51results about How to "Precise Segmentation Effect" patented technology

The invention discloses a remote sensing image ship integrated recognition method based on deep learning. The method comprises the steps of image classification, target detection and image segmentation. Compared with the prior art, the method has the advantages that the modern artificial intelligence deep learning model is combined with the traditional image processing method to detect and segmentthe ship with the eye remote sensing image; The remote sensing image segmentation method based on deep learning can accurately identify ships in a sea area, is suitable for various processing environments, has a good anti-interference capability for a complex environment, and can accurately segment the detected and identified ships.

The invention relates to a convolutional neural network road scene classification and road segmentation method. A dual-task joint structure model based on the convolutional neural network is established.and composed of an encoder and two decoders. Feature information sharing is achieved through end-to-end training.Classification of various road scenes such as urban roads, rural roads and highwaysis completed, road drivable areas in the scenes are segmented, and the dual goals of road scene classification and road area segmentation are achieved. Based on a convolutional neural network, scene classification and road extraction in road scene perception are combined. Real-time output is obtained through an end-to-end training mode, the functions of an intelligent driving assistance system areimproved, and an effective and reliable driving assistance function is provided. A mode of combining a full convolutional network and a conditional random field is adopted for a part extracted from adrivable area in the model, a high-resolution image optimization output result can be generated, and a relatively accurate segmentation effect is achieved.

The invention relates to the field of medical image processing and pattern recognition, and provides a computer-assisted lump detecting method based on a mammary gland magnetic resonance image. The computer-assisted lump detecting method based on the mammary gland magnetic resonance image aims at solving the problems that in the prior art, the lump partition effect is not good, and the accuracy, the sensitivity and the specificity in a classification test are not high. The computer-assisted lump detecting method includes the following steps: S1, extracting an interest area from the mammary gland magnetic resonance image; S2, extracting an initial lump area from the interest area in a separated mode, and determining the contour line of the initial lump area; S3, calculating the weight distribution of characteristic parameters of the initial lump area; S4, selecting the characteristic parameters, with the weight coefficients larger than a standard weight coefficient, of the initial lump area, and carrying out training classifying to obtain optimized characteristic parameters; S5, inputting the optimized characteristic parameters into a classifier, analyzing the optimized characteristic parameters with a support vector machine classification method, determining a final lump area, and displaying the final lump area for a user. The detecting method has the good lump partition effect, the accuracy, the sensitivity and the specificity in the classification test are effectively improved, the detecting result serves as a second opinion to be provided for a radiologist, and the misdiagnosis rate and the missed diagnosis rate of the radiologist can be effectively reduced.

The invention provides a remote sensing image thin and weak target segmentation method. Firstly, data enhancement and corresponding preprocessing are carried out on an original remote sensing image, U-net is improved by means of the dense connection thought of DenseNet, and a Dense-Unet network structure is provided. Dense convolution is used in a network structure, the cascade relation between convolution channels is enhanced, through a symmetric structure and a jump connection thought, the connection between features of all layers is further tighter, and thin and weak target features can belearned more effectively. In order to ensure the real-time performance of final network identification and reduce the parameter quantity, a bottleneck layer and a batch normalization layer are introduced behind each dense block. And the objective function is adjusted by using the cost-sensitive vector weight, so that the problem of unbalanced segmentation target categories is solved, and the segmentation precision is further improved. And finally, a plurality of independent models are trained by using an ensemble learning method, the independent models are combined, and target category information is jointly predicted in the picture.

The invention relates to a fabric flaw detection method based on wavelet transformation and a genetic algorithm. The method comprises the following steps: preprocessing acquired images with flaws; obtaining multi-dimensional sub-images by performing wavelet decomposition on the preprocessed images; obtaining optimal flaw edge information by performing fusion on the sub-images; calculating a threshold for the sub-images by use of the genetic algorithm, and performing threshold segmentation on the fused images by use of the threshold; and performing morphological processing on the images after threshold segmentation. According to the invention, the cloth flaw segmentation effect after processing is accurate, the segmentation speed is fast, and original forms of the flaws are well reserved.

The invention relates to the technical field of computer vision, image processing, pattern recognition and the like and particularly relates to a multiple foreground object image interactive segmentation method. The image interactive segmentation method includes the steps of (1) constructing an image pixel similarity matrix, (2) acquiring image pixel label information; (3) constructing a spectral clustering segmentation model by combining the image pixel similarity matrix and the image pixel label information and solving to obtain a preliminary segmentation result, (4) constructing spatial smoothing constraint, and (5) constructing a markov random field model by combining the preliminary segmentation result and the spatial smoothing constraint and solving to obtain a final segmentation result. According to the multiple foreground object image interactive segmentation method, advantages of grabcut methods and linear constraint spectral clustering methods are integrated, simultaneously defects of the grabcut methods and the linear constraint spectral clustering methods are overcome, and images with randomly distributed colors and multiple foreground objects can be segmented merely by labeling an extremely small quantity of pixel points.

The invention discloses a partial differential equation-based nano-particle size measurement method. The method includes the following steps that: 1) a nano-particle image I is inputted, pixel-level multiplication is performed on the filtering result of a mean curvature flow model and a PM model, so that a filtered image u can be obtained; 2) a region scalable fitting (RSF) model is adopted to segment the image u; 3) pixel calibration is carried out, the actual size of each pixel in the image can be obtained; 4) inadherent particles are selected out by means of the convexity (Cconv) of a target; and 5) least-square circle fitting is performed on the boundaries of the particles, so that the diameters of spherical nano-particles can be obtained, and the diameter rc of the internally tangent circle of the nano-particles, the diameter ri of the externally tangent circle of the nano-particles, and the sphericity of the nano-particles are obtained, wherein the sphericity of the nano-particles can be represented by an expression S=ri / rc. The method of the invention can be widely applied to the high-tech fields such as catalytic science, medical drugs, new materials, electric power industry and compound materials which require nano-particle size measurement technology.

The invention belongs to the field of medical image processing and especially relates to a nucleus segmentation method based on white blood cell detection. The method comprises the following steps: to begin with, summarizing the size of diameter of each kind of white blood cell to set a plurality of sizes of sliding windows; then, selecting a candidate window for each sliding window according to gradient information of each white blood cell picture; calculating characteristic value of each candidate window according to the color contrast information, boundary density contrast information and gradient contrast information of the white blood cell pictures, and inputting the characteristic values to an epsilon-SVR model, and selecting a final detection window; intercepting the region of the final detection window of each white blood cell picture as a final positioning subgraph; and then, selecting threshold segmentation nucleus through a subgraph gray statistical chart polynomial curve fitting method, and then, carrying out hole filling and isolated point removal through morphologic processing to enable the segmentation result to be more accurate.

The invention discloses an electronic connector detection method based on machine vision, and belongs to the technical field of computer program design comprising image acquisition, analysis and processing. Actual targeted images are processed by various methods of image smoothing, image segmentation, image matching, edge extraction and color recognition, complete geometrical characteristics of defects of the electronic connector are acquired, and automatic detection can be realized. According to experiments, system detecting results is consistent with visual detecting results, accuracy of the detection is guaranteed, and detecting efficiency of the electronic connector is improved greatly.

The invention provides a tunnel three-dimensional geometric reconstruction method considering data driving segment segmentation and model driving segment assembly. The method comprises the following steps: (1) a mapping method from tunnel three-dimensional point cloud expansion to a two-dimensional image; (2) a tunnel image segmentation method; and (3) tunnel segment reconstruction based on modelmatching. The invention provides a method for reconstructing a three-dimensional tunnel geometric model by coupling data and a model driving thought, which is used for reconstructing a multi-scale shield tunnel geometric model. Based on a data driving thought, firstly, point cloud is converted into a binary image, and then tunnel point cloud is segmented on a loop-by-loop scale and a segment scalerespectively by utilizing morphology and a template matching algorithm. The sealing grooves and the bolt holes between the segments are more obvious in the binary image, and the image itself also contains the mapping topological relation between the original point cloud and the pixels. Based on the observation that each ring of a tunnel only comprises one capping block; a segment segmentation problem is converted into a template matching problem based on least square constraint of a binary image; and then, based on the thought of model driving, the reconstruction problem of the shield tunnelsegment on the scale is converted into the matching problem of the segment point cloud and the model base, the types, sizes and positions of various segments are obtained, and assembling of the tunnelsegments is completed.

The invention belongs to the technical field of computer vision, image treatment and pattern recognition, and particularly relates to a video image splitting method based on restrain spectral clustering and a markov random field. According to the method, a similarity matrix between two pixels is calculated based on the static characteristics of images, and the motion characteristic is added in a spectral clustering framework to be used as a restrain. In comparison with the conventional simple weighting method, two pieces of information with different reliability degrees are utilized well in the treatment method. Besides, as motion information coding is used as a restrain, only sparse point motion trains are required, so that reliable long-time motion information can be utilized to obtain an accurate and dense splitting result. Further, space smooth information of each pixel is coded into corresponding restrains by constructing a markov random field model, so that the video image splitting effect is relatively precise.

The invention provides a character segmentation method and device and a computer readable storage medium. The character segmentation method comprises the following steps: acquiring a character regionimage, and converting the character region image into a grayscale image; performing clustering analysis on the grayscale image by using a fuzzy C-means clustering algorithm, and performing binarization processing on the grayscale image according to a clustering analysis result to obtain a binary image; obtaining at least one character positioning block from the binary image by using a projection method; and performing character segmentation on the character region image according to the position information of the character positioning block. By utilizing the method and the device, a more accurate character segmentation effect can be realized for the character region image with poor image quality.

The invention provides a video target segmentation method based on motion attention, and the method comprises the steps: adding a channel feature map outputted by a channel attention module and a position feature map outputted by a motion attention module, and obtaining a segmentation result of a current frame, wherein the input of the channel attention module is the feature map Ft of the currentframe and an appearance feature map F0 of a target object provided by the first frame; enabling the channel attention module to calculate the association between the input feature map Ft and the F0 channel, wherein the output channel feature map reflects the object with the appearance closest to the target object in the current frame, the input of the motion attention module is the current frame feature map Ft and the position information Ht-1 of the target object predicted by the memory module in the previous frame motion attention network; enabling the motion attention module to calculate the association between the positions of the input feature map Ft and the Ht-1, wherein the output position feature map reflects the approximate position of the target object in the current frame. According to the invention, two factors of appearance and position are combined to realize more accurate segmentation of the video target.

The invention discloses a method for establishing an intracranial vessel simulation three-dimensional stenosis model based on transfer learning. The method comprises the following steps: acquiring a bright blood image group and an enhanced black blood image group of an intracranial vessel; preprocessing each bright blood image and the corresponding enhanced black blood image to obtain a first bright blood image and a first black blood image; performing image registration by using mutual information based on Gaussian distribution sampling and a registration method of an image pyramid; groupingthe registered bright blood images to obtain MIP images in all directions; obtaining a two-dimensional blood vessel segmentation image based on the MIP image and the fundus blood vessel image; performing back projection synthesis on the two-dimensional blood vessel segmentation image to obtain first three-dimensional blood vessel body data, and obtaining an intracranial blood vessel simulation three-dimensional model by utilizing the second three-dimensional blood vessel body data; and for each section of blood vessel in the model, obtaining a numerical value of a target parameter representingthe stenosis degree of the section of blood vessel, and marking the intracranial blood vessel simulation three-dimensional model by utilizing the numerical value of the target parameter to obtain a simulated three-dimensional intracranial blood vessel stenosis analysis model.

The invention provides an adaptive fuzzy C-means image segmentation method based on a potential function and mainly aims at solving the problems that classifying number needs to be previously set and the segmentation efficiency is low in the existing fuzzy C-means segmentation method. The method is realized by the following steps: (1) inputting a to-be-segmented image; (2) obtaining the histogram potential function and the maximum potential remnant height of the to-be-segmented image; (3) obtaining the histogram c-factorial remnant potential function of the to-be-segmented image; (4) combining pseudo potentials; (5) obtaining the clustering center and the classifying number of the to-be-segmented image; (6) carrying out fuzzy classification on the pixel dots of the to-be-segmented image; and (7) outputting an image segmentation result. The adaptive fuzzy C-means image segmentation method has the advantages of capability of obtaining optimal image classifying number in an adaptive way, high segmentation efficiency, strong region consistency in the segmentation result, smooth margin and the like. The adaptive fuzzy C-means image segmentation method can effectively segment natural gray images and infrared images and can be used for target recognition and tracking.

The invention relates to an image segmentation method based on an image data field. Conventional monochromatic image segmentation methods are generally based on the discontinuity and similarity of a pixel gray value; and a conventional segmentation algorithm is difficult to get the ideal segmentation effect under the condition of non-uniform illumination. Firstly, the input image is subjected to pre-processing filtering and then segmented by a watershed algorithm based on the eight-direction gradient, and then the average gray value of the edge point is obtained by the edge detection of the segmented image and is set as the seed value, and meanwhile, an image data filed is constructed, and the potential value function of the pixel point is calculated, and at the end, based on the seed value and the image data field, the feed filed automatic growth is carried out. The segmentation effect and the segmentation time of the image being nonuniform in illumination are greatly improved, and the demand of image identification can be met.

The invention discloses an intestinal tract lesion segmentation method combining a multi-scale U-shaped residual encoder and an overall reverse attention mechanism, and the method comprises the steps of taking the multi-scale U-shaped residual encoder as a backbone network to extract the features of an inputted intestinal tract lesion image, and introducing a multi-scale residual block for improving the segmentation reliability to generate an initial prediction image, wherein the U-shaped residual block filled in each stage of the backbone network can directly extract multi-scale features step by step under the condition of keeping a high-resolution feature map and reducing memory and calculation cost; enhancing the shallow features by using an overall attention mechanism which contributes to segmentation of the whole significant intestinal tract focus and refinement of more accurate boundaries to obtain an enhanced initial prediction graph; introducing a reverse attention mechanism for establishing the relationship between the region and the boundary clues to mine more boundary clues and make up the possible error part of an overall attention mechanism refining boundary. According to the invention, better intestinal tract lesion segmentation precision is achieved.

The invention relates to a probability graphic model image segmentation method based on flag fusion. The method comprises the steps of (1) acquiring a target image and multiple training images, conducting registration on the target image and the training images, and obtaining a corresponding deformation field; (2) acquiring the manual segmentation flag image of each training image subjected to registration, and obtaining a corresponding candidate flag according to deformation field reflection; (3) establishing a probability graphic model according to the training images and target image subjected to registration; (4) calculating the flag value of each pixel point in the target image by means of the local weighted voting flag fusion algorithm and the maximum posterior probability according to the probability graphic model, so that segmentation of the target image is achieved. Compared with the prior art, the method has the advantages that segmentation can be achieved automatically and segmentation precision is high.

The invention discloses a vegetable leaf disease image segmentation method and system and a computer readable storage medium. The vegetable leaf disease image segmentation method includes the following steps that 1, superpixel clustering processing is conducted on vegetable leaf images to obtain superpixel segmentation images; 2, significance region detection processing is conducted on the superpixel segmentation images to obtain lesion significance images; 3, the lesion significance images are processed to obtain non-lesion pixel sets; 4, region growing an region merging are conducted on pixel points in the non-lesion pixel sets to obtain lesion areas and normal areas of the vegetable leaf images. By the adoption of the vegetable leaf disease image segmentation method and system and the computer readable storage medium, leaf diseases in greenhouses can be rapidly and effectively detected and extracted under the conditions that no chemical reagents are in use and diseased vegetable leaves are not damaged, and therefore the vegetable leaf disease image segmentation method and system and the computer readable storage medium can be well applied in monitoring greenhouse vegetable diseases.

The invention discloses an eyelid topology morphological feature extraction method based on deep learning. The method specifically comprises the following steps: collecting an electronic digital photo of a normal person, processing the electronic digital photo, constructing an ROI image training set, and inputting the ROI image training set into a to-be-trained convolutional neural network to obtain a trained convolutional neural network; positioning an eye region-of-interest (ROI) position by using a facial recognition method for a to-be-detected electronic digital photo to obtain a to-be-detected ROI image; inputting the to-be-detected ROI image into the trained convolutional neural network to output an image with an eyelid contour line and a cornea contour line, determining a circular scale and a pupil center of the to-be-detected electronic digital photo, and extracting eyelid topology morphological characteristics of a single eye. According to the method, the eyelid and cornea structures are segmented by using the convolutional neural network; after the center of the pupil is determined by using Mean Shift clustering, the parameters of the related structure of the eyelid are automatically calculated, so that the accuracy equivalent to that of manual measurement is obtained.

The invention relates to an OCT (Optical Coherence Tomography) image segmentation method based on a random forest and a composite active curve, and is designed for accurately segmenting a retina layer and hydrops. The OCT image segmentation method based on the random forest and the composite active curve comprises the following steps that: obtaining OCT image features to train a random forest classifier, and obtaining a final SF1 through a composite active curve algorithm; extracting 24 features of the OCT image; and using the random forest classifier. The OCT image segmentation method based on the random forest and the composite active curve has the advantages of being simple in operation and accurate in detection results. The existing problems of low identification rate, poor segmentation effect and the like of a lesion OCT image segmentation algorithm are overcome.

The invention discloses a multi-scale-based local statistic active contour model (LSACM) level set image segmentation method. The offset field B epsilon, a variance sigma i epsilon and a level set function Phi(x) in an LSACM level set method are initialized. The quantity L(x) for describing a local area characteristic in a multi-scale LSACM method is calculated. A differential characteristic d(X) which describes a multi-scale local area is calculated. The maximum response M of a high pass filter in the multi-scale LSACM method is calculated. A local area simulation gray Ci epsilon is updated. The offset field B epsilon is updated. The variance sigma i epsilon is updated. The purpose of curve evolution is achieved through solving the partial differential equation minimum value corresponding to a multi-scale LSACM level set energy function. If a set number of iterations is achieved, the iteration operation is stopped, the curve evolution is ended, and if the number of iterations is not achieved, the iteration is continued. The invention provides the multi-scale LSACM level set method, a gray uneven image can be effectively segmented, and the phenomena of excessive segmentation and insufficient segmentation in an image segmentation method are improved.

The invention belongs to the technical field of cervical cancer cell screening, adn discloses an intelligent screening system and method for cervical cancer cell pathology negative exclusion. The intelligent screening system for the cervical cancer cell pathology negative exclusion comprises an image acquisition module, an input module, a main control module, an image processing module, a filtering module, an identification module, a data storage module and a display module. By means of the image processing module segmentation method from coarse to fine, on one hand, the processing speed is guaranteed, and on the other hand, an accurate segmentation effect is obtained. Meanwhile, the artificial recognition pressure is reduced through the recognition module, and the cost for artificially judging the probability of suffering from cervical cancer is also reduced. The identification efficiency of the image block (TCT digital slice) of the TCT slide scanning image is improved; evaluation iscarried out through a statistical machine learning model, cost is saved, and convenience and effectiveness are achieved.

The invention provides a malignant nodule edge detection image processing method based on a benign thyroid template. The method comprises the steps: learning an obtained thyroid three-dimensional image data set through a neural network, and forming a parameterized benign thyroid template; matching with the benign thyroid template according to the labeling information and the image information of the thyroid nodule ultrasonic image to be recognized, and performing coarse positioning on the nodule by combining the three-dimensional image and the feature information of the benign thyroid templateto obtain an initial contour curve of the nodule; adopting a local binary fitting model to iterate and evolve the initial contour curve to obtain an evolved contour curve; and performing effective contour selection on the evolved contour curve to obtain a final peripheral contour curve and a segmentation result, and performing fine positioning on the position of the nodule. The method is higher in autonomous recognition capability, better in adaptability to input of different thyroid images, better in segmentation effect, more accurate in numerical value implementation and higher in segmentation adaptability to images with uneven gray levels.

The invention discloses an image segmentation method and system based on fuzzy C-means and probability label fusion, and the method comprises the steps: carrying out the lesion segmentation of an obtained brain nuclear magnetic resonance image through employing a fuzzy C-means algorithm employing a superpixel as a clustering center, and obtaining initial lesion segmentation results of different modal images; and based on a multi-modal probability label fusion algorithm, performing probability label fusion on the initial lesion segmentation results of the different modal images to obtain a final lesion segmentation result. The segmentation of each mode of the nuclear magnetic resonance image is carried out by using an improved fuzzy C-means algorithm based on superpixels. And the segmentation advantages of different modes are fused to generate an optimal segmentation result. The segmentation of the three modes of the nuclear magnetic image has advantages, and the method fuses the segmentation results of different modes to obtain a more accurate segmentation effect.

The invention discloses a three-dimensional CNV growth prediction method and device and a quantitative analysis method, and the method comprises the steps: sending an original OCT sequence into a three-dimensional segmentation model, and segmenting an original CNV region; selecting a registration reference, sequentially pairing the registration reference with the original OCT sequence, sending theregistration reference into a three-dimensional registration model, and obtaining a registration parameter sequence; sequentially carrying out image transformation on the original OCT sequence by using the registration parameter sequence to obtain a registered OCT sequence; and sending the registered OCT sequence into a three-dimensional CNV growth prediction model to obtain a three-dimensional CNV prediction area. The three-dimensional CNV region can be accurately segmented, the CNV region can be accurately predicted, and the CNV growth region can be quantitatively analyzed and predicted.

The invention provides an image segmentation model training method, an image segmentation method and an image segmentation device. The method comprises the steps that firstly, a sample image library is acquired, each group of sample images in the sample image library is used for training an image segmentation model until it is determined that image segmentation model training is completed, and theexecuted training process comprises the steps that pixel values of pixel points of the sample images in the sample image library are adjusted, and a first image corresponding to the sample images isobtained; inputting the first image into an image segmentation model, and outputting a second image after the first image is segmented; calculating a loss value of image segmentation model training according to the second image and the mark image; when the loss value meets the preset condition, it is determined that training of the image segmentation model is completed, and through the method, theimage segmentation accuracy is improved.