42results about How to "Improve detection robustness" patented technology

The invention discloses a multi-task deep convolutional neural network-based vehicle color identification system. The system comprises a high-definition camera mounted above lanes on a road, a trafficcloud server and a vehicle color visual detection subsystem; the vehicle color visual detection subsystem comprises a vehicle locating detection module, a license plate locating detection module, a license plate background color identification module, a color difference calculation module, a vehicle color correction module and a vehicle color identification module; the vehicle locating detectionmodule, the license plate locating detection module and the vehicle color identification module share a same Faster R-CNN deep convolutional neural network; by adopting the deep convolutional neural network, vehicles on the road are quickly segmented; license plates on the road are quickly segmented by further adopting the deep convolutional neural network through using vehicle images; and space position information of the vehicles and the license plates in a road image is given. The multi-task deep convolutional neural network-based vehicle color identification system provided by the invention is relatively high in detection precision and relatively high in robustness.

The invention relates to a cloth defect detection method based on Fourier transform and image morphology. The cloth defect detection method comprises the following steps: (1) collecting an original image of cloth by using a linear array camera; (2) preprocessing the original image; (3) extracting image texture features by using LAWS texture filtering; (4) adopting a GMM classifier model to carry out defect pre-judgment; (5) constructing a band elimination filter by using a Gaussian filter; (6) generating a defect image by using Fourier transform and inverse transform; (7) extracting flaw positions and areas by adopting image morphology; And (8) outputting a cloth detection result. According to the method, more than ten types of cloth flaws including broken wefts, broken warps, broken holes, floating wefts and the like can be detected in real time, the detection speed is high, the accuracy rate is high, and the detection robustness is enhanced through the image processing method integrating the airspace and the frequency domain.

The invention provides a device for prompting blind persons to avoid obstacles. The device comprises a depth sensor, an image sensor, a sound collector, an obstacle detection component and an interaction feedback component. Whether the obstacles are available on roads at the fronts of the blind persons or the blind persons can pass the roads can be detected by the obstacle detection component according to information which is acquired by the various sensors, and control signals can be transmitted to the interaction feedback component according to detection results. The device has the advantages that the information of the various sensors is fused, and accordingly the obstacle detection robustness and identification precision can be improved.

The invention provides a barrier detection device based on pattern projection. The device comprises a pattern generator, an image sensor and a barrier detection component, wherein the pattern generator is used for projecting a pattern to a pavement; the image sensor acquires pavement image information with the pattern; the barrier detection component judges whether a barrier exists on the pavement by analyzing a change situation of the pattern in the acquired image. According to the device, a pattern projection mode is adopted during barrier detection, so that the complexity of barrier analysis and detection is greatly simplified, the cost is reduced, the robustness of barrier detection is improved, and the detection speed and the precision are improved.

The invention provides a forest fire disaster detection method based on videos. Based on unique growth change features of forest fire disaster smoke, by judging growth change conditions in a candidate smoke region, forest fire disaster smoke detection is achieved, so it is not easy for objects in other colors and similar shapes to interfere with the smoke detection; effects imposed by other types of moving objects can be effectively eliminated; the method is quite high in detection robustness; in the provided detection method based on early videos of the forest fire disasters, accumulation regions are applied and comparison is performed for many times in a certain frame number interval based on the accumulation regions, so judgment and analysis for growth change of the regions are quite stable; and when a forest fire disaster happens early, through detection of the smoke, problems of automatic detection and alarming of the fire disaster based on a remote-distance monitoring video are effectively solved.

The invention relates to a mine belt anomaly detection method and device based on video processing. A high-definition camera is installed at a coal drop port between the belt conveyors. The camera shoots a picture directly opposite the coal drop port, and the video information of the coal drop port is collected in real time, and the processing process is visualized. The method comprises the following steps of: preprocessing a frame image extracted from a monitoring video of a coal nozzle, wherein the preprocessing comprises image grayscale transformation and filtering; Calculating the ratio ofthe number of pixels nc of coal particles in the preprocessed image to the total number of pixels n of the delimited detection area to obtain the relative area lambda s; When the relative area lambdas is greater than the first threshold value Ts1, determining that there is a large coal block or a coal piling condition in the image; the first threshold value Ts1 has a range of 0<Ts1<1. The invention can intelligently detect the possible large coal lump or coal piling at the coal nozzle, and the method can enhance the detection robustness, reduce the calculation amount, reduce the calculationerror, and speed up the operation.

The invention belongs to the technical field of image processing, and particularly discloses a checkerboard corner detection method based on a neural network and application of the checkerboard cornerdetection method. The method comprises the following steps: acquiring a multi-angle image of a checkerboard to be detected, and preprocessing the multi-angle image to obtain a grayscale image; basedon a neural network, carrying out angular point detection on the gray level image so as to determine candidate angular points; removing the candidate angular points which do not meet the geometric standard of the checkerboard to determine the angular points to be detected; determining sub-pixel positions of the checkerboard corner points to be detected according to the obtained corner points to bedetected; and performing region screening on the obtained to-be-detected corner points, determining an index matrix of each to-be-detected corner point, and further obtaining a plurality of checkerboard corner point sequences which are not crossed with one another, thereby finishing corner point detection of the to-be-detected checkerboard. The method has high detection robustness for interference such as illumination shielding and noise offset, has high positioning stability for image bending and distortion caused by uneven lenses or wide-angle camera shooting, and can complete detection formultiple chessboards at the same time.

The invention discloses an image boundary detection method and a device, relates to the technical field of image processing, and aims to solve the problem that palm boundaries cannot be accurately detected in the prior art. The method mainly comprises the steps of performing equalization processing on a to-be-detected image according to a preset reference color to generate an equalized image; performing binarization processing on the averaged image to generate a boundary image; extracting an outer contour image of an object in the boundary image according to an edge detection algorithm; judging whether the outer contour image is a target image according to a support vector machine SVM algorithm; and if the judgment result is yes, marking the outer contour pixel points of the outer contourimage in the to-be-detected image. The method is mainly applied to the image recognition process.

The invention discloses a face fusion detection method based on a color texture dual-channel convolutional neural network and a recurrent neural network, and the method comprises the following steps:1, carrying out the face region extraction of an input image, carrying out the resampling of a face region, enabling the face region to be in a fixed size, and enabling the face region to be divided into non-overlapped image blocks; 2, preprocessing each image block to extract a color component and a texture component; 3, inputting the color and texture components extracted from each image block into the trained dual-channel convolutional neural network to obtain high-dimensional feature expression; 4, taking the high-dimensional feature expressions of all the image blocks obtained in the step3 as the input of the spatial recurrent neural network to obtain the output score of the network; comparing the output score with a preset threshold value, and judging whether the input image is a face fusion image or not. According to the method, the detection performance under the conditions that the number of training samples is limited and the image acquisition environment is complex is effectively improved, and the robustness of malicious post-processing operation is enhanced.

The invention discloses a contact point real-time detection algorithm based on a monocular infrared image. The method comprises the steps of using a bow template intercepted from a first original image of an infrared video frame sequence for carrying out template matching on a monocular infrared original image obtained in real time, and determining the position of a pantograph; according to the position of the pantograph, respectively extracting an image of a contact line above the pantograph and an image of an area of interest of the pantograph from the monocular infrared original image; respectively performing image enhancement operation on the contact line image above the bow and the image of interest of the pantograph to obtain a contact line enhanced image and a pantograph enhanced image; performing pixel normalization operation and parameter optimization on the contact line enhanced image and the pantograph enhanced image respectively, and extracting a contact line point set anda pantograph carbon slide plate contour point set; and respectively carrying out linear parameter estimation on the contact line point set and the pantograph carbon slide plate contour point set, andcalculating an intersection point of a contact line fitting straight line and a pantograph slide plate fitting straight line to realize the contact point positioning.

The invention discloses a pedestrian detection method, system and terminal for multi-modal information fusion in an image. The method comprises the steps of obtaining a corresponding top view according to an original head-up view; predicting a pedestrian target probability and a pedestrian target frame width for each point on the top view; obtaining a corresponding depth map through the head-up image, and estimating the height of a pedestrian target frame based on the depth; integrating the pedestrian target frame width and the pedestrian target frame height into a target candidate frame, anduniformly encoding the target candidate frame to a semantic space through an encoder to obtain a pedestrian target probability under a head-up view; and performing weighted average on the pedestrian target probabilities of the head-up view and the top view to obtain a final pedestrian target prediction result. According to the method, the pedestrian shielding problem under a head-up view angle iseffectively avoided, complex background interference is effectively removed, the method is suitable for variable situations in a real application environment, the detection robustness is enhanced, thefalse detection and missing detection probability is reduced, and the detection capability of pedestrian targets in video images is effectively improved.

The invention relates to an optical remote sensing image marine ship target extraction and identification method. The method comprises the following steps: inputting a visible light remote sensing image; performing global detection on the sea surface in the remote sensing image by using a visual saliency method based on covariance joint features to generate a single-scale saliency map; performing downlink sampling on the generated single-scale saliency map, establishing a multi-scale saliency map, and obtaining a final saliency map; calculating a detection threshold value of a sea surface ship target, binarizing the saliency map, and realizing coarse segmentation of the saliency map; establishing a training set and a test set; the feature vectors can be used for training and testing; training a model according to the established framework and positive and negative samples in the training set, and classifying the feature vectors; and according to the established framework and positive and negative samples in the test set, testing the candidate area, and completing the extraction and identification of the optical remote sensing image marine ship target. The sea surface target area can be efficiently searched, the false alarm rate is greatly reduced, and the detection accuracy is improved.

The invention discloses an article taking and placing detection system, method and device. The article pick-and-place detection system comprises an article pick-and-place cabinet, an image acquisitionunit and a pick-and-place detection unit, markers are arranged on the edge of an entrance of the article taking and placing cabinet. The image collecting unit is used for collecting images of the entrance of the article taking and placing cabinet. And the pick-and-place detection unit is connected with the image acquisition unit and is used for carrying out article pick-and-place detection basedon the marker information in the image acquired by the image acquisition unit. The marker is arranged at the edge of the entrance and exit of the article taking and placing cabinet, the marker information in the entrance and exit image of the article taking and placing cabinet is detected, article taking and placing detection is carried out based on the detection result, target article detection and tracking are not needed, the requirements for scenes and target articles are reduced, and frequent changing scenes and taking and placing at any speed can be coped with; detection robustness is high. In addition, pick-and-place detection is carried out through the marker information, so that the authenticity and the accuracy of the detection pick-and-place operation are improved.

The invention discloses a warehouse cargo multi-QR code integrated visual positioning method, and relates to the field of image detection, in particular to a multi-QR code integrated visual positioning method. The technical problem that an existing detection method cannot accurately position the QR code because of complex environment where the QR code is located and multi-angle shooting is solved.The method comprises the following steps: filtering an image; converting the QR code into a grey-scale map, binarizing the grey-scale map, roughly locating the position of a position detection graphaccording to closed contour information in the binary image, finely locating the position detection graph according to the characteristic that the areas of three position detection graphs in the sameQR code presented in the image are similar, and finally locating the QR code. According to the method, a plurality of two-dimensional codes in a scene can be accurately positioned from a plurality ofshooting angles in a complex environment, the robustness is relatively high, and the method has very good performance in actual use.

The invention provides an aluminum profile surface defect detecting method based on machine vision. The method comprises the steps of dividing an aluminum profile area, detecting aluminum profile surface defects and uploading defect information and pictures, wherein the divided aluminum profile area comprises a grained area and a non-grained area. The method realizes real-time detection of the aluminum profile surface defects based on machine vision and computer technology, and can be used for positioning the defect part more conveniently, quickly and accurately compared with existing artificial defect detection, so that the detection efficiency is improved, human input is greatly reduced, and the cost of an enterprise in the production process is reduced; and the method can avoid interference of human subjective factors and enterprise reputation can be maintained.

The invention relates to a cloth defect detection method based on Fourier transform and image morphology, comprising the following steps: (1) collecting the original image of the cloth with a line array camera; (2) preprocessing the original image; (3) Using LAWS texture filtering to extract image texture features; (4) Using GMM classifier model for defect pre-discrimination; (5) Using Gaussian filter to construct a band-stop filter; (6) Using Fourier transform and inverse transform to generate defect images; (7) Use image morphology to extract the defect position and area; (8) Output the cloth detection results. The present invention can detect more than ten types of cloth defects in real time, including broken weft, broken warp, hole, floating weft, etc., with fast detection speed and high accuracy, and the image processing method integrating air domain and frequency domain enhances the robustness of detection .

The invention relates to a behavior recognition method based on wearable equipment and a human skeleton. The method includes the following steps: (1) acquiring an overall behavior state of a human body based on wearable equipment; (2) acquiring human skeleton information based on a depth camera; (3) extracting spatial features of the human skeleton through a graph convolutional network; (4) extracting time sequence information of the skeleton sequence by using a long-short-term memory network; (5) fusing the space information extracted by the GCNN and the time sequence information extracted by the LSTM; and (6) carrying out classification and identification through an SVM classifier, and identifying specific behaviors of the human body. According to the invention, the wearable device is utilized to obtain the overall position and motion condition of the human body, the deep learning network model is utilized to identify and classify the human skeleton information, and thereby the human behaviors are accurately identified.

The invention provides a personnel leaving detection system. The personnel leaving detection system comprises an ultra wide band main node and at least one ultra wide band auxiliary node which are arranged in a vehicle, wherein each ultra wide band auxiliary node is used for collecting channel impulse response (CIR) information generated by communication with the ultra wide band main node, and sending the CIR information to the ultra wide band main node; the ultra wide band main node is used for detecting whether a vehicle is flamed out and locked or not: if yes, initiating periodic ultra wide band broadcasting to establish communication with all the ultra wide band auxiliary nodes, and after CIR information sent by all the ultra wide band auxiliary nodes in the current broadcasting period is received, operating a personnel leaving detection algorithm to judge whether the personnel are left or not according to all the received CIR information. Due to the fact that the ultra-wideband signals are not easily affected by human body and environment shielding, and the ultra-wideband universal chip has a CIR information analysis function and does not need chip firmware refreshing, hardware cost is low, detection robustness is high, in addition, biological characteristic data collection does not need to be carried out, and user privacy is also protected.

The invention discloses a hopper real-time positioning method for semitrailer material filling. The method comprises the following steps: preprocessing a current hopper video frame image to obtain a semitrailer grayscale image; processing the semitrailer grayscale image to obtain halo image data; carrying out vehicle hopper outer corner feature extraction on the halo image data to obtain vehicle contour image data; performing gaussian blur on the halo image data, and robustness optimal modeling is performed by using the vehicle contour image data; performing auto-quotient graph processing by using the vehicle contour image data and the robustness optimal modeling result; coordinate regression is conducted on the autoquotient graph processing result, and car hopper corner point coordinatesare obtained; and taking the next frame of the video frame image sequence to execute the processing of the steps 1-6, and repeating the steps to obtain the car hopper corner point coordinates of the video frame image sequence, thereby realizing the real-time positioning of the car hopper. By utilizing the method and the device, the workload of manual parameter debugging can be reduced, and the detection robustness under the illumination change condition is improved.

The invention provides a chromosome stability evaluation method and device based on liquid biopsy, terminal equipment and a storage medium, and the chromosome stability evaluation method comprises the following steps: respectively carrying out capture sequencing on a to-be-detected plasma sample and a selected reference plasma sample, and carrying out preprocessing operation to obtain a Bam file; according to the Bam file, performing statistics on reads numbers of bin levels of the to-be-detected plasma sample and the reference plasma sample; scoring the chromosome stability of the to-be-detected plasma sample according to the reads number of the bin levels; and evaluating the chromosome stability of the to-be-tested plasma sample according to the score. According to the method, the score of the chromosome stability is obtained according to the reads number of the bin levels, and then the chromosome stability of the to-be-detected plasma sample is evaluated according to the score, so that the detection robustness and sensitivity are improved.

The embodiment of the invention discloses a mycobacterium tuberculosis detection model training method and device, equipment and a storage medium. The method comprises the following steps: acquiring a first preset number of first images and annotation information of each first image, namely the position of mycobacterium tuberculosis in the first image; then, cutting out a second preset number of second images with different sizes from each first image, namely taking part of the first images as the second images, and zooming all the second images containing mycobacterium tuberculosis into third images with the same size to simulate the shooting effects of different shooting devices; then, constructing a sample data set for training the network by using all the third images; and finally, training a preset target detection network according to the sample data set to obtain a mycobacterium tuberculosis detection model. Therefore, according to the embodiment of the invention, the sample input into the preset target detection network can simulate different shooting equipment to shoot the image containing the mycobacterium tuberculosis, so that the detection robustness of the preset target detection network is improved.

The invention provides a device for prompting blind persons to avoid obstacles. The device comprises a depth sensor, an image sensor, a sound collector, an obstacle detection component and an interaction feedback component. Whether the obstacles are available on roads at the fronts of the blind persons or the blind persons can pass the roads can be detected by the obstacle detection component according to information which is acquired by the various sensors, and control signals can be transmitted to the interaction feedback component according to detection results. The device has the advantages that the information of the various sensors is fused, and accordingly the obstacle detection robustness and identification precision can be improved.

The invention relates to the field of signal detection, and particularly relates to a target detection method based on symmetrical spectral characteristics and sub-symmetrical characteristics and a target detection device thereof and a system. The target detection method comprises the steps that a set of echo data are received and multiple echo data of the echo data are selected to act as master data, and multiple echo data of the set of echo data apart from the master data are selected to act as auxiliary data, wherein the master data and the auxiliary data have the same reverberation covariance matrix, and the covariance matrix has the predetermined sub-symmetrical characteristics; the master data and the auxiliary data are transformed based on the sub-symmetrical characteristics and the symmetrical spectral characteristics of the reverberation covariance matrix; the detection statistics is calculated according to the transformation result of the master data, the maximum likelihood estimation of the reverberation covariance matrix and the target nominal guide vector; and the detection statistics is compared with the predefined threshold, and existence of the target is judged according to the comparison result. The space-time detection robustness under the background of partial uniform reverberation can be effectively enhanced by using the symmetrical characteristics of the power spectrum and the sub-symmetrical characteristics of display.

The present invention provides a distance measuring method based on the monocular vision of the front vehicle driven by the plane geometric model of the lane, utilizes the CCD camera to collect the front target vehicle image, and utilizes the fusion Haar-like feature and Adaboost algorithm to identify the front target vehicle obtained in step 2 For the target vehicle ahead in the image, use the particle filter method to track the target vehicle ahead obtained in step 3, and construct the longitudinal distance measurement model of the lane plane geometry in the frame image according to the target vehicle ahead in each frame of image obtained above , to obtain the longitudinal perception distance y of the target source in the frame image, construct a vehicle ranging error dynamic compensation model according to the target vehicle ahead in each frame image obtained above, and obtain the longitudinal measurement error value z, according to the target target obtained in step 5 The source longitudinal perception distance y and the longitudinal measurement error value z calculate the longitudinal distance Y between the front target vehicle and the own vehicle in the frame image W (P).

The invention provides a multi access detection method used for Doppler tolerance, comprising the steps as follows: 1) at the emission end, a driving waveform used for detection is formed and emitted out; the driving waveform cuts the wideband hyperbolic frequency modulation waveforms into N sub-pulses; subsequently, the waveform is composed according to HFM-Costas formed by the resorting of N-order Costas coding sequence; 2) at the receiving end, the echo receiving detection of the driving waveform is carried out. The method of the invention has the technical effects that: compared with the driving waveform compared with the traditional multi access detection method, the HFM-Costas composite waveform in the novel method has better detection robustness on the Doppler tolerance; aiming at the interference-non-interference mixed detection receiving machine of the composite waveform, the Doppler passage searching is not needed, the structure is simple and the hardware cost is low.

Disclosed is a palm vein recognition system. Candidate regions are detected through MSER, key points are positioned and scored through edge images, ROI is extracted, and authenticity of the palm ROI is judged through an HOG + SVM classifier; during feature extraction, curvature filtering, Gamma filtering and contrast-limited histogram equalization are firstly adopted for preprocessing, and then Gabor wavelet coefficient features, LGBP features and index features are extracted; during comparison, the index features are firstly used for comparison, the first 5% of the registration library is selected as a candidate set, then sampling points, directions and frequencies are selected for comparison, 20% of the candidate set is selected again, and finally the registration features with the highest similarity are selected through one-to-one comparison.

The present invention provides a forest fire detection method based on video. Based on the unique growth and change characteristics of forest fire smoke, forest fire smoke detection is realized by discriminating and analyzing the growth and change of alternative smoke areas, so that smoke detection is not easily affected by other methods. The interference of objects with similar colors and shapes can also effectively eliminate the influence of other forms of moving objects, and has high detection robustness; at the same time, the forest fire early video detection method of the present invention uses the accumulation area, and within a certain frame number interval Multiple comparisons based on the cumulative area make the discriminant analysis of regional growth changes more stable; the technology of the present invention can effectively solve the automatic fire detection of remote monitoring video through the detection of smoke when forest fires occur in the early stage and alarm issues.

The invention discloses a complex injection molding product flash defect identification method based on multi-template low-rank decomposition. The method comprises collecting the defective injection molding product sample images, extracting foreground gray level images, obtaining template images from the foreground gray level images by adopting a method of decomposing a low-rank matrix, extractinghole areas of the template images and the sample images on the basis of contours, and matching the templates and the hole areas of the samples on the basis of position similarity; constructing hole area angle feature descriptors corresponding to the template and the sample image, and calculating similarity to determine whether a flash defect exists or not. According to the present invention, a defect-free product does not need to be obtained in advance, the method is not influenced by tiny deviation during shooting, the influence of image brightness nonuniformity, noise and the like is small,and the detection robustness is high.

The present invention provides a liquid biopsy-based chromosome stability evaluation method and device, a terminal device and a storage medium, wherein the chromosome stability evaluation method includes: respectively capturing and sequencing a plasma sample to be tested and a selected reference plasma sample, and then The Bam file is obtained by preprocessing; the number of reads at the bin level of the plasma sample to be tested and the reference plasma sample are counted according to the Bam file; the chromosome stability of the plasma sample to be tested is scored according to the number of reads at the bin level; the plasma to be tested is scored according to the score. Chromosomal stability of the samples was evaluated. It obtains a chromosomal stability score based on the number of reads at the bin level, and then evaluates the chromosomal stability of the plasma sample to be tested according to the score, improving the robustness and sensitivity of the detection.