31results about How to "Reduce misclassification rate" patented technology

The invention belongs to the technical field of surface defect detection, and particularly provides a synthesis method of a neural network training sample in part surface defect detection. The training sample synthesis method comprises the following steps: S1, acquiring an image of a defective part sample; S2, acquiring a defective image from the image of the defective part sample; And S3, extracting image features of the defective image, and adding the disturbance into the image features to generate a training sample. Through the method, the training sample is obtained; only a small number ofdefective parts need to be obtained; A small number of images of the surface of a defective part are obtained; flaws existing in the image are extracted to obtain images of various flaws; According to the method, the image features of the flaws are extracted from the images with various flaws, and then corresponding disturbance is added to the image features to generate a large number of trainingsamples, so that the training requirements of the neural network are met, and the problems that the training samples for neural network training are difficult to obtain and a large number of trainingsamples cannot be obtained are solved.

The invention discloses an SAR image varying detecting method based on artificial immunity multi-target clustering, and mainly solves the problems of low accuracy and low efficiency of an SAR image varying detection result. The realizing steps of the method comprise: (1) reading-in two time-phase SAR images; (2) constructing difference images of the two time-phase SAR images; (3) performing gray value-based self-adaptive immunity multi-target clustering on the difference images, and dividing the images into a varying type, a non-varying type and a to-be-identified type; (4) carrying out non-subsample wavelet conversion-based immunity clone multi-target clustering on the to-be-identified type to obtain a group of clustering center of the to-be-identified type; (5) performing minimum distance classifying on the to-be-identified type according to the group of clustering center to obtain one group of varying detection outcome images; (6) calculating the target function values of the varying detection outcome images; (7) selecting the minimum target function value according to the target function value; and (8) taking the varying detection result corresponding to the minimum target function value as the final detection result. The SAR image varying detecting method based on the artificial immunity multi-target clustering has the advantages of high detection efficiency and high detection precision.

The invention belongs to the technical field of surface flaw detection, and specifically provides a method for synthesizing training samples of a neural network in part surface flaw detection. The method for synthesizing training samples includes: Step S1: Obtain an image of a sample with defects; Step S2: Obtain a defect image from the image of a sample with defects; Step S3: Extract the image features of the defect image and add disturbances to the image features to generate Training samples. Obtaining training samples through this method only needs to obtain a small number of parts with defects, and by obtaining a small number of images of the surface of defective parts, extract the defects existing in the image to obtain images of various defects, from images of various defects Extract the image features of defects, and then add corresponding perturbations to each image feature to generate a huge number of training samples, which meets the training needs of the neural network and solves the difficulty of obtaining training samples for neural network training, and it is impossible to obtain a large number of training samples The problem.

The invention discloses a multi-objective SAR image segmentation method based on feature fair integration, the realization steps of which are: (1) reading in a synthetic aperture radar SAR image; (2) extracting a grayscale feature map; (3) extracting grayscale co-occurrence Matrix feature map; (4) synthetic total feature map; (5) normalization processing; (6) calculation of superpixel features; (7) population initialization; (8) feature fair integration; (9) calculation of individual fitness in the population value; (10) optimize the fitness value; (11) segment the image. The invention adopts multi-feature fair integration, reduces the misclassification rate of image segmentation, and improves the accuracy of image segmentation.

The invention discloses a face expression recognition method based on sparse representation based on double dictionaries and multi-feature fusion decision-making. First, features are extracted from face image samples without expression and face images with specific expressions, and a nominal dictionary and a feature dictionary are constructed according to the features. ; For the image to be recognized, by extracting the corresponding features, use the nominal dictionary to perform sparse coding on it, and then combine the coding coefficient results with the nominal dictionary to obtain the reconstructed expressionless image features, and subtract the features before and after reconstruction The features containing only the information of expression characteristics are obtained, and the feature dictionary is used to sparsely encode the features to obtain the coding coefficient vector; based on the feature dictionary, an auxiliary decision-making fusion dictionary is trained for different types of features, and based on the sparse representation, the different types of features are calculated. The encoded coefficient vectors are classified and judged, and the judgment results of various features are obtained; the final recognition results are obtained by voting; this method can effectively overcome the influence of face, illumination, occlusion and other changes on expression recognition.

The invention discloses a gray level image segmentation method based on multi-objective fuzzy clustering, relating to the technical field of image processing and mainly solving the problem of lower accuracy rate of gray level image segmentation. The gray level image segmentation method comprises the steps of: after graying an image, randomly generating a plurality of clustering centers according to a generated grey level histogram, and constituting the clustering centers into a parent antibody population. The gray level image segmentation method is characterized in that a dense separation effectiveness function as an evaluation criteria is combined with a fuzzy optimization function in a fuzzy C-mean value method to form a multi-objective optimization problem, the whole parent population is iterated for multiple times by adopting an immune clone multi-objective evolutionary algorithm, simultaneously searched from multiple directions, and calculated in parallel so as to finally acquire an optimum clustering center, and a classifying result is output. Therefore, the detail information in the gray level image is effectively reserved, the wrong fraction is reduced, the gray level image segmentation precision is improved, and a good platform is provided for subsequent operation of gray level image segmentation. The gray level image segmentation method can be used for extracting and obtaining the detail information of the gray level image.

The invention provides a gray scale image segmentation method based on a sequencing K-mean algorithm according to the defect of difficulty in retaining image details in a category with more pixels in existing K-mean algorithm. The method comprises the following steps: (1) reading in a noise-free gray scale image G and randomly assigning each cluster center; (2) calculating a histogram HL of the read-in gray scale image G; (3) calculating the distances of each gray scale to each cluster center; (4) sequencing the distances of each gray scale to each cluster center; (5) storing the sequenced distances; (6) assigning each gray scale to a cluster center category which is nearest therefrom; (7) updating the cluster centers according to the sequenced distances of each gray scale to each cluster center; and (8) determining whether an iteration stopping condition is achieved according to the updated cluster centers and the non-updated cluster centers, if yes, terminating clustering and outputting the results of clustering to complete image segmentation. The method provided in the invention has the advantages of high precision of image segmentation and capability of being used to extract and obtain detail information of a gray scale image.

The invention discloses a pixel number clustering-based fuzzy C-average value gray level image splitting method, and mainly solves the problem of low accuracy of splitting of the gray level image. The method is realized by the steps of (1) reading a gray level image and counting a gray level histogram; (2) randomly initializing a clustering center; (3) calculating the Euclidean distance between each gray level and each clustering center; (4) calculating the total number of the pixels contained between each gray level and each clustering center by the Euclidean distance; (5) judging the type of each gray level by the total number of the pixels to obtain a classified result; (6) calculating each type of gray level average value by the classifying result to be used as a new clustering center; (7) calculating a membership matrix according to the clustering center; (8) updating the clustering center by the membership matrix; (9) repeating the steps (3)-(8) until the terminal condition is met, and outputting an updated clustering center; and (10) classifying the gray images by the updated clustering center to obtain a splitting result image. The pixel number clustering-based fuzzy C-average value gray level image splitting method has the advantage of high image splitting precision and can be used for extracting the detail information of the gray level image.

The invention discloses a multi-spectral image classification method based on threshold self-adaptation and convolutional neural network, which inputs multi-spectral images of satellites to be classified in different time phases and different bands, and converts the marked parts of the same band images of all cities All pixels are normalized; the selected 9 bands are stacked into an image as a training data set; a classification model based on convolutional neural network is constructed, and the training data set is used to train the classification model to obtain a probability model based on OSM. The model and the confidence strategy adjust the softmax output results to obtain the final classification model, and finally upload the test results to the IEEE website to obtain the classification accuracy. The multi-spectral image classification method provided by the present invention makes full use of the characteristics of multi-spectral image with multiple bands, large data volume, and large information redundancy, and solves the problem that it is difficult to classify complex types of ground objects, and can not only improve the classification accuracy , Reduce the misclassification rate, and can also improve the classification speed.

The invention provides a blood vessel extraction method for a live tumor angiogenesis image under an optical resolution photo-acoustic microscope (OR-PAM), and belongs to the field of image processing. First, a tumor angiogenesis image is preprocessed using a variety of enhanced algorithms, which enhances the contrast of small blood vessels effectively and reduces the influence of noise to the greatest degree. Then, a threshold is selected automatically to binaryze the image using an Otsu algorithm, which reduces the mis-classification probability. Finally, the centerline of a blood vessel is measured using a two-step fast marching method, so the calculated centerline can be located in the blood vessel more accurately. Blood vessel extraction using the traditional level set blood vessel extraction algorithm is too slow, while the running speed of the blood vessel extraction algorithm based on the fast marching method is increased greatly. Experimental results show that the method can be used to extract the blood vessel information of a tumor angiogenesis image more effectively, and has a good extraction effect for the information of small blood vessels.

The invention discloses a cross-belt sorting method, a sorting control system and a cross-belt sorting system, wherein the sorting method includes S1, receiving the top surface image of the sorting trolley, and determining the state of the sorting trolley; S2, determining the state of the sorting trolley When there is only one item on the trolley, determine the sorting slot corresponding to the barcode information on the item and its position coordinates on the sorting trolley; The coordinates of the four vertices of the smallest circumscribed rectangle projected on the sorting cart; S4, calculate the sorting start point based on the virtual coordinates and the target sorting grid; S5, control the sorting cart to start when the sorting cart reaches the sorting start point sorting. This scheme considers the position of the item on the sorting trolley, and accurately calculates the sorting start point based on the minimum circumscribing rectangle of the item, which can effectively improve the accuracy of sorting and reduce the system misclassification rate from one thousandth to ten thousand Less than one-third, greatly improving the reliability of sorting.