36results about How to "Improve segmentation results" patented technology

The invention provides a superpixel method for medical image segmentation. The method comprises the steps of: processing a medical image into a superpixel; for the medical images obtained after superpixel segmentation, using bilateral filtering to preserve the edge and filter the noise to reduce the error rate of the network model; configuring a network framework, and constructing a convolution network for the medical images obtained after superpixel segmentation by iterative training parameters. Based on the linear iterative clustering segmentation method, this method applies the thought of the U-Net network to the post-optimization of super-pixels, which makes up the defect of inaccurate segmentation of inner edge of super-pixel, increases the standard layer to improve the weight sensitivity of each network layer, improves the convergence performance of the network, and makes the segmentation result closer to the actual value. Because the anatomical structure and pathological tissueof medical images are very clear, the medical images segmented by SLIC algorithm can obtain more comprehensive super-pixel, and the edge accuracy of super-pixel can be further improved by convolutionnetwork.

The invention relates to the technical field of artificial intelligence, and discloses a three-dimensional point cloud semantic segmentation method, device and equipment and a medium, and the method comprises the steps: carrying out the point cloud division and quantitative discrimination of to-be-predicted three-dimensional point cloud data through employing a preset space cell, and obtaining target point cloud data; inputting the target point cloud data into a point cloud semantic category prediction model to perform semantic category probability prediction to obtain a point cloud semantic category probability prediction value of the target point cloud data, wherein the point cloud semantic category prediction model is a model obtained by training based on a Point SIFT neural network module and a Point Net + + neural network; and determining a target semantic category of each point in the target point cloud data according to the point cloud semantic category probability prediction value. According to the method, rapid and accurate logic division is carried out on the point cloud of the complex large-scale target object, the recognition precision of point cloud segmentation is improved, fine features of the complex target object can be well processed, and the accuracy of semantic category prediction is improved.

The invention provides a weak supervised image semantic segmentation method based on iterative mining common features of objects, belonging to the technical field of pattern recognition. In the training phase, the method acquires the training data set, constructs and trains the multi-label classification network, and acquires the initial seed area corresponding to each training image. Then, the superpixel region and the region label of each training image are obtained for training the region classification network, and the updated region label of the superpixel region is obtained for trainingthe semantic segmentation network. After iteration, when the performance of the semantic segmentation network converges, the trained semantic segmentation network is obtained. In the use stage, the color image is input into the trained semantic segmentation network, and the network outputs the semantic segmentation results of the image. The invention can realize reliable pixel-level semantic segmentation under the condition of only image class label, reduces the time and labor cost of data labeling, and has wide application prospect.

The invention belongs to the field of image processing, computer vision, deep learning and semantic image segmentation, and particularly discloses a method for segmenting high-resolution combined viewdermoscopy images on the basis of global cavity convolution. The method includes constructing high-resolution combined view feature extraction networks and semantic segmentation networks on the basisof cavity convolution; carrying out training by the aid of cross entropy and jaccard approximation coefficient combined loss functions; carrying out data enhancement and post-processing during prediction. The method has the advantages that sufficiently comprehensive context information can be sieved by combined view by the aid of the global cavity convolution, high-resolution images can be retained, accordingly, sufficiently detailed information can be captured, and the dermoscopy images can be accurately segmented.

The invention discloses a method for splitting images based on clustering of immunity sparse spectrums, mainly solving the problem that a spectrum clustering method has poor stability and high complexity. The method comprises the following steps: (1) extracting features of images to be split; (2) normalizing feature data to eliminate the magnitude effect among data; (3) carrying out real coding onthe normalized feature data; (4) randomly generating initial species groups of the coded data and calculating the affinity; (5) cloning based on the affinity size of an antibody; (6) carrying out Gaussian mutation on the cloned antibody species groups and selecting the antibody of the highest affinity as an input for the next round; (7) iterating the set maximum iterations to obtain a final selected sample subset; (8) carrying out greed spectrum dimensionality reduction on the selected sample subset, clustering the dimensionally reduced data and outputting the final image splitting result. Compared with the prior art, the method has the advantages of no need of priori knowledge, high accuracy and low calculation complexity. The method can be used for object detection and object identification.

The invention discloses a multi-objective organ-at-risk automatic segmentation method, device and system based on deep learning. The method comprises the steps of receiving an input image of a patient; carrying out format conversion on the input image of the patient, and converting the input image into JPEG format data; inputting the JPEG format data into an Overfeat positioning detection networktrained according to a physician manual segmentation result, and automatically selecting a region of interest containing multi-objective organ-at-risk; inputting the automatically selected region of interest into an FCN initialization segmentation network, and carrying out contour inference; carrying out the coordinate processing of the initial boundary contour obtained through the contour inference and the received artificial mark boundary, mapping the initial boundary contour and the received artificial mark boundary to an input image, extracting a DAISY feature, and obtaining a DAISY feature image; and inputting the DAISY feature image into a deep belief network trained according to a physician manual segmentation result to obtain an accurate segmentation boundary of organ-at-risk, namely a segmentation result.

The invention discloses an object segmentation method and an object segmentation system based on mixed marks. The method comprises the following steps of: receiving marking strokes of pixels in an image which comprises a target object, wherein the marking strokes comprise foreground strokes for the target object, background strokes for a background and undetermined strokes which are undetermined for a foreground or the background; on the basis of the foreground strokes and the background strokes, establishing a foreground model and a background model to classify the pixels in the image, which are not marked by the undetermined strokes, into foreground pixels or background pixels; and according to the foreground pixels and the background pixels, which are classified, calculating probability values with which one or more pixels corresponding to the undetermined strokes belong to the foreground strokes, determining the pixels of which the probability values are smaller than lower limit as the background pixels, and determining the pixels of which the probability values are greater than upper limit as the foreground pixels. An object in the image is interactively segmented on the basis of various marks, so complicated boundary situation can be handled, and a high-accuracy segmentation result is obtained.

The invention discloses a remote sensing image segmentation method based on a disparity map and a multi-scale deep network model, which mainly solves the problems of low segmentation precision and weak robustness of the existing remote sensing image segmentation method. The implementation scheme includes: reading a data set, and generating a training data set of remote sensing image segmentation;constructing a multi-scale fusion segmentation network model; using the training data set to train a segmentation network model, and storing seven models with different iteration times; obtaining seven different segmentation result graphs by using the stored segmentation network model; carrying out majority voting on the seven different segmentation result graphs, and carrying out super-pixel processing on the voted result graph to obtain a preliminary segmentation result graph; obtaining a disparity map of the test scene by using an SGBM algorithm; and optimizing the preliminary segmentationresult graph by using the disparity map to obtain a final segmentation result. Compared with an existing method, the method has the advantages that the segmentation precision and robustness are obviously improved, and the method can be widely applied to urban and rural planning and intelligent urban construction.

The invention discloses a tooth orthodontic treatment monitoring method based on an intraoral image and a three-dimensional model. The tooth orthodontic treatment monitoring method comprises the following steps that step 1, a preprocessed intraoral image and a dental jaw three-dimensional digital model of a patient at a P1 stage are obtained; step 2, tooth-gingiva and tooth-tooth segmentation is performed on the dental jaw three-dimensional digital model based on the intraoral image of the patient in the P1 stage, a segmented dental jaw three-dimensional digital model is obtained; step 3, a preprocessed intraoral image of the patient in a P2 stage is obtained; and step 4, the segmented dental jaw three-dimensional digital model is converted according to the intraoral image of the patient in the P2 stage, and a dental jaw three-dimensional digital model of the P2 stage and coordinate transformation data and morphological change data of the dental jaw models from the P1 stage to the P2 stage are generated. According to the tooth orthodontic treatment monitoring method based on the intraoral image and the three-dimensional model, coordinate transformation and form change conditions ofteeth of the patient after tooth orthodontic treatment can be known only by shooting the intraoral image, so that a scanner is prevented from being used for scanning an oral cavity after treatment.

The invention relates to an AMD lesion OCT image classification segmentation method and system based on a bidirectional guide network, and the method comprises the following steps: obtaining an OCT image, and dividing the OCT image into a training set, a verification set and a test set; constructing a mask complementary convolutional neural network for classification of OCT images; adopting a Grad-CAM algorithm to calculate a class activation graph of the mask complementary convolutional neural network, and obtaining the output of the class activation graph; constructing a class activation graph guided U-shaped segmentation network for segmenting a lesion area in the OCT image; training the network through the training set and the verification set to obtain an optimized mask complementary convolutional neural network and a class activation graph guided U-shaped segmentation network; and substituting the test set into the optimized mask complementary convolutional neural network and the U-shaped segmentation network guided by the class activation graph to realize classification and segmentation of the OCT image. According to the method, OCT images containing glass membrane warts, CNV and normal retinas can be accurately classified, and an accurate segmentation result of a lesion area is generated.

The invention provides a three-dimensional shape segmentation method and system based on weight energy adaptive distribution. The method comprises the steps of training a deep neural network and performing a segmentation prediction process on a to-be-segmented three-dimensional model, the training process comprising the steps of: segmenting the three-dimensional model into n small blocks, randomlyselecting a triangular patch on each small block to represent the small block, and determining a segmentation label corresponding to each triangular patch through a segmentation label; extracting a feature vector of each triangular patch; calculating the minimum value of the geodesic distances of the triangular patches under the same three-dimensional model through the segmentation labels to obtain weight energy distribution, calculating and obtaining the soft label of each triangular patch, and taking the soft labels of the triangular patches under all three-dimensional models as the outputof deep neural network training; and training a deep neural network with a random inactivation layer by using the input and the output. The method has the advantages of high accuracy, strong robustness, strong learning expansion capability and the like.

The invention discloses a remote sensing image building accurate segmentation method. The method comprises the following steps: constructing a building extraction network comprising a feature extraction module, a cavity convolution module, an attention module, an up-sampling module and a convolution prediction module; based on the training sample set, adopting a multi-scale composite loss function combining Dice Loss and BCE Loss to train the constructed building extraction network; and inputting a remote sensing image to be extracted into the trained building extraction network to obtain a building extraction result. The method has the remarkable effects that the feature learning and generalization ability is high; the network is low in complexity and easy to train; and the building extraction precision is high.

The invention discloses a semantic segmentation method based on an efficient convolutional network and a convolutional conditional random field. The method comprises the following specific steps of: 1, inputting an RGB image with any size, and performing semantic extraction on the original RGB image by adopting an encoder network consisting of a down-sampling module and a one-dimensional non-bottleneck unit to obtain a matrix consisting of characteristic patterns; 2, adopting a deconvolution layer and a one-dimensional non-bottleneck unit to semantically map the discriminative features learned by the encoder network to a pixel space to obtain a dense classification result; and 3, adopting the convolutional conditional random field network layer and pixel point information of the original RGB image and pixel point classification information obtained by the decoder network to classify pixel point semantic features again, so that the purpose of output result optimization is achieved. A brand new coding and decoding network is adopted to classify the end-to-end pixel points, and a segmentation result is re-optimized through a convolutional conditional random field network with high use efficiency.

The invention discloses a ship target segmentation method based on a multi-scale similarity guidance network, and mainly solves the problem that the ship target segmentation result is poor under the condition of small samples in the prior art. According to the scheme, the method comprises the following steps: constructing an original data set by using existing SAR image ship target segmentation data sets which are in different regions and contain different imaging modes; constructing the original data set into a small sample segmentation training data set and a small sample segmentation test data set; constructing a multi-scale similarity guidance network composed of a feature extraction branch for supporting the image, a feature extraction branch for querying the image, a similarity guidance module and a generation branch; training the network by using a small sample training set; and inputting the small sample test set into the trained network to obtain a segmentation result of the ship target. Compared with other small sample semantic segmentation methods, the method has the advantages that the number of data needing to be labeled on the target domain is effectively reduced, and the small sample semantic segmentation effect is improved. The method can be used for intermediate processing of SAR image interpretation.

The invention discloses a small sample image segmentation method based on a guide network and a full-connection conditional random field, and the method comprises the steps: carrying out the group division of an obtained to-be-segmented image, and obtaining a support image and a query image; marking positive sample points and negative sample points in the support image to obtain a foreground information feature map and a background information feature map containing positive and negative sample positions; based on the support image, the foreground information feature map and the background information feature map, extracting task features by adopting a guide network; performing preliminary segmentation according to the task features and the query image to obtain a preliminary segmentation result; carrying out edge refinement on the preliminary segmentation result based on a full-connection conditional random field to obtain a final segmentation result; inferring potential features of the support image by optimizing the guide network; performing preliminary segmentation on the query image without pixel annotation according to the potential features; and according to the preliminary segmentation result, carrying out finer segmentation through a full-connection conditional random field so as to obtain a relatively high segmentation result.

The invention discloses a two-dimensional image auxiliary segmentation method based on three-dimensional point cloud of an overhead contact system cantilever system. The method comprises the followingsteps: step 1, acquiring three-dimensional point cloud data of the overhead contact system cantilever system; step 2, performing uniform resampling on the three-dimensional point cloud data by adopting voxel filtering; step 3, converting the uniformly resampled three-dimensional point cloud data of the cantilever system into a two-dimensional image; step 4, segmenting the two-dimensional image obtained in the step 3, then sequentially carrying out image closed operation and median filtering processing, and returning to the three-dimensional point cloud to obtain a segmentation result of eachlinear part of the three-dimensional point cloud of the cantilever system. According to the method, two-dimensional images are adopted for auxiliary segmentation, efficient segmentation results of alllinear parts of the overhead contact system cantilever system are finally returned, the method has the advantages of being good in noise immunity, high in robustness and high in precision, and the segmentation effect of the overhead contact system cantilever system is improved; and the consumption of manpower and material resources is reduced, and the influence of weather and experience judgmentof operators is avoided.

The invention belongs to the technical field of point cloud instance segmentation, and discloses a label optimization point cloud instance segmentation method. The method comprises the steps: firstly carrying out the feature extraction of a point cloud through a graph convolution neural network; establishing an instance label matrix for the training set, and performing label propagation on the instance label matrix by using a label propagation algorithm to obtain an optimized instance label matrix; and finally, performing instance segmentation on the point cloud by combining the label matrix and the optimized instance label matrix. According to the method, the similarity relationship of the points in the point cloud is fused, so that the supervision of the label matrix has better distinguishing representation, and a better segmentation result is obtained. According to the method, the global shape information and the local feature information of the large-scene point cloud model are considered, and the effect of integrating the global information is achieved; a label propagation algorithm is used for optimizing a label matrix, and the point cloud segmentation result is improved through the combined action of original data set labels and optimized labels.

The invention discloses a plant leaf segmentation method, and relates to the field of image processing. The method comprises the steps that a sample data set is constructed, sample images in the sample data set are inputted into a convolutional neural network, the convolutional neural network comprises a Backbone network, an RPN network and a plurality of cascaded blade segmentation modules, each blade segmentation module comprises a ROIAlign network and a Head network, each Head network comprises a classification branch, a segmentation branch and a detection branch, a plant leaf segmentation model is obtained through training of the sample data set based on a convolutional neural network, a to-be-segmented image is input into the plant leaf segmentation model, a leaf segmentation result of the to-be-segmented image is obtained, and the to-be-segmented image can adopt a multi-scale segmentation strategy. The method can be used for effectively segmenting sheltered leaves, unclear-edge leaves and small-scale leaves, and the application of deep learning in the field of plant leaf segmentation is promoted.

The invention discloses a semantic segmentation method based on an efficient convolutional network and a convolutional conditional random field. The specific steps of the present invention are as follows: 1. Input an RGB image of any size, and use an encoder network composed of a down-sampling module and a one-dimensional non-bottleneck unit to extract semantics from the original RGB image, and obtain a matrix composed of feature maps; 2. Using a deconvolution layer and a one-dimensional non-bottleneck unit, the discriminative features learned by the encoder network are semantically mapped to the pixel space to obtain dense classification results; 3. Using a convolutional conditional random field network layer, combined with the original The pixel information of the RGB image and the pixel classification information obtained by the decoder network classify the semantic features of the pixels again, so as to achieve the purpose of optimizing the output results. The present invention uses a brand-new encoding and decoding network to classify pixel points end-to-end, and re-optimizes the segmentation result by using a highly efficient convolution conditional random field network.

The present application discloses a brain injury region segmentation method and apparatus for magnetic resonance images acquired on different scanning devices. The method includes receiving a first nuclear magnetic resonance image to be processed on a first scanning device and a second nuclear magnetic resonance image to be processed on a second scanning device; Converting the first to-be-processed nuclear magnetic resonance image and the second to-be-processed nuclear magnetic resonance image to a common domain by a converter respectively; And performing brain injury segmentation on the converted MRI image in the common domain. The present application solves the technical problem of poor effect of lesion segmentation processing on magnetic resonance images from different scanning apparatuses. The present application can better segment brain lesions on MR images acquired on different scanning devices, thereby accurately detecting cerebral infarction regions.

The invention provides an interactive image segmentation method and an interactive image segmentation system based on a concentrated mistaken segmentation area, comprising the following steps of: carrying out foreground and background matting processing on initial segmentation of an input image to obtain foreground and background matting images; generating an under-segmentation and over-segmentation geodesic distance guide map for the input image and the under-segmentation and over-segmentation indication points; according to the input image, the initial segmentation image and the under-segmentation and over-segmentation geodesic distance guide image, extracting the characteristics of the whole image; extracting under-segmentation and over-segmentation region features according to the background and foreground matting images and the under-segmentation and over-segmentation indication points; and performing feature fusion on the under-segmentation and over-segmentation region features and the full-image features to obtain a corrected segmentation image. According to the method, the priori knowledge and the learning ability of the neural network are combined, the accuracy and interpretability of image segmentation are improved, the method serves as a means for obtaining segmentation data annotation, annotation can be completed only through several times of click interaction, andpixel-by-pixel annotation is avoided.

The present application discloses a method and a device for segmenting brain injury regions in multi-equipment cerebral infarction MRI images. The method includes receiving a first MRI image of cerebral infarction to be processed on a first scanning device and a second MRI image of cerebral infarction to be processed on a second scanning device; The resonance image and the second MRI image of cerebral infarction to be processed are converted to a common domain; and brain injury segmentation is performed on the converted MRI image of cerebral infarction in the public domain. The present application solves the technical problem that the lesion segmentation processing effect on cerebral infarction MRI images from different scanning devices is poor. The application can better segment brain lesions on MR images collected on different scanning devices, and then accurately detect cerebral infarct regions.

The invention discloses a multi-component semantic segmentation method based on a four-cavity cardiac tangent plane of a multi-disease fetus. The method comprises the steps of collecting an ultrasoniccardiac sequence image, copying and expanding the sample size of the collected multi-position ultrasonic cardiac sequence image, and segmenting the sample size; segmenting the plurality of componentsbased on the four-cavity heart tangent planes of the plurality of disease types; using a proportion balance strategy to obtain a better segmentation effect, especially in a group with a low data volume proportion. The method is the first method for achieving segmentation of multiple core components in the ultrasound cardiogram of the fetus with multiple diseases, greatly assists doctors in statistical analysis of multiple key indexes of the heart of the fetus, and is beneficial to remote medical treatment, especially to improvement of the medical level of grassroots remote areas.