367 results about "Superpixel segmentation" patented technology

The present invention pertains to the technical field of digital image processing. Disclosed are an unsupervised super-pixel segmentation method and system assisted by collaboration between an atrous pyramid and an attention mechanism. The method comprises: combining image RGB channel information with position information of a pixel point; constructing a channel attention module by using an attention mechanism; processing the result of the attention mechanism by using atrous spatial pyramid pooling; constructing a loss function, constructing a clustering loss term, and, by using a spatial smoothing loss term, constructing a reconstruction loss term; updating model parameters by using an Adam optimizer, and using for super-pixel generation an effective depth feature obtained from the last separation; and obtaining the maximum value of a channel dimension by using an argmax function, converting a processing result of the argmax function into a two-dimensional array, and, on the basis of a defining condition, completing adaptive super-pixel segmentation in a CPU. According to the present invention, the complexity is low, the adaptive and generalization capabilities are high, and effective support is provided for improving image processing efficiency and accuracy.

The invention provides a small target identification method and system based on YOLOv5, and the method comprises the steps: collecting a multi-scale image in a target range, and forming an original data set; performing image enhancement on the original data set through superpixel segmentation and adversarial enhancement operation to obtain an enhanced data set; performing multi-scale frequency domain aliasing enhancement on the enhanced data set through frequency domain decomposition and frequency band exchange operation to obtain a to-be-detected data set; improving the YOLOv5 model to obtain a small target recognition model; and performing small target identification on the to-be-detected data set through the small target identification model to obtain an identification result. According to the method, small target recognition is realized through the dynamic feature pyramid and the double-path detection head in combination with cross-level kernel sharing and a space-frequency double-domain attention mechanism, the problem of detail loss caused by a traditional static feature pyramid is solved, the recognition precision is improved, and the false detection rate is reduced.

The invention discloses an optical performance detection system and method for a camouflage material, and relates to the technical field of camouflage material detection, and the system comprises a structure identification and illumination control module which is used for collecting a surface image of a camouflage material sample, identifying the surface structure of the sample, starting a full-spectrum light source, and simulating step-by-step illumination from weak light to strong light; carrying out RGB imaging based on the standard illumination; and the region division module is used for applying a superpixel segmentation algorithm to the image, automatically dividing sub-regions, performing hyperspectral reflectivity scanning on each region, synchronously acquiring a light intensity image in a polarization direction, calculating a Stokes vector and constructing a polarization enhanced image. According to the method, frequency domain quantitative analysis of microcosmic interference fringes and accurate mask extraction of polarization abnormal areas can be realized, optical exposure risks of camouflage materials under complex illumination and visual angle conditions are effectively quantified through multi-angle reflection sampling and scoring function modeling, and the detection precision, the area pertinence and the automation level are remarkably improved.

The invention belongs to the technical field of computer vision, and relates to a lightweight image feature point detection and matching method and related equipment thereof. The method comprises the following steps: performing feature extraction on a target image through a lightweight SuperPoint neural network model to generate feature information; performing super-pixel segmentation processing on the image to generate a super-pixel segmentation result; performing initial matching on the feature information through a lightweight SuperGlue neural network model to generate an initial matching pair; performing outlier filtering on the initial matching pair based on a superpixel segmentation result, and determining an optimized target matching pair; performing geometric consistency optimization on the optimized target matching pair by adopting an adaptive exponential moving average algorithm to generate a final matching result; the lightweight SuperPoint neural network model is generated by carrying out hierarchical progressive pruning processing on an original SuperPoint neural network model; the lightweight SuperGlue neural network model is generated by carrying out attention head pruning processing and graph network layer pruning processing on an original SuperGlue neural network model.

The invention relates to the technical field of metal material detection, in particular to a titanium alloy bar machining quality detection method and system, and the method specifically comprises the steps: carrying out the superpixel segmentation of the section of a titanium alloy bar, constructing a mirror image edge texture saliency value of a superpixel block based on the proportion of edge points in the superpixel block and the edge gradient, and carrying out the detection of the machining quality of the titanium alloy bar. Constructing the mirror image texture intensity of the titanium alloy bar by combining the gray level distribution concentration characteristics of the titanium alloy bar, and performing image segmentation on the section of the titanium alloy bar; and in each segmented region, based on the edge line fuzzy degree difference and the pixel gray level difference of each super-pixel block and the rest of super-pixel blocks, combining a mirror image edge texture saliency value to construct a segregation evaluation value of each super-pixel block, and based on the segregation evaluation value, carrying out titanium alloy bar segregation quality defect detection. The titanium alloy bar segregation defect identification capability is enhanced, the titanium alloy bar segregation defect detection accuracy under the titanium alloy bar mirror image scene is improved, potential problems can be found earlier, missing detection is avoided, and the defective rate is reduced.

The invention discloses a hyperspectral image segmentation method based on semi-supervision and super-pixel progressive growth. The method comprises the following steps: carrying out normalization processing on a hyperspectral image to be segmented; taking the normalized hyperspectral image as input, constructing a full convolutional network comprising an input layer, a convolutional layer, a batch normalization layer and an output layer, and taking the network output as a semantic segmentation result corresponding to the hyperspectral image; performing superpixel segmentation on the normalized hyperspectral image, and endowing each initial superpixel area with an initial pseudo label; combining the similar initial superpixel areas, and updating the pseudo labels corresponding to the combined superpixel areas; the method comprises the following steps: minimizing a loss function, reversely optimizing feature extraction of a feature extractor, performing fine adjustment on a full convolutional network by using a small number of labeled samples, and continuously optimizing a superpixel region to converge a model to obtain a final segmentation result.

The invention discloses an unmanned aerial vehicle power distribution network equipment identification method based on mutual neighbor density peak value clustering, and relates to the technical field of power distribution network equipment inspection, and the method comprises the steps: obtaining an original RGB image of a power distribution line; outputting the contrast characteristic coefficient of each channel; constructing a saturation retention item; calculating a contrast feature retention item, constructing a total energy function, solving an optimal channel weight by adopting a discrete search strategy, and outputting an initial grayscale image; sequentially carrying out weighted guide filtering, morphological reconstruction and super-pixel segmentation operation; performing two-dimensional wavelet decomposition on the super-pixel segmented image, and extracting a feature vector; and calculating the local density and the relative distance, selecting a clustering center, and completing sample distribution based on the shared mutual neighbor similarity to realize a power distribution network equipment identification effect. According to the method, the problems of detail loss, noise interference, edge breakage, disordered classification of multi-scale equipment and the like under complex illumination are effectively solved, and the identification precision and efficiency of unmanned aerial vehicle inspection are effectively improved.

The invention provides an annotation method and system based on forest land remote sensing data. The method comprises the following steps: acquiring multi-source remote sensing data of a forest land; processing the multi-source remote sensing data; an improved Mask R-CNN network is adopted to extract optical image features, the weight of a near-infrared band is enhanced through a compression and excitation SE module, and feature fusion is carried out to obtain a multi-modal enhanced feature map; carrying out feature extraction by adopting a 3D convolutional network to obtain a multi-level three-dimensional space feature map; inputting the processed data into a Transform model to obtain a forest growth curve graph; performing initial labeling based on the feature map and the curve map to obtain an initial labeling result; performing morphological filtering on the initial labeling result to remove small-area noise, and optimizing a labeling boundary in combination with a superpixel segmentation algorithm to obtain an optimized labeling result; and generating a vector annotation file according to the optimized annotation result. Multi-dimensional information complementation of spectral features, three-dimensional structure features and time sequence dynamic features is achieved, and data availability and reliability are remarkably improved.

The invention relates to a multi-modal remote sensing image change detection method and system and a readable storage medium, and the method comprises the steps: carrying out the normalization of a multi-modal remote sensing image, carrying out the superpixel segmentation through employing simple linear iteration clustering, and employing the pixels of superpixels as nodes to construct an original image structure; inputting the original graph structure into a graph attention network to extract graph structure features of superpixels; performing style regularization on the graph structure features, and performing style randomization on the graph structure features and adjacent superpixels to obtain new style features of the superpixels; inputting the new style features into a graph structure decoder to obtain a super-pixel reconstructed graph structure; training a graph attention network by using the content consistency of the reconstructed graph structure and the original graph structure to obtain a target graph attention network; and inputting an original image structure obtained by processing a to-be-detected multi-modal remote sensing image before and after change into the target image attention network to obtain image structure features before and after change, and obtaining a change binary image by using an Otsu method. The change detection precision is effectively improved.

The invention relates to the field of image processing, in particular to an oilstone static pressure forming quality detection method based on image processing, and the method comprises the steps: obtaining an infrared image set of an oilstone sintering and cooling stage, carrying out the superpixel segmentation to obtain a superpixel region of each frame of image, matching the superpixel regions of adjacent frames, and building a corresponding relation, and calculating a temperature anomaly characteristic value of each region. And analyzing the temperature change trend of each region, calculating a correlation coefficient, and introducing distance and temperature anomaly weight to carry out weighted fusion to obtain a comprehensive correlation index. And in combination with the temperature anomaly characteristic value and the comprehensive correlation index, evaluating the possibility value of occurrence of pore anomaly in each superpixel region, and judging whether pore distribution anomaly exists in the oilstone or not, thereby completing quality detection. According to the method, through superpixel segmentation and matching, in combination with temperature anomaly characteristics and comprehensive correlation analysis, precise detection of oilstone internal pore anomaly is realized, and the reliability of static pressure forming quality detection is improved.

The invention discloses a multisource remote sensing image surface water body extraction method based on a graph neural network. The SAR image is used as a main part, and the optical image, the digital elevation model and the local incident angle information are fused. Firstly, radiation correction, geometric correction and noise suppression are carried out on SAR data, and multi-source data spatial resolution matching is carried out. And then converting the image into a graph structure by adopting a self-adaptive Bayesian superpixel segmentation method. Then, geometric, texture and physical features are constructed for the superpixel nodes, and feature dimensions are optimized through a feature selection algorithm; and finally, constructing a classification aggregation graph convolutional neural network, wherein the core lies in a self-adaptive microaggregation mechanism of the classification aggregation graph convolutional neural network. According to the mechanism, neighborhood superpixels can be dynamically divided according to categories and feature fusion is carried out, and the recognition capability of the model on water body boundaries and regional heterogeneity is remarkably improved. According to the method, multi-source remote sensing data, superpixel graph structure conversion and an optimized graph convolutional network are comprehensively utilized, and the precision and robustness of water body extraction in the complex surface environment are effectively improved.

The invention discloses a semiconductor photoelectric chip anomaly detection method and device and computing equipment, and the method comprises the steps: loading a first target chip image, based on a first color model, of a target semiconductor photoelectric chip, and converting the first target chip image into a second target chip image based on a second color model, the second color model comprises a red channel, a saturation channel and a brightness channel; performing superpixel segmentation on the second target chip image to obtain a target chip superpixel image, and collecting a feature vector of each target superpixel in the target chip superpixel image; predicting a plurality of target key areas in the target chip superpixel image by using the target visual large model; based on the feature vector of each target superpixel, generating a corresponding OCSVM model for each target key area; and performing anomaly detection on any target key area based on the OCSVM model corresponding to any target key area. According to the invention, accurate anomaly detection can be carried out on the single key area of the semiconductor photoelectric chip.

The invention belongs to the technical field of target detection, and discloses a passive millimeter wave target detection method and system based on regional clustering, and the method comprises the steps: dividing a brightness temperature image into a plurality of superpixel regions through a superpixel segmentation algorithm according to the similarity between target regions and the similarity between background regions, calculating a saliency image based on the regional statistical characteristics; according to the spatial distribution difference of boundary superpixels and internal superpixels, regional local direction centrality measurement is introduced to carry out clustering analysis, so that a candidate target region is highlighted; edge compensation is carried out on the preliminarily screened candidate target area; building a total variation optimization model by using regularization construction constraints; the local features of the compensated candidate target region are enhanced by solving the total variation optimization model; and finally, realizing accurate target detection and extraction through threshold segmentation operation. According to the invention, the accuracy of target detection and the integrity of target contour extraction can be improved.

The invention relates to the technical field of image recognition, and discloses a crop lodging image data detection method and system, and the method comprises the steps: carrying out the self-adaptive illumination correction processing of the original image data of crop lodging, and obtaining an illumination compensation image of the original image data; performing multi-scale super-pixel segmentation on the illumination compensation image to obtain a crop area image of the illumination compensation image; extracting texture features and shape features in the crop region image, and performing feature fusion on the texture features and the shape features to obtain a multi-modal fusion feature set of the crop region image; performing multi-dimensional feature collaborative analysis on the multi-modal fusion feature set to obtain an initial discrimination result of the multi-modal fusion feature set; performing confidence coefficient optimization on the initial judgment result in combination with a spatial context relationship to obtain a target lodging detection result of the crop region image; according to the invention, the efficiency of crop lodging image data detection can be improved.

The invention discloses an urban landscaping tree species classification method based on a Pleiades Neo ultrahigh resolution multispectral satellite image. According to the invention, based on the Pleiades Neo image, an NDVI threshold method is adopted to rapidly identify the urban vegetation area; using a Unet < 3 + > model to separate vegetation in the shadow area and the illumination area; a shadow region tree species classification integration model is constructed by combining a superpixel segmentation algorithm and introducing a weighted dictionary and an attention mechanism, so that the classification precision is effectively improved; and carrying out tree species classification on the illumination area by using a deep learning algorithm and making an overall tree species distribution diagram. Aiming at the shadow interference problem in a high-resolution image, a set of tree species classification scheme comprehensively considering vegetation spectrum characteristics of an illumination region and a shadow region is developed by combining a deep learning algorithm to deeply mine potential information, a more accurate and feasible tree species distribution acquisition approach is provided for fine management of urban landscaping, and the urban landscaping quality is improved. And important data support is provided for health assessment and planning of the urban ecosystem.

The invention relates to the technical field of geological survey, in particular to an unmanned aerial vehicle multispectral geological survey method and system, and the method comprises the steps: fusing a multispectral image, a digital elevation model, geophysics and historical geological data, systematically constructing a geological feature priori knowledge model, including lithology, construction and alteration feature libraries, and mapping with multispectral data; multi-scale geologic features are extracted through adaptive wavelet transform and morphological analysis, and feature weight adaptive adjustment is achieved; geological units are accurately divided by adopting geological scene perception superpixel segmentation and combining geological boundary constraint and similarity recursion combination; identifying an interference mode, generating an adaptive filtering matrix, and enhancing image quality; cooperatively interpreting multi-source information by using a deep auto-encoder network to generate a high-precision geological interpretation map and a confidence map; geological professional knowledge is introduced, so that the geologic body recognition accuracy is remarkably improved; the adaptive flight control strategy ensures the consistency of complex terrain data, and improves the precision and efficiency of geological survey.

The invention discloses a food waste detection method and system based on image processing, belongs to the field of image recognition, and aims to realize efficient and automatic recognition and quantification of kitchen waste. According to the method, residual food images are collected at multiple periods and multiple angles in a kitchen garbage can or a dinner plate recovery area through high-resolution and multi-spectral imaging equipment, preprocessing is carried out in combination with an improved Retinex algorithm and a space self-adaptive denoising technology, and the image quality is improved. Afterwards, fine segmentation of a food area is achieved through a multi-scale super-pixel segmentation and graph segmentation algorithm, and multi-category intelligent recognition is conducted on remaining food through a recognition network fused with multi-modal features. The system further combines stereoscopic vision and Monte Carlo sampling to dynamically and accurately count the volume or weight of various residual foods. The method has the advantages of high adaptability and accurate statistical result, and can provide data support for catering management, resource recovery, nutrition evaluation and the like.

The invention relates to the technical field of remote sensing image processing, in particular to a land consolidation boundary line division method and system based on machine vision. The method comprises the following steps: obtaining a plurality of superpixel blocks of a satellite remote sensing image; according to the elevation information distribution of different pixel points in each super-pixel block in the digital elevation model, obtaining the terrain height change degree of each super-pixel block; according to the elevation information difference between each pixel point in each super-pixel block and other pixel points in the corresponding neighborhood range in the digital elevation model and the gradient distribution of the corresponding pixel points in different directions in the digital elevation model, obtaining the overall water flow direction value of each super-pixel block; further obtaining the merging possibility between different super-pixel blocks; and obtaining a superpixel optimal block, and dividing the arrangement boundary line of the to-be-divided land region. According to the method, the superpixel segmentation process of the land region to be divided is accurately obtained, so that the accuracy of boundary line division arrangement is improved.

The invention discloses a photovoltaic fault positioning method and system based on an unmanned aerial vehicle, and particularly relates to the technical field of fault positioning. A thermal imaging camera and a visible light camera carried by an unmanned aerial vehicle collect a photovoltaic module image along a preset track, and a thermal radiation image and a visible light image are obtained; performing registration fusion on the thermal radiation image and the visible light image to generate multi-modal image data; performing super-pixel segmentation on the multi-modal image to form a plurality of analysis sub-regions; according to the temperature gradient distribution and color feature difference of the sub-regions, identifying temperature anomaly and color anomaly regions; analyzing a propagation path of the abnormal features through spatial neighborhood topology, and judging a diffusion range of the propagation path; according to the method, a temperature abnormal area and a color abnormal area are comprehensively scored in combination with the diffusion range of a propagation path, a fault area is marked, and spatial positioning information is output, so that high-precision positioning of a photovoltaic module fault is realized, and the automation and intelligence level of fault detection is improved.

The present invention discloses a polarimetric SAR change detection method, device, equipment and medium, which relate to the field of radar image processing technology. The method acquires a time-series PolSAR image, uses JBLD divergence to calculate the similarity measure of the multi-phase covariance matrix; calculates a time-series edge intensity map based on the similarity measure; uses the time-series edge intensity map to initialize the cluster center and introduces a dynamic edge constraint mechanism to suppress superpixels from crossing the image edge during the iteration process, and outputs the superpixel segmentation result; constructs an image topology representation that integrates time-series feature similarity, spatial adjacency and cross-phase cross-feature similarity; constructs an energy function containing node cost and edge cost, solves the energy minimization problem through quadratic pseudo-Boolean optimization, and obtains a change detection map. The present invention can avoid errors caused by regional discontinuity and boundary fuzziness in superpixel segmentation, and exhibits strong robustness in both natural objects and complex urban building scenes.

The invention discloses a polarimetric SAR change detection method, device and equipment and a medium, and relates to the technical field of radar image processing. The method comprises the following steps: acquiring a time sequence PolSAR image, and calculating similarity measurement of a multi-temporal covariance matrix by adopting JBLD divergence; calculating a time sequence edge intensity graph based on similarity measurement; initializing a clustering center by using a time sequence edge intensity graph, introducing a dynamic edge constraint mechanism in an iteration process to inhibit super-pixels from crossing image edges, and outputting a super-pixel segmentation result; constructing image topological representation fusing time sequence feature similarity, spatial adjacency and cross-time phase cross feature similarity; and constructing an energy function containing node cost and edge cost, and solving an energy minimization problem through quadratic pseudo Boolean optimization to obtain a change detection graph. According to the method, errors caused by discontinuous regions and fuzzy boundaries in super-pixel segmentation can be avoided, and high robustness is shown in natural ground features and complex urban building scenes.

The invention provides a hyperspectral image classification method fusing a hybrid multi-hop graph convolutional network, and belongs to the technical field of hyperspectral image processing, and the method comprises the steps: carrying out the dimension reduction processing of original HSI image data, and obtaining the HSI image data after the dimension reduction; segmenting the HSI image data after dimension reduction to obtain a superpixel segmentation result; inputting the original HSI image data into a convolutional feature pyramid network to obtain a first feature representation, and inputting the superpixel segmentation result into a hybrid multi-hop graph convolutional neural network to obtain a second feature representation; splicing the first feature representation and the second feature representation, and inputting the spliced feature representation into a mixed feature fusion module to obtain a fused feature; and inputting the features into a softmax classifier to obtain a final classification result. Local spatial-spectral features in the superpixel can be learned, long-distance associated information of the captured image can be captured, and rich multi-scale features and semantic information in the image can be effectively extracted; and higher classification precision is realized with less parameter quantity.

The invention relates to the technical field of remote sensing image processing, in particular to a joint sparse representation hyperspectral image classification method based on dual domain constraints, which comprises the following steps: preprocessing: carrying out spectral feature-based wave band grouping on hyperspectral image data, and then carrying out MNF data dimension reduction on the grouped hyperspectral image data; extracting a main component feature map by using morphology; performing superpixel segmentation on the hyperspectral image by using an improved watershed algorithm, and performing FCM clustering on the hyperspectral image; adaptive selection of the neighborhood is carried out through weight calculation under double constraints of the obtained superpixel neighborhood and the clustering field; multi-view angles of the superpixel field, the clustering field and the constraint field are used for joint sparse representation; and a majority voting method is adopted to integrate classification results, and the classification results are adjusted through a correction rule, so that the classification effect of the hyperspectral image is improved, and the classification precision of edge pixels is improved.

The invention relates to the technical field of three-dimensional map recognition, and discloses a semantic segmentation-based low-altitude three-dimensional map element autonomous recognition method and system. The method comprises the following steps: acquiring a low-altitude remote sensing image and preprocessing to generate a multi-channel image matrix containing geographic coordinates and spectral characteristics; extracting multi-scale features through a pyramid feature extraction network in combination with cavity convolution, and obtaining an adaptive weighted feature tensor through a cascade attention mechanism fusion channel and a spatial weight; adopting a bidirectional feature fusion strategy to generate fusion features, and outputting an initial category probability distribution diagram by a semantic segmentation header network; obtaining a refined mask through edge perception optimization and superpixel segmentation correction, and mapping the refined mask to a three-dimensional coordinate system to generate a vector layer with a semantic tag; a constraint rule is deduced through a topological relation inference engine, logic conflicts are eliminated through rule-driven post-processing, finally, a standardized three-dimensional map element database meeting the geographic information standard is generated, and efficient, accurate and autonomous recognition of low-altitude three-dimensional map elements is achieved.

The invention discloses a non-reference image quality evaluation method and device based on multi-perception feature fusion and dynamic enhancement, and the method comprises the steps: obtaining the quality information of a multi-domain distorted image based on superpixel segmentation and Gaussian kernel filtering texture generation; quality related feature extraction is performed on multi-domain information through a semantic perception module and a distortion perception module, bidirectional modulation is performed on global visual features of an original distorted image through a cross attention mechanism, and dynamic fusion of multiple perception features is realized; through a parallel feature enhancement unit formed by local adaptive filtering of a visual self-attention block and a dynamic residual block, dynamic allocation of perception modes to different content areas is realized; generating a weighted quality score consistent with human visual perception through a weighted dual-path regression device; and outputting a predicted score consistent with the human score from the distorted image through the three sub-networks. According to the method, the problems of insufficient adaptability to complex content of a distorted image and low local distortion sensitivity are effectively solved.

The invention relates to the technical field of image data processing, in particular to a cashmere product quality detection method. The method comprises the following steps: acquiring a first grayscale image of the surface of a cashmere product to be detected, and performing super-pixel segmentation on the first grayscale image to obtain a plurality of pixel blocks; determining shape similarity and gray feature similarity between the first pixel block and the second pixel block to determine a target parameter value between the first pixel block and the second pixel block; under the condition that the target parameter value is smaller than a preset parameter threshold value, combining the first pixel block and the second pixel block to obtain a second grayscale image; and inputting the second grayscale image into a pre-trained cashmere quality detection model to obtain a quality detection result of the second grayscale image. Through the technical scheme, a more accurate quality detection result of the cashmere product to be detected can be obtained.

The invention relates to an optical and SAR image registration method and system based on local distortion division. The method comprises the following steps: acquiring optical and SAR images, performing superpixel segmentation, calculating superpixel feature similarity, dividing a local distortion region and a general region, inputting the local distortion region and the general region into a twin multi-feature capsule network for feature extraction, and constructing to obtain feature descriptors; and obtaining a first registration result corresponding to the local distortion region and a second registration result corresponding to the general region through the transformation model, completing super-pixel replacement reconstruction, and obtaining a registration result. By defining superpixel features and calculating the similarity of the superpixel features, local distortion regions and general regions of optical and SAR images can be quickly divided; according to the method, feature extraction is performed in a twin multi-feature capsule network, a corresponding transformation model is constructed to obtain a corresponding registration result, and replacement reconstruction is performed, so that the defect that a single registration model is easily influenced by local distortion can be overcome, and the image registration precision is improved.

The invention discloses a semi-supervised medical image automatic segmentation method based on pseudo label optimization. The method comprises the following steps: step 1, preprocessing and dividing a data set; 2, constructing a semi-supervised training framework which comprises a pseudo label refining module, a triple loss module and a mutual correction framework; refining processing is carried out through an SLIC superpixel segmentation method and an information entropy voting mechanism; a triple loss function is controlled to be minimized; a mutual correction loss function is controlled to be minimized; step 3, constructing a supervision loss function based on the result of the supervision test; the total loss function convergence is controlled; and step 4, segmenting the verification set, and evaluating segmentation results based on two indexes of a Dice coefficient and an ASD average surface distance. According to the method, a semi-supervised training framework is constructed based on a pseudo-label refining module, a triple loss module and a mutual correction framework, the number of pseudo-labels is increased, the spatial consistency and accuracy of the pseudo-labels are improved, and the capture capability and segmentation precision of boundary information are optimized.

The invention provides a medical image cell segmentation method based on progressive pseudo-label optimization in order to solve the problem of learning deviation caused by unreliable pseudo-label learning and the limitation that a fixed-form pseudo-label cannot provide reliable cell morphological characteristics in an existing weak supervision method. The method comprises the following specific steps: 1) utilizing weak supervision point labeling information, generating two initial pseudo labels with complementarity through a clustering algorithm and a superpixel segmentation algorithm, and respectively guiding a training process of a double-branch network; 2) utilizing a dynamic threshold and watershed algorithm to realize growth and boundary division of a cell region in the pseudo tag, so that the pseudo tag is gradually close to the real form of a cell; 3) designing a bidirectional cross supervision mechanism, and realizing knowledge migration and collaborative optimization through high-confidence prediction results of the two branch networks; and 4) uncertainty estimation and a difficult sample attention loss function are designed, the feature learning ability of the network to difficult samples with low contrast, fuzzy boundary and the like is enhanced, and the precision of weak supervision cell segmentation is effectively improved.

The invention discloses an SAR (Synthetic Aperture Radar) image target detection method and device combining superpixel segmentation and fusion. The method comprises the following steps: inputting an SAR image; uniformly segmenting the SAR image into a preset number of initial super-pixel regions, and taking the central point of each initial super-pixel region as an initial clustering center; updating a clustering center based on joint probability distribution of local entropy and gray level distribution; according to a self-adaptive pixel similarity measurement criterion, the pixels are distributed to the nearest clustering center; according to the self-adaptive pixel similarity measurement criterion, the similarity of pixel points is evaluated in a manner of combining a gray distance and a spatial distance and introducing a dynamic weight factor; constructing a gray-level co-occurrence matrix in six directions, calculating texture similarity among superpixels, and fusing the superpixels of which the texture similarity values are smaller than a set threshold value; and performing multistage rule screening on the SAR image after superpixel fusion, distinguishing target and background superpixels, and outputting a target detection result. According to the invention, the precision of the SAR image target detection technology is improved, and the calculation complexity is reduced.