrial anomaly detection accuracy and the positioning precision are effectively improved.

The invention provides a deep water gravity flow gas field reservoir water saturation earthquake prediction method based on TDCN-TransNet. The method comprises the following specific steps: a TDCN-TransNet network architecture comprises an input processing layer, a dual TWT position code, a TDCN module, a Transformer encoder layer and an output layer; mapping the features to a hidden dimension in an input processing layer through linear full connection; a dynamic Gaussian convolution kernel and a CBAM attention mechanism are embedded in the TDCN, and modeling of a nonlinear relation between an elastic parameter and saturation is enhanced through multi-scale time sequence feature extraction; tCBlock is integrated on a Transform encoder layer, and long-distance sequence dependence is captured by using attention pooling and an SE channel excitation mechanism; a TWT position code is introduced between the TDCN and the Transform, and sequence position information is enhanced; constructing a combined loss function of a mean square error and L2 regularization; finally, training and reasoning are conducted based on the model, and high-precision water saturation prediction is achieved. The method provided by the invention has high generalization and accuracy, and shows excellent prediction performance in a complex geological environment of a deepwater gravity flow reservoir.

The present application relates to a kind of solar filament fine structure enhancement method based on line Gauss filter, belong to image enhancement field.To overcome the current H-alpha image in solar filament fiber because of low contrast, leading to artificial marking difficult, the problem that the fiber attribute information is not accurate enough is extracted, the present application proposes a kind of image enhancement method with line Gauss convolution as core.Firstly, H-alpha image is normalized to preliminary stretch image contrast, while suppressing noise using Laplace-Gauss operator to enhance edge, then using line Gauss filter to enhance the contrast of solar filament fiber.In post-processing stage, using the adaptive histogram equalization of restriction contrast and top hat, bottom hat transformation method to improve the uneven problem of line Gauss filter enhancement effect.Finally, line Gauss filter is used again to further improve the contrast of solar filament fiber.After the enhancement of the method, the filamentary structure of solar filament fiber is clear, which makes it possible to objectively and accurately measure the filamentary structure attribute information of solar filament.

The invention relates to an intelligent restoration method for handwriting archive fonts, and the method comprises the steps: decomposing a handwriting archive target image, obtaining channel images corresponding to all color channels, for each channel image, determining a character enhancement coefficient of each pixel point in a character region based on the image feature value of each pixel point in the character region of the channel image, and restoring the character enhancement coefficient of each pixel point in the character region. And then, based on each character enhancement coefficient, correcting a preset Gaussian convolution scale parameter to obtain a target Gaussian convolution scale parameter corresponding to each pixel point in each channel image, and based on a Retinex algorithm and each target Gaussian convolution scale parameter, carrying out character enhancement processing on the corresponding channel image. According to the method, channel images subjected to character enhancement processing are obtained, the channel images subjected to character enhancement processing are combined, the character enhancement image after the target image of the handwritten file is restored is obtained, and restoration of the target image of the handwritten file is achieved.

The present invention relates to a method for detecting and correcting cutterhead images under uneven illumination within a shield machine. The method comprises: first, inputting an unevenly illuminated image under low illumination, converting the image from RGB to HSV space, and performing continuous convolution operations on the extracted illumination components using a small-scale Gaussian convolution kernel to obtain the illumination component distribution of the scene; second, determining the gamma value of each pixel position in combination with the local contrast and gradient magnitude, dynamically adjusting the gamma value according to the local features of the image, and converting the corrected image from HSV to RGB space; finally, performing a weighted adaptive histogram equalization algorithm on the image, performing guided filtering to enhance the image, and outputting the image. The method has a significant effect on correcting cutterhead images in unevenly illuminated areas within the dim space of a shield machine, improving local contrast, taking into account both the global characteristics and local detail information of the cutterhead image, suppressing the influence of noise, and making the cutterhead image enhancement result clearer and more accurate.

This invention discloses a visual-tactile fusion multimodal tactile sensor and a three-dimensional pressure field reconstruction method, relating to the technical field of tactile perception. The sensor includes a visual acquisition unit, a tactile acquisition unit, a gel component, a data processing unit, and a display and interaction unit, achieving dual-modal collaborative perception. The three-dimensional pressure field reconstruction method sequentially executes steps such as force system initialization, synchronous acquisition, image preprocessing, pressure calibration calculation, cross-modal registration, dynamic Gaussian convolution fusion, 2D-to-3D mapping, point cloud and mesh reconstruction, real-time display and interaction, data recording, and reset maintenance. It utilizes dynamic Gaussian convolution kernels to simulate pressure diffusion and generates a three-dimensional pressure field model based on the Open3D library. This invention reduces the cost of the gel component, and the reconstruction delay of the three-dimensional pressure field does not exceed 8 milliseconds, making it suitable for fields such as robot operation, intelligent prosthetics, and human-computer interaction.

The application discloses a spraying robot nozzle flow adjustment method based on a Gaussian convolution kernel and reinforcement learning, and comprises the following steps: S1, measuring two-dimensional plane data of the distance between a spraying robot nozzle and an unsprayed wall surface; S2, defining a nozzle flow two-dimensional distribution function; S3, defining a two-dimensional Gaussian function as a convolution kernel; S4, weighting the convolution kernel by using the nozzle flow function, and performing convolution operation on the distance data and the two-dimensional Gaussian function; S5, calculating the roughness of the wall surface after spraying according to the convolution operation result; S6, optimizing the nozzle flow function by using reinforcement learning according to the roughness, and stopping the optimization when the roughness reaches the requirement; and S7, outputting the nozzle flow function result and spraying the wall surface according to the result. The method of convolution calculation and reinforcement learning is used to adjust the nozzle flow of the spraying robot, the problem of 'fixation and lagging adjustment' of the nozzle flow control can be solved, and thus good spraying effect is realized.

The embodiment of the invention relates to the technical field of laser radars, and discloses a centroid extraction method, a laser radar, a robot and a storage medium, and the centroid extraction method comprises the steps: firstly obtaining an original echo signal curve outputted after the laser radar detects a target object, dividing the original echo signal curve into N continuous signal section curves, for each signal section curve, determining an asymmetric Gaussian convolution kernel matched with the signal section curve, and then performing convolution operation on the asymmetric Gaussian convolution kernel and an original energy value in the signal section curve to obtain a first target energy value corresponding to the signal section curve, and so on until M first target energy values are obtained, and finally, extracting the centroid according to the M first target energy values and the corresponding pixel positions. According to the method, the asymmetric Gaussian convolution kernel is adopted for convolution operation, the local form of the signal is precisely fitted, the asymmetric Gaussian convolution kernel is differentially matched, the accuracy of centroid extraction is improved, and then the ranging accuracy of the laser radar is improved.

The invention discloses an online detection method for trace impurities in a benzene product, and belongs to the field of online analysis of trace impurities in high-purity chemical organic products. The method comprises the following steps: collecting Raman spectrums of a benzene product mixture with known content and a pure substance benzene component based on the same measurement condition; carrying out pretreatment on the collected Raman spectrum; performing spectrum peak decomposition based on a Lorentz Gaussian convolution model on the preprocessed Raman spectrum to obtain Raman spectrum signals of the benzene component, the toluene and the non-aromatic hydrocarbon component; constructing a quantitative analysis model based on the effective peak height of the impurity peak and the relative concentration of the impurity; and calculating the mass percent of the impurities in the to-be-detected sample through the quantitative analysis model. The method effectively solves the problem of mutual overlapping of benzene peaks and impurity component signals in the Raman spectrum by utilizing a spectrum analysis technology, can remarkably improve the detection precision of trace impurities and reduce the detection limit, and has the advantages of low demand quantity of training samples, strong model extrapolation, short analysis time, accurate prediction result and the like.

The invention relates to the technical field of radar data processing, in particular to a group initiation method based on cyclic random Hough transformation and DBSCAN (Density-Based Spatial Clustering of Applications with Noise), which comprises the following steps of: S1, acquiring a set of adjacent three frames of echo trace points of a radar, and executing random Hough transformation on the first two frames of trace points in the set of the adjacent three frames of echo trace points to form a characteristic parameter matrix; s2, performing dimension accumulation and Gaussian convolution condensation on the characteristic parameter matrix, extracting trace points corresponding to peak values according to a condensation result to form a temporary group, and updating a trace point set; and S3, performing DBSCAN spatial clustering on the temporary group to obtain a clustering result, and performing group track confirmation on the clustering result through a third frame echo point. The method not only can solve the problem of track initiation of the dense group, but also has relatively high clutter resistance, is small in calculation amount, and has relatively high engineering application value.

The present invention belongs to the field of image enhancement technology, and particularly relates to an infrared image enhancement method based on circular convolution and multi-clustering space, comprising the following steps: S1, establishing a multi-scale circular convolution module, establishing a multi-scale convolution module with a plurality of Gaussian convolution kernels having different radii, and connecting multiple multi-scale convolution modules in a densely connected manner to process image details; S2, using the K-means algorithm to perform clustering analysis on the image gray values, and selecting the initial clustering center by the maximum contribution rate method; S3, using a two-dimensional Gaussian function and a rotation matrix to construct the kernel function at the i th angle as the template kernel function to extract the contour information of the image at different angles. Compared with the existing image enhancement methods, the method of this embodiment outputs an image with clearer contours and richer detail information, and the quality of the original image has been significantly improved in terms of details, brightness, etc.

The invention discloses a visual sense and tactile sense fusion multi-mode tactile sensor and a three-dimensional pressure field reconstruction method, and relates to the technical field of tactile sensing, the sensor comprises a visual sense acquisition unit, a tactile sense acquisition unit, a gel assembly, a data processing unit and a display and interaction unit, and bimodal collaborative sensing is realized; in the three-dimensional pressure field reconstruction method, the steps of force system initialization, synchronous acquisition, image preprocessing, pressure calibration solution, cross-modal registration, dynamic Gaussian convolution fusion, two-dimensional to three-dimensional mapping, point cloud and grid reconstruction, real-time display and interaction, data recording, reset maintenance and the like are executed in sequence. Simulating pressure diffusion by using a dynamic Gaussian convolution kernel, and generating a three-dimensional pressure field model based on an Open3D library; the cost of the gel assembly is reduced, the reconstruction delay of the three-dimensional pressure field does not exceed 8 milliseconds, and the method is suitable for the fields of robot operation, intelligent prostheses, man-machine interaction and the like.

The embodiment of the invention provides a long-distance dependent image feature extraction method and related equipment, and belongs to the technical field of computer vision and deep learning. The method comprises the following steps: compressing an input feature map; on the basis of the compressed feature map, independently predicting a ternary parameter () for controlling the scale and direction of a Gaussian kernel for each pixel position; generating a self-adaptive two-dimensional Gaussian convolution kernel covering a whole image range for each pixel according to the parameters; global weighted summation is carried out on the convolution kernel and the same-channel full-image features to obtain a convolution response containing a long-distance dependency relationship; and finally, recovering the feature dimension and fusing the feature dimension with the input feature residual error. According to the method, the Gaussian kernel with the adjustable direction and scale is dynamically generated at the pixel level, the local receptive field limitation of traditional convolution is broken through, long-distance semantic dependency and directional structural features in the image can be efficiently and explicitly modeled, and the low calculation complexity is kept while the visual task performance is remarkably improved.

This application proposes an image correction method, apparatus, device, and storage medium based on local Gaussian convolution. It acquires regional images of the same area on several consecutive grains, divides each regional image into image blocks corresponding to various positions, determines the sharpness index of each image block to determine template image blocks, and uses local image blocks as processing units. Determining the template image blocks for correction based on the sharpness index helps avoid noise amplification. Based on the template image blocks, the Gaussian kernels of all other image blocks at the same position are optimized to obtain target Gaussian kernels corresponding to all other image blocks. Convolution is then performed on the corresponding other image blocks using the target Gaussian kernels, which helps to determine accurate image block convolution results with low overhead. Finally, the corrected image corresponding to each regional image is obtained through fusion and reconstruction based on each template image block and the image block convolution results, reducing noise and providing a high-quality corrected image for subsequent processing.

This invention discloses a polyp segmentation system based on multi-scale feature fusion and edge awareness, belonging to the field of polyp segmentation technology. The system first extracts features from intestinal polyp images at multiple scales, from deep to shallow, forming feature maps with at least four levels of resolution, laying the foundation for subsequent multi-level representation. Then, multi-scale feature refinement is performed on the higher-resolution feature maps, introducing a multi-receptive-field structure to enhance the expression of local structure and texture details, while retaining the lowest-resolution features as the starting point for fusion. Furthermore, cross-scale feature fusion is performed in ascending order of resolution, achieving progressive supplementation of semantic features at different scales. Finally, edge aggregation processing using unified channels, unified sizes, and fixed Gaussian convolution kernels strengthens the consistency of structural boundary region representation. The overall scheme has the advantages of clear structural hierarchy, full feature utilization, and strong ability to express polyp edge details.

The present invention discloses a method for predicting a load sequence of a thermal power unit using an AGC prediction signal, comprising the following steps: S1, collecting parameter features and target features of the thermal power unit; S2, inputting the parameter features into a unit load curve prediction model; S3, first mapping the parameter features through two feature mapping layers; S4, then performing convolution processing on the parameter features using a one-dimensional convolutional neural network with a Gaussian convolution kernel to extract the inertial features of the data; S5, normalizing the target features and inputting them into the unit load curve prediction model; S6, first processing the data features using two weighted layers; S7, then performing feature merging; S8, further merging the inertial features extracted in step S4 with the data features obtained by merging the features in step S7 to obtain a thermal power unit operating power prediction sequence and curve. The present invention solves the problem that existing thermal power unit load prediction methods only consider internal factors of the thermal power unit but fail to consider external factors of the unit.

The invention provides a lightweight wood defect segmentation method based on multi-dimensional boundary perception and guidance, and the method comprises the steps: obtaining an RGB image of a wood surface defect, extracting an initial boundary through a Laplacian operator based on a pixel-level label, and replacing an initial boundary neighborhood value with a Gaussian convolution kernel to generate a Gaussian boundary label; inputting the image into a Segform-based coding-decoding network, and outputting a multi-scale fusion feature map through decoupling features of a space attention flow and a channel attention flow; based on a multi-scale fusion feature map, outputting a defect segmentation map through a segmentation decoding network, and predicting a boundary probability map through a boundary supervision branch; calculating segmentation loss by adopting cross entropy loss and Dice loss, and calculating boundary prediction loss in combination with a Gaussian boundary label and a cosine similarity function; performing model optimization by combining boundary prediction loss, segmentation loss and auxiliary loss, and balancing boundary positioning and segmentation precision through weighted summation; and iteratively executing a boundary prediction and segmentation process, and outputting a defect category segmentation map in real time.

The present invention discloses an improved MSR subway wheel tread image enhancement method, device and storage medium, and relates to the field of subway vehicle detection technology. The present invention first adaptively calculates the Gaussian convolution kernel size based on the grayscale statistics based on the overall grayscale mean and grayscale distribution of the image. Secondly, the convolution calculation method is optimized, and according to the light intensity distribution characteristics of the point light source in the plane, the light compensation amount of the current position is added to the single convolution calculation. At the same time, the direction information based on the current position is introduced into the light compensation amount, and the direction compensation values ​​are added up, down, left and right to form an asymmetric Gaussian convolution kernel. Finally, when the final reflected image is obtained, the image is linearly stretched in intervals, the highlight and low-brightness areas are suppressed, and the grayscale range of some target components is gamma-corrected in intervals to highlight the area of ​​interest, thereby solving the grayscale interval compression problem.

The embodiment of the invention provides an image generation method and device, electronic equipment and a storage medium, and relates to the technical field of image denoising. The method comprises the following steps: acquiring a to-be-processed image in an original RAW format; based on an interpolation algorithm, converting the to-be-processed image into a standard red-green-blue sRGB format to obtain a to-be-blurred image; for each color channel of the to-be-blurred image, performing Gaussian convolution blurring on the pixel value of each pixel position in the to-be-blurred image under the color channel to obtain a blurred pixel value of each pixel position for the color channel; for each pixel position, acquiring a fuzzy pixel value of the pixel position for the corresponding to-be-utilized color channel to obtain a fuzzy RAW image; the to-be-utilized color channel corresponding to one pixel position is a color channel at the pixel position in the to-be-processed image; and adding noise into the fuzzy RAW image to obtain a noise RAW image. The noise distribution in the noise RAW image is closer to the real noise distribution.

The application discloses a distortion image correction method and device, electronic equipment and storage medium, and relates to the field of biological identification, wherein the correction method comprises the following steps: acquiring the pixel coordinates of each pixel point in a target distortion image; inputting the pixel coordinates of each pixel point into an image correction model, outputting the initial correction coordinates corresponding to each pixel point, and calculating the deformation convolution kernel size based on the initial correction coordinates; inputting each pixel coordinate into a Jacobian calculation formula to obtain the Jacobian value of each pixel coordinate, and performing inverse mapping calculation on the Jacobian value to obtain the inverse-mapped Jacobian value; calculating the Gaussian convolution kernel of each pixel point; inputting the Gaussian convolution kernel into an anti-aliasing correction model to output the optimized correction coordinates of each pixel point, and obtaining a target correction image. The application solves the technical problem that, in the related art, the distortion correction effect is poor when a distorted image is corrected, and high-precision image analysis and application cannot be met.

The application discloses an automatic feature extraction method suitable for battery state of health (SOH) evaluation and life prediction, and comprises the following steps: S1, performing a cyclic charge-discharge aging test on the battery under a fixed temperature and different charge-discharge rates, and collecting current information and voltage information of the battery in the cyclic process; S2, obtaining the capacity corresponding to the cycle through ampere-hour integration under different charge-discharge rates, and establishing a capacity increment curve corresponding to the working condition; S3, performing Gaussian convolution on the IC curve under a one-dimensional time scale, and establishing a Gaussian space and a Gaussian difference space corresponding to the working condition; S4, constructing towers of different scales, calculating the Gaussian difference space of each tower, and finding local extreme points in adjacent difference spaces; and S5, processing the automatic feature database, selecting similar points as the same sequence, and performing correlation analysis on the battery SOH, and selecting a feature sequence with high correlation as the automatic feature. The method has low calculation complexity, and can directly obtain a feature point sequence with high correlation with the SOH.

The present invention provides a method, device, medium and equipment for detecting laser dead pixels. The method includes: acquiring a luminous point image of a transmitting laser, determining target pixels in the luminous point image, and obtaining a to-be-detected image according to the target pixels; creating a Gaussian convolution kernel model, performing convolution processing on the to-be-detected image based on the Gaussian convolution model to obtain a feature map; performing non-maximum suppression processing on the feature map to determine a plurality of reference local maximum values; screening the plurality of reference local maximum values to obtain a target maximum value; the pixel point corresponding to the target maximum value is a dead pixel; thus, regardless of the model and type of the transmitting laser, as long as the luminous point image of the transmitting laser is acquired, the Gaussian convolution kernel model can be used to determine the dead pixel; this testing method has good versatility and can improve the testing efficiency.

Disclosed is a Gabor wavelet-fused multi-scale local level set ultrasonic image segmentation method. In the method, non-uniformity of the grayscale of an ultrasonic image is taken as a texture having cluttered directions, the multi-directional property of Gabor wavelets is used to process the image, and intermediate images in different filtering directions are fused by taking maximum values, so as to obtain an intermediate image having a weakened texture effect and an enhanced difference between a foreground and a background. For the feature of a weak edge of an ultrasonic image, a concept of multi-scale is used to improve the conventional LIC method, Gaussian convolution kernels having different variances are set, and a final edge is obtained by means of average fusion.

The invention discloses a lattice-based two-party collaborative signature generation method and system based on Gaussian convolution, and the method comprises the steps of system initialization, key generation, collaborative signature and verification, and employs the Gaussian convolution technology to avoid restarting, effectively prevent private key information leakage, and improve the collaborative signature efficiency. The method has the advantages of being high in safety, complete in function, high in operation efficiency and the like, has a great promotion effect on application of the anti-quantum cryptography algorithm, and is suitable for various application scenes such as block chains and PKI systems under the background of distributed key management.

The invention discloses a wide-width different-source satellite image registration method and system based on random block optimization, and the method comprises the steps: carrying out the division of overlapped subblocks, carrying out the calculation of edge feature information entropy, dynamically dividing an image subblock into a strong information block and a weak information block, and providing a structural feature basis for the subsequent matching. According to the strong and weak information mutual feedback random sampling matching strategy, strong and weak information blocks are randomly sorted and alternately extracted, a strong and weak information alternate matching sequence is constructed, uniform distribution of homonymy point pairs is effectively guaranteed, and the local optimum problem is avoided. And constructing an efficient multidirectional feature descriptor: proposing a description method based on a joint multidirectional feature tensor, and remarkably improving the adaptability to the nonlinear radiation difference of the heterogeneous image through gradient magnitude calculation, high-dimensional multidirectional decomposition, three-dimensional Gaussian convolution and normalization processing. And fast Fourier transform similarity measurement: the feature matching process is accelerated by using frequency domain conversion, and the calculation efficiency is greatly improved. The method has high algorithm efficiency and reliable matching success rate.

The application discloses a mobile part detection system and method using single-vision and galvanometer cooperation, belongs to the technical field of part detection, and comprises a galvanometer module, a visual sensing module, a motion control module and a processing module. The galvanometer module comprises a rotating motor and a galvanometer, the galvanometer is connected with the rotating motor and rotates under the driving of the rotating motor, the visual sensing module comprises an industrial camera, the industrial camera is used for shooting through the galvanometer on the rotating motor, and the image of a measured part in a moving state on a conveying belt can be collected. The application can acquire a relatively complete image of the part by tracking and collecting the part image multiple times through single-vision technology and the galvanometer, can better perform edge processing on the part by performing gray processing and Gaussian convolution processing on the collected part image and then performing edge detection on the part image after convolution, and can more accurately judge whether the part has defects.

The application discloses an orthodontic diagnosis report automatic generation method based on image and text alignment. First, the orthodontic lateral film is pretreated, and the image features are extracted by using Gaussian convolution. At the same time, the state space model is used to model and extract features of the orthodontic report. Then, the image and text features are aligned through the cross attention mechanism, so that the image feature extraction module can more accurately identify the semantic information related to the image, thereby closely associating the image features with the corresponding language description. Finally, the extracted orthodontic knowledge features are input into the text decoder to generate high-quality orthodontic diagnosis reports. The application significantly improves the accuracy and efficiency of the diagnosis report generation, and solves the problems of low quality and low efficiency in the traditional manual diagnosis process.

The application relates to the technical field of line laser center line extraction, and discloses a method for extracting a line laser center line of a shoe outsole surface, which comprises the following steps: collecting an original image of the shoe outsole, extracting a green channel in the original image, and then performing a Gaussian convolution operation to obtain a pretreated image; then, quantile of each column is calculated, and whether the position is a laser stripe is judged according to whether the quantile of each column tends to be normally distributed, so that the line laser center line of the shoe outsole is obtained. The application can accurately and quickly judge different regions such as a bright area and a dark area of the laser stripe, and solves the problem that the light intensity threshold method cannot accurately determine the position of the laser stripe of an irregular and rough object.

The present invention relates to an image processing method and image processing chip. By converting the process of processing a high-resolution image once into the process of processing multiple low-resolution images, there is no limit on the original image resolution and it can adapt to images of various resolutions. In this way, a wide image is divided into multiple blocks of smaller images. During the calculation of the Gaussian convolution of each small image block, the width of each row of image stored in the intermediate calculation process becomes smaller, thereby significantly reducing the hardware circuit area of ​​the image processing chip and ensuring the calculation accuracy of the Gaussian convolution filter. In addition, by setting a multiplexing area, the method ensures the continuous and uninterrupted search of extreme points at different block boundaries, achieving seamless connection between the extreme point searches of adjacent blocks, and ensuring the continuity and accuracy of data processing.

The present invention provides a light spot extraction method based on a non-uniform scale Hessian matrix. The method can ensure the accuracy of the light spot center while performing sub-pixel extraction of the light spot center, thereby meeting the requirements of precise image processing. The method comprises the following steps: S1, obtaining a non-uniform light spot image, binarizing the light spot image, and determining an initial scale value of a Gaussian convolution kernel; S2, establishing a Gaussian convolution template of a non-uniform scale according to the initial scale value; S3, constructing a determinant operator of a normalized Hessian matrix according to the Gaussian convolution template; S4, obtaining the Gaussian template scale when the determinant operator value is maximum, and using the image point coordinates at the Gaussian template scale as the pixel-level coordinates of the light spot center; and S5, performing sub-pixel center extraction in the obtained light spot image center area according to the pixel-level coordinates of the light spot center, thereby obtaining the sub-pixel coordinate position of the light spot center.