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

The invention discloses an intelligent garbage classifying method based on machine learning. The intelligent garbage classifying method based on machine learning comprises the following steps that anautomatic garbage classifying system is provided and comprises a garbage sorting device, a plurality of garbage collecting devices, a plurality of sensors and a computer; a plurality of garbage categories to be divided are preset on the computer, garbage data corresponding to the garbage categories is input, and based on the input garbage data, a machine learning model is used for establishing corresponding garbage characteristics of the various garbage categories; according to the multiple garbage categories given by the computer in a default manner, a user selects the needed garbage category; materials to be classified are put into the garbage sorting device, the sensors collect the data of the materials, and the computer carries out characteristic extraction on the collected data, compares the extracted material characteristics with the garbage characteristics in the model and automatically distinguishes garbage category division results of the materials and the result credibility;and according to the garbage category division results of the materials, the garbage sorting device automatically sorts the materials to the corresponding garbage collecting devices.

The invention discloses a deep thin interbedded reservoir quantitative characterization method based on seismic grading sensitive attribute fusion, which effectively increases the efficiency of analyzing deep thin interbedded reservoir seismic attributes, and greatly increases prediction precision. The deep thin interbedded reservoir quantitative characterization method comprises the steps of: extracting reasonable small time window segment attributes of a geologic target of an interest interval in a regional mode on the geological background of large time window segment, carving a favorable reservoir finely to be fused in the overall background, and acquiring an attribute prediction map of the geologic target. The deep thin interbedded reservoir quantitative characterization method comprises two key steps that: 1, performing multi-attribute correlation dimension reduction by utilizing conventional seisms and optimizing target sensitive attributes of the interest interval; 2, grading the attributes based on seismic grading firstly, determining reasonable small time window segments of the target region on different backgrounds, refusing the attributes, and predicting reservoir range in a research region quantitatively by refusing the graded sensitive attributes under small time window parameters. The grading and fusion are implemented by adopting a support vector machine (SVM) algorithm under real drilling data constraints.

The invention discloses a method for edge detection of a digital image, so that a technical problem that a current Canny operator edge detection method is restricted by a selected threshold can be solved. According to a technical scheme employed in the invention, a gray value matrix of an image is obtained and gray values are normalized to an interval [0, 1]; the size of a template is defined and memberships um (x (i, j)) of all pixels in the template are calculated according to a membership function; and fuzzy entropies of all pixels in a window are calculated according to a fuzzy entropy formula, and fuzzy entropy information measure is calculated and is compared with a threshold so as to determine whether there is a boundary. Because a pixel point in the template is used as a processing unit and a fuzzy algorithm is utilized, noises in a filtered image are reduced and a fuzzy phenomenon existing in the filtered image is significantly improved. Compared with a threshold segmentation method in a background technology, the provided method has the following advantages: there is no need to select a threshold; an error rate of a segmented image is low; and a visual effect of the image can be substantially improved.

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 improved hyperspectral image classifying method comprising the following steps of: (1) selecting the spatial neighbor of the central pixel of a hyperspectral image to be classified to obtain a high-quality spatial neighbor area by using a selection strategy based on watershed segmentation area or a minimal spanning tree neighbor selection strategy; (2) extracting the spatial feature of the high-quality spatial neighbor area; and (3) predicting a spectral space joint model to fuse a spatial feature and a spectral feature in a synthetic kernel way, to train a classification model, and to predict a hyperspectral image test set tag. The method selects different spatial neighbor selection modes, satisfies requirement for speed and precision in hyperspectral classification, solves a problem of same object with different spectra and different objects with same spectrum in the hyperspectral image by mining and utilizing spatial information, enhances the robustness of the original spectrum classification in virtue of high-quality spatial neighbor and feature information, and has a high use value.

The invention discloses a feature fair integration-based multiple target SAR image segmentation method. The method comprises the steps that 1, a synthetic aperture radar SAR image is read; 2, a gray-scale feature map is extracted; 3, a gray-scale co-occurrence matrix feature image is extracted; 4, a total feature image is synthesized; 5, normalization processing is conducted; 6, superpixel features are calculated; 7, puperpixel characteristics is conducted; 8, feature fair is integrated; 9, an individual fitness value in a population is calculated; 10, the fitness value is optimized; 11, the image is segmented. Accordingly, multi-feature fair integration is adopted, the wrong segmentation rate of image segmentation is lowered, and the accuracy of image segmentation is improved.

The invention discloses a building information extraction method and system based on rule and sample fusion, and the method comprises the steps: obtaining worldview 2 single-chip data of a corresponding region; preprocessing the extracted worldview2 single-chip data, so as to obtain the preprocessed worldview2 single-chip data; according to the preprocessed worldview 2 single-chip data, carrying out object-oriented building information extraction based on samples, and obtaining first building area information; performing rule-based object-oriented building information extraction according to the preprocessed worldview 2 single-chip data to obtain second building area information; and fusing the first building area information and the second building area information to obtain building areafusion extraction information. In the embodiment of the invention, the building without obvious errors is extracted, and the building is good in intuitive feeling, clear in contour and regular in shape.

The invention discloses a cross belt sorting method, a sorting control system and a cross belt sorting system. The sorting method comprises the steps that S1, a top surface image of a sorting trolleyis received, and the state of the sorting trolley is determined; S2, when determining that only one object exists on the sorting trolley, a sorting grid corresponding to bar code information on the object and position coordinates of the sorting grid on the sorting trolley are determined; S3, virtual coordinates are converted according to the position coordinates, wherein the virtual coordinates are four vertex coordinates of a minimum enclosing rectangle of projection of the object on the sorting trolley; S4, a sorting starting point is calculated according to the virtual coordinates and the target sorting grid; and S5, when the sorting trolley reaches the sorting starting point, the sorting trolley is controlled to start sorting. According to the scheme, the position of the object on thesorting trolley is considered, the sorting starting point is accurately calculated based on the minimum enclosing rectangle of the object, the sorting precision can be effectively improved, the systemerror sorting rate is reduced to be lower than one ten thousandth from one thousandth, and the sorting reliability is greatly improved.

The invention belongs to the technical field of art design application, and discloses an art design application platform based on a new media platform and a data information processing method. The artdesign application platform based on the new media platform comprises an input module, a new media data acquisition module, a main control module, a design module, a three-dimensional modeling module, an animation generation module, an art resource testing module, a publishing module and a display module. Through the animation generation module, a large amount of time consumed when art modifies the two-dimensional animation is completely saved, and the requirement for art is greatly reduced. Meanwhile, whether the art resources conform to the specifications, whether the art resources conformto the design requirements and whether some program editing errors exist or not is automatically detected through the art resource testing module, and the workload of testers in the art resource detecting process is greatly reduced.

The invention discloses a double-dictionary and multi-feature fusion decision facial expression recognition method based on sparse representation, and the method comprises the steps of firstly, extracting features from an expression-free face image sample and a specific expression face image sample, and constructing a nominal dictionary and a feature dictionary according to the features; Image tobe identified, extracting the corresponding features, performing sparse coding on the features by adopting a nominal dictionary, combining a coding coefficient result with the nominal dictionary to obtain reconstructed expression-free image features, subtracting the features before and after reconstruction to obtain the features only containing expression feature information, and performing sparsecoding on the features by adopting the feature dictionary to obtain a coding coefficient vector; training auxiliary decision fusion dictionaries for different types of features on the basis of the feature dictionaries, and performing classification judgment on the coding coefficient vectors calculated from the different types of features on the basis of sparse representation to obtain judgment results of the various types of features; obtaining a final identification result in a voting mode. The method can effectively overcome the influence of face, illumination, shielding and other changes on expression recognition.

The invention relates to a face pore detection method based on a deep neural network, and the method comprises the steps of: firstly constructing a face pore sample detection data set, and employing a deep semantic segmentation network to detect face pores to obtain a preliminary pore detection result graph; then predicting a face partition through a deep semantic segmentation network, and performing post-processing on a partition graph; performing post-processing on the preliminary pore detection result image in combination with the face partition to obtain a final pore detection result; and on the basis, calculating the number of pores in each partition in combination with face partition and pore detection results. According to the invention, the problems of high omission ratio, high error detection rate, insufficient algorithm generalization ability and the like of the existing face pore detection technology are solved.