69 results about "Gaussian blur" patented technology

The invention relates to the technical field of image recognition, in particular to a packaging film defect detection system based on image recognition, which comprises a candidate region extraction module, a form enhancement module, a feature extraction module, a fuzzy evaluation module and a defect judgment module. According to the method, a self-adaptive binarization value is generated by calculating a pixel local neighborhood gray average value and a standard deviation, background texture noise and illumination non-uniform interference are effectively suppressed by using a dynamic threshold, a defect entity contour is reconstructed by combining polygonal broken line approximation and morphological filling operation, key geometric angular points are reserved, and internal fracture textures are repaired. Constructing a multi-dimensional feature vector set covering a texture energy entropy value and illumination uniformity, deeply analyzing a defect surface microscopic optical structure, introducing a Gaussian fuzzy membership function and an information entropy weight distribution mechanism, objectively quantifying the affiliation probability of feature components to different defect categories, and determining the defect surface microscopic optical structure. The limitation of a single rigid criterion is broken through, and accurate classification of fuzzy boundary defects is realized.

The invention discloses a robot imitation learning method based on a diffusion model, and belongs to the technical field of robot learning and body intelligence. Comprising the following steps: carrying out standardization processing on input image observation data and a robot state, and carrying out image enhancement by adopting gray scale transformation, random erasure and Gaussian blur; based on the enhanced image and the robot state, training a diffusion model to generate an action sequence, including forward diffusion, conditional feature fusion, noise prediction and joint loss optimization; in real-time control, features are extracted through an image enhancement network, sampling is accelerated through DDIM to generate an action sequence, noise scheduling coefficients are dynamically adjusted based on visual feature differences, and closed-loop optimization is achieved. According to the method, the image enhancement technology and the diffusion strategy are deeply fused, joint optimization of visual features and action sequences is achieved, the action generation accuracy and robustness of the robot under the complex visual interference and small sample conditions are remarkably improved, and the success rate is improved by 18% compared with a base line under 40 sample sizes.

The invention provides an e-commerce commodity graph differentiation generation method and device, equipment and a medium, and the method comprises the steps: obtaining an original to-be-shelved commodity graph of e-commerce, and finally obtaining the disassembly information; the method comprises the following steps: adjusting an original commodity image into 980 * 980 pixels, performing Gaussian blur processing at the same time, and then converting image pixel data into a Base64 coded character string; constructing a composite cue word according to the image data, the disassembly information and a format required by generation; inputting the composite cue word into a model to generate a new scene description; inputting the new scene description, the original commodity graph and the synthesis requirement into a model to generate a synthesis cue word, and then inputting the original commodity and the synthesis cue word into an image-text model to generate a differentiated commodity graph; setting the output resolution, the number of iterations and the learning rate of the SuperIR model, and inputting the differentiated commodity graph into the SuperIR model to generate a differentiated commodity optimization graph; the defects of low efficiency, poor scene adaptation, main body distortion and difficulty in repeated avoidance of a traditional method are overcome.

The application discloses a GAN neural network multiple distortion suppression model based on a coordinate attention mechanism, obtains an original neutron radiographic image from a neutron source; proposes a brand-new neutron radiographic image degradation model; adds Gaussian blur, Gaussian noise, Poisson noise and noise obeying gamma distribution to the clear neutron radiographic image; uses real white spot noise to train the GAN neural network, simulates white spot noise and randomly adds the white spot noise to the clear neutron radiographic image; constructs a multiple distortion suppression model of the GAN neural network based on the coordinate attention mechanism; uses Huber loss to train the constructed GAN neural network based on the coordinate attention mechanism; inputs a real neutron radiographic image containing multiple distortions into the trained multiple distortion suppression model as input, predicts the original image of the real neutron radiographic image containing multiple distortions, and obtains a target result.

The present application belongs to the field of terahertz imaging and image processing, and particularly relates to a terahertz image data expansion method based on a space-frequency degradation model. The method mainly comprises the following steps: pre-screening and pre-processing a high-resolution image database to establish a high-resolution image library; simulating a terahertz Gaussian beam to perform fuzzy degradation processing on the high-resolution image using a Gaussian fuzzy kernel; simulating terahertz wave source fluctuation and sensor noise to add Gaussian noise; simulating low resolution of terahertz imaging, and reducing the image resolution using a bicubic downsampling method; simulating mutual interference of multiple terahertz waves during imaging, and using a scheme of adding a frequency domain mask to change the frequency domain features; simulating the case that low image contrast is caused by low terahertz wave source power, and using a gray scale compression method to compress the gray scale histogram of the image to obtain a final low-resolution image data set. The present application can effectively expand the super-resolution training set of terahertz images, thereby improving the training effect of the terahertz image super-resolution reconstruction task based on the deep learning method.

The invention belongs to the field of flame detection, particularly relates to a flame detection method and system based on flame dynamic characteristics and electronic equipment, and aims to solve the problems of unstable detection effect and false alarm in the existing method. The method comprises the following steps: inputting a flame picture into a flame detection model to obtain a flame detection frame area; in the current frame, a flame detection frame area is cut out to serve as a first picture; cutting a flame detection frame area detected by a previous frame or a next frame corresponding to the frame as a second picture; respectively carrying out Gaussian blur processing on the grey-scale map of the first picture and the grey-scale map of the second picture to obtain a first image and a second image; respectively carrying out binary image processing on the first image and the second image based on the gray value threshold K; and judging whether flames exist in the field or not based on the image contour area after binarization image processing. And the final detection effect, the detection efficiency and the accuracy of the detection result are improved.

This invention discloses an automatic computer protection system and method, relating to the field of computer file security protection. The technical solution includes: detecting user idle time; identifying sensitive tabs in the browser when the user is about to leave, and extracting the ORB feature point set of their content area as a content fingerprint; creating a Gaussian blur overlay window when the user has left, tracking the positional changes of the content area through feature point matching and affine transformation, and dynamically adjusting the position and size of the overlay window so that the overlay always follows the movement of the file content; and destroying the overlay window when the user returns, restoring the file to normal display. This invention, through content area tracking based on visual features and progressive blurring, solves the problem of sensitive file content being exposed and background tasks being interrupted when a user is briefly away, achieving fine-grained protection at the file and tab levels, and helping to improve the security of computer files in the office environment.

The invention provides a protection pressing plate state judgment method based on OpenCV, and the method comprises the steps: A, sequentially carrying out the gray conversion, Gaussian blur contour detection, Canny edge extraction, structural element denoising and contour refinement description of a protection pressing plate opening and closing state template image, and obtaining a template refinement contour; step B, executing the processing flow of the step A on the to-be-identified protection pressing plate image to obtain a to-be-identified refined contour; and step C, calculating the 1-7-order Hu invariant moment similarity of the refined contour of the template and the refined contour to be identified, and judging the opening or closing state of the protection pressing plate based on the 1-7-order Hu invariant moment similarity and a preset threshold value. According to the invention, the problems of low efficiency and insufficient accuracy during substation secondary screen cabinet inspection are solved.

This application belongs to the field of image processing technology and discloses a method and apparatus for generating hairstyle replacement videos based on a diffusion model. The method includes: acquiring a hairstyle image dataset and training a diffusion model to obtain a hairstyle generation model; acquiring a user-inputted image of a person and a target hairstyle prompt; using skeletal keypoint detection technology to identify the person image and obtain skeletal pose data; using an image segmentation algorithm to divide the hair region in the person image to obtain a hair mask; performing smoothing and feathering processing on the hair mask using Gaussian blur and feathering algorithms; inputting the hair mask and the target hairstyle prompt into the hairstyle generation model, and using the skeletal pose data as constraints for the hairstyle generation model to obtain a hairstyle replacement image; and generating a hairstyle replacement video based on the person image and the hairstyle replacement image. This application can improve the realism and naturalness of hairstyle replacement images and videos.

The invention discloses an image deblurring method and device based on blind deconvolution, and relates to the technical field of engine turbine blade strain measurement image deblurring. In order to solve the defect that the dual requirements for image quality and processing efficiency in engine blade strain detection are difficult to meet in the prior art, the technical scheme provided by the invention comprises the following steps: acquiring images of a blade in a static state and at different rotating speeds, taking a static image as a reference, and taking a dynamic image as a to-be-processed image; initializing a blind deconvolution parameter based on a Gaussian blur kernel; performing blind deconvolution iterative optimization under the constraint of a still image, and jointly estimating a clear image and a blurred kernel to obtain a preliminary deblurred image; de-noising is carried out; and adaptively selecting a gamma value according to the histogram mean value difference between the de-noised image and the still image, and executing gamma correction to enhance the contrast and details. According to the method, a deblurring-denoising-enhancing link is formed, the image definition can be effectively recovered under the condition of high-speed rotation, and the strain detection precision and stability are remarkably improved.

This invention relates to a privacy protection method for baby care cameras based on instance segmentation detection, comprising: Step 1, region detection, using an instance segmentation detection algorithm to identify and locate the adult region, the baby region, and the background region; Step 2, background region processing, the identified background region is first reduced in size, then Gaussian blurred, and then enlarged back to its original size to obtain a Gaussian blurred background region, which is then filled back into the original background region. The privacy protection method for baby care cameras based on instance segmentation detection obtained by this invention has the following advantages: the algorithm segments the camera image into adult, baby, and background regions, providing a foundation for subsequent processing; reducing the background region saves computational power required for subsequent processing, improving processing efficiency; and by combining modern image processing technology and artificial intelligence algorithms, it achieves efficient and aesthetically pleasing privacy protection measures.

The application provides a building edge automatic fusion transition method and device based on depth detection, which comprises the following steps: dividing a virtual scene into a plurality of sub-regions and generating a mask image sequence; performing Gaussian blur processing according to the mask image sequence to determine a fusion transition region; when rendering a terrain material, determining a corresponding target mask image according to a position index of a current pixel point in the mask image sequence; and performing automatic fusion transition on a building edge according to a final color value obtained by mixing calculation of the fusion transition region and the target mask image, so that the mask image can be automatically generated, the building edge and the terrain can be mixed in material, a natural edge transition effect can be obtained, the sense of reality is enhanced, and the atlas sampling overhead is low and the calculation amount is small.

This invention discloses a lidar depth completion method based on hierarchical minimum surface reconstruction, relating to the fields of computer vision and image processing technology. The method includes: performing depth value inversion and morphological dilation on a sparse depth map to obtain a scene dilated depth map and a full-resolution effective mask; then downsampling to obtain a coarse-scale scene dilated map and a coarse-scale mask; using an iterative convolution kernel of the Laplacian operator to obtain a coarse-scale filled map; upsampling the coarse-scale filled map to obtain an upsampled depth map, and then fusing it with the scene dilated depth map to obtain a fused depth map; locating all remaining holes in the fused depth map using connected component analysis, and calculating the mean depth of the effective pixels within the annular pixel band around each remaining hole; using this mean depth to fill the corresponding remaining holes to obtain a hole-free depth map; and after global Gaussian blurring, performing a depth value inversion operation to obtain a scene dense depth map, thereby achieving high-precision 3D reconstruction.

The invention provides a vegetation detection method and system, computer equipment and a storage medium, and belongs to the field of remote sensing data processing, and the method comprises the steps: collecting vegetation UAV image data in a sample region according to five different flight heights, and collecting vegetation UAV image data in a research region according to a single flight height; performing position and quantity labeling of vegetation targets on the vegetation UAV image data collected in the sample area, and performing Gaussian blur simulation and area resampling processing on the labeled image data in sequence to obtain multiple groups of simulation image data consistent with the image resolution characteristics of the research area; the target detection model is trained through the multiple sets of simulation image data; inputting the collected vegetation UAV image data of the real single flight height of the research area into the trained five sets of target detection models, and outputting respective vegetation detection frame coordinates and confidence degree results; and fusing each vegetation detection frame and the confidence result, and outputting the number of vegetation plants corresponding to the single flight height image of the research area. According to the method, cooperation and unification of large-range coverage and high-precision identification are realized.

The invention belongs to the technical field of wine identification, and discloses a wine batch identification method and system in a complex environment, and the method comprises the steps: carrying out the training of a low-illumination image enhancement model based on a data enhancement modulation method; obtaining a to-be-recognized image, and obtaining a low-light enhanced image by using the low-light image enhancement model; training the image definition enhancement model according to down-sampling and Gaussian blur algorithms; based on the low-light enhanced image, using an image definition enhancement model to obtain a definition enhanced image; training the wine identification model by using a two-stage identification framework and combining an additive angle interval loss function to respectively obtain a wine bottle target detection model and a single-bottle wine image identification model; and identifying the definition enhanced image to obtain wine category information. The method has excellent performance in the aspects of multi-wine batch identification, wine identification under complex illumination conditions, low-definition wine identification, new category expansion and open set identification.

The application provides a laser projector production line light adjustment and focusing detection method and system, which comprises the following steps: S1, image preprocessing is performed by using Gaussian blur filtering, edge detection, contour discovery and ellipse fitting operation; S2, spot identification is performed by using edge protection filtering, Gaussian curve fitting, binarization processing and ellipse fitting operation; S3, the size and gray scale of the spot are calculated to complete spot checking; S1', light adjustment grid image preprocessing is performed by using Otsu binarization, contour discovery, minimum circumscribed rectangle and image interception operation; and S2', grid checking is completed by image interception, grid gray scale calculation and checking. The application solves the technical problems of difficult accurate identification of spot boundary and difficult quantification of grid line width, improves the efficiency and yield of the light adjustment and focusing process, saves the labor cost, and brings economic benefits to customers.

This invention relates to the field of image processing technology, specifically disclosing an image privacy processing method based on region segmentation. The method includes: segmenting the original image to generate a region mask, marking visible and invisible regions with the region mask; using the image corresponding to the visible region as input to a keypoint detection model, outputting the keypoint coordinates of the target object; determining the target object region and non-target object regions based on the keypoint coordinates, further dividing the target object region into a key feature region and other regions; applying pixelation processing to the key feature region and Gaussian blur processing to the other regions to generate a privacy processing region; and a processing unit fusing the images of the non-target object region, the privacy processing region, and the invisible region to obtain a fused image. This region-segmentation-based image privacy processing method solves the problem of effectively distinguishing between visible and invisible regions, protecting the privacy of the target object, and retaining necessary information during image acquisition and processing.

This invention provides a geometric correction method for high-orbit staring SAR based on multi-angle weighted image simulation in mountainous areas. The method involves selecting a DEM image D for the corresponding region; determining the coordinates of the four corner points of image D; calculating the corresponding simulated image value for each pixel in image D; performing multi-scale decomposition on each SAR multi-angle image using a non-downsampling pyramid to obtain high-frequency and low-frequency sub-bands for each image; fusing the high-frequency and low-frequency sub-bands using Gaussian blur; calculating the fusion coefficients of the high-frequency and low-frequency sub-bands; and then determining the grayscale value of the fused image; finally, matching the simulated image with the SAR image using the SIFT algorithm to obtain the geometrically corrected image. This invention effectively alleviates the problem of the significant impact of elevation on the geometric correction accuracy of images during high-orbit SAR imaging in mountainous areas.

The invention relates to a multi-modal perception and excitation micro-payment high-precision map system, which comprises a vehicle-mounted data acquisition and lightweight perception module which is deployed in a vehicle TBOX or a user mobile terminal and is used for acquiring multi-source sensing data of a road environment in real time; the input end of the privacy protection module is in communication connection with the output end of the vehicle-mounted data acquisition and lightweight sensing module, and the privacy protection module is used for intercepting a multi-frame image original data packet 5-15 seconds before and after an event occurs when the event is detected, performing Gaussian blur or black block shielding processing on a face and license plate area in the image, and transmitting the face and license plate area to the vehicle-mounted data acquisition and lightweight sensing module; and meanwhile, generating an encrypted position hash based on the desensitized geographic coordinates. According to the method, a lightweight multi-mode sensing model deployed in a vehicle-mounted TBOX, a smart phone APP or automobile data recorder firmware is used as a distributed sensing node, and local fusion of multi-source heterogeneous data is utilized, so that the system can identify a long-tail event on a resource-limited edge device in real time, and the sensing boundary of a map for instantaneous change of the real world is expanded.

The invention discloses an anti-resonance optical fiber identification and high-precision axis alignment method based on end face characteristics, and relates to the technical field of optical fiber fusion splicers, and the method comprises the following steps: S1, collecting end face images of multiple types of anti-resonance optical fibers, the multiple types of anti-resonance optical fibers including a basic type and a complex nested type; preprocessing the acquired image, wherein the preprocessing comprises graying, Gaussian blur denoising, Sobel operator gradient calculation, binaryzation and morphological operation; s2, preprocessing the acquired image, including graying, Gaussian blur denoising, Sobel operator gradient calculation, binaryzation and morphological operation; according to the anti-resonance optical fiber identification and high-precision axis alignment method based on the end face features, the end face feature differences of anti-resonance optical fibers of different structures are deeply excavated, a two-stage classification framework of large class distinguishing-subdivision identification is innovatively constructed, and the adaptability and accuracy of classification identification of the multi-structure anti-resonance optical fibers are remarkably improved.

The application discloses a pedestrian re-identification method and system based on Fourier frequency domain disturbance data generation. The method realizes data generation by modifying the color and texture of pedestrian clothes, and specifically comprises the following steps: an adaptive intensity adjustment module is used to dynamically calculate disturbance weights according to the brightness, contrast and texture richness of the input image to ensure the visual authenticity of the generated samples; a frequency domain amplitude disturbance module is used to decouple the phase and amplitude components by using Fourier transform, and high and low frequency division is performed on the amplitude spectrum to respectively realize controlled disturbance of color and texture; and a clothing mask matching module is used to generate a soft edge mask in combination with human body analysis and Gaussian blur, so that the modification range is accurately limited in the clothing area to avoid the risk of identity information loss. The above scheme can effectively expand the cross-clothing training data, and significantly improve the generalization performance and retrieval accuracy of the pedestrian re-identification model in complex actual environments.

The invention relates to a training data set acquisition method and a microscopic particle analysis method, and the method comprises the steps: extracting a plurality of particles from an original training data set with labels as seeds, removing a background through an image segmentation method, and storing the background as a transparent channel image; gaussian blur processing is carried out on an original image, and a diversified background image pool is constructed; and placing a plurality of extracted particles in various backgrounds through a plurality of rules to obtain a plurality of combinations, and generating new image data with labels. And mixing the generated new image data with the original training data set in proportion to form a mixed data set. According to the method, a large amount of labeled data can be obtained through expansion of a small amount of original labeled data, and the generated new image data is formed by crossed combination of particles and backgrounds, so that the sample quality is high, the problem of high cost of deep learning labeling in the field is solved, and the industrial application cost is greatly reduced.

The invention discloses a method and device for generating a high dynamic range image based on FPGA double-exposure multi-scale fusion, and belongs to the technical field of image processing.The method comprises the steps that a double-exposure LDRI is obtained, and then an input image data stream is obtained; the method comprises the following steps: calculating the contrast, saturation and exposure satisfaction of an image based on an input image data stream, and carrying out normalization processing on a weight value formed by the contrast, the saturation and the exposure satisfaction to obtain a normalized weight map; completing the construction of a Gaussian pyramid of the normalized weight map; completing the construction of a Laplacian pyramid for the obtained input image data stream; and performing same-scale fusion on each layer of the output image subjected to Gaussian pyramid construction and each layer of the output image subjected to Laplacian pyramid construction with the same resolution, performing Gaussian blur and up-sampling reconstruction on the fused pyramid, and restoring the image to the initial resolution. According to the invention, the image fusion efficiency and the operation speed of the whole system are improved.

This application proposes a video stream cloud-edge collaborative analysis method and device. The method includes: an FPGA edge computing device acquiring a video stream captured by a camera device; the FPGA edge computing device performing real-time key region detection on the video stream using an internally deployed lightweight neural network model to obtain key region detection results and extracting low-resolution feature vectors of the key regions; the FPGA edge computing device, through its built-in video encoding and decoding unit, performing lossless encoding and high bit rate compression on the key region video stream based on the key region detection results, and performing Gaussian blur preprocessing and high quantization parameter compression on the non-key region video stream to obtain compressed video data; and uploading the compressed video data, key region location metadata, and low-resolution feature vectors to a cloud GPU server. This can improve real-time performance and network efficiency.

The present application provides a kind of ghost imaging monitoring method and system based on environmental coupling simulation and image enhancement, the method comprises: setting initial light source, the initial light source is handled with splitter matrix model to obtain probe light path light source and reference light path light source, based on the probe light path light source and reference light path light source, respectively construct probe light path and reference light path;In probe light path, lens, object to be measured and bucket detector are set in sequence, and surface detector is set in reference light path, high gauss blur is introduced in probe light path to simulate heavy fog environment, and random phase screen based on Perlin noise is introduced to simulate turbulent environment.The present application builds ghost imaging system through Python, uses Gaussian simulation to simulate heavy fog environment and turbulent environment, and combines the concept of histogram equalization in digital image processing, the imaging quality of ghost imaging is optimized by non-linear mapping gray scale chart, and the imaging quality of ghost imaging system under long-distance condition can be improved.

The application provides a method, device, equipment and medium for layering and identifying commodity surface packaging text, which comprises the following steps: performing gray scale conversion and Gaussian blur noise reduction on a commodity packaging image; calculating pixel depth statistical features and gray gradient features of the gray scale image; setting a depth threshold value set based on equal interval quantile of the gray value, setting a basic gradient threshold value based on the mean and standard deviation of the gradient amplitude, and fine-tuning the basic gradient threshold value according to the hierarchical index to obtain gradient threshold values of each level; obtaining text masks of each level according to joint screening of the depth threshold value and the gradient threshold value; and performing morphological optimization and connected domain analysis on the text masks to mark text regions of different levels on the original image. The application can adaptively identify commodity packaging text of different light and dark levels and has high accuracy.

This invention provides a method for LiDAR relocalization in underground space based on Gaussian fuzzy maps, belonging to the field of LiDAR relocalization technology. The method includes: S101, acquiring underground space point cloud data through a robot and denoising the data; S102, processing the denoised point cloud data using a Gaussian kernel function to generate a Gaussian fuzzy NDT map; S103, quantifying the impact of point cloud distribution on LiDAR positioning reliability using density ratio exponentiation; and S104, predicting and correcting robot pose using an unscented Kalman filter algorithm and a normal distribution transformation algorithm. This invention achieves centimeter-level positioning accuracy (RMSE 0.077m) and real-time stable operation in underground space, significantly improving registration success rate, error tolerance, and computational efficiency compared to existing methods.

The application discloses a kind of self-adapting video sharpening method, device, equipment and storage medium based on look-up table method, method includes: extracting Y component in the image of video frame to be sharpened;Gradient and gradient intermediate value of image are calculated;Based on look-up table method initialization, corresponding sharpening weight is obtained by searching look-up table method according to the pixel value of gradient image and gradient intermediate value;Smoothed image is obtained using Gaussian blur;High-frequency information image is calculated according to Y component image and smoothed image;Sharpening mask is calculated according to high-frequency information image and sharpening weight, and the image after outputting sharpening mask;Sharpening mask image and smoothed image are merged, and the output Y component image after sharpening is merged with UV component.Utilize look-up table method to solve the problem of large amount of adaptive sharpening calculation, realize the adaptive sharpening adjustment of image high-frequency information, reduce the generation of noise in the case of clear picture, and can be applied to real-time video processing.

This invention relates to the field of UAV pose map optimization technology, and particularly to a method for optimizing the pose map of a UAV monocular camera. The method includes: blurring different image frame data, constructing a Gaussian pyramid, and downsampling; determining candidate feature points by subtracting Gaussian blurred images of adjacent scales; accurately determining the positions of candidate feature points on the image frame; removing low-contrast candidate feature points and edge response feature points to obtain the final feature points; and using the LM algorithm to adjust the UAV attitude parameters and optimize the pose map for discontinuous image frame data. The process of Gaussian difference point interest detection on continuously acquired image frame data suppresses noise in the image frames, effectively smoothing noise in the image. The resulting pose map is smoother with smaller gradient changes, improving the robustness of attitude estimation. The pose map optimization process expands the scope of pose map optimization, allowing optimization even for discontinuous images.

The invention discloses a transferable face adversarial attack method based on input transformation. The method comprises the following steps: training an adversarial transformation network based on a convolutional neural network (CNN) by using a preprocessed data set; inputting an original image into the adversarial transformation network, simulating an input transformation operation through the adversarial transformation network, and outputting a transformed image; in each iteration, n points are randomly selected from the face feature points, a shielding area is generated, and Gaussian blur is applied to the shielding area; and iteratively generating an adversarial sample in combination with a fast gradient descent method, and attacking the target model by using the adversarial sample. The adversarial transformation network is used for simulating an input enhancement strategy, and the problem that the generalization of unknown changes is poor due to the fact that most input transformation methods are single and fixed transformation modes is solved; the excessive sensitivity of the model to some specific discrimination areas is reduced by shielding the key points, and the adversarial sample mobility can be effectively improved by combining an adversarial transformation network.