32results about How to "Meet real-time detection requirements" patented technology

The invention discloses a Newton's theorem-based method for detecting ellipses in shot object images. According to the method of the invention, in a pre-processing stage, invalid edge information is removed according to arc dimension constraints; edge gradient information is extracted, the slope of each edge point is calculated, and edges are divided into four categories according to the slope information; arcs which satisfy a condition of constituting an ellipse are aggregated according to the relative positional relationships of endpoints and Newton's theorem-based geometric constraints; theleast squares method is directly adopted to estimate the parameters of arcs belonging to the same set; and the validity of fitted ellipses is verified, ellipses that do not satisfy a verification condition are removed, and ellipses requiring repeated detection are aggregated. Since a large number of invalid edges are removed, dimensions, relative positions and Newton's theorem-based geometric constraints are adopted, the precision and speed of ellipse detection are significantly improved.

The invention discloses an artificial intelligent image recognizing based detection system for a printed circuit board. The system comprises a housing, a conveyor belt, an air pump, a motor, a telescoping rod, a pushing plate, a detector, a receiver, a sensor and a controller, wherein a feeding port and a discharging port are correspondingly formed in the front surface and the rear surface of thehousing; a motor and a defective product outlet are correspondingly arranged at the left end and the right end of the housing; the air pump is fixed to the top part of the housing; the controller is fixed to the outer surface of the housing. With the adoption of the system, general defects of a bare printed circuit board can be quickly and accurately positioned, including short circuit, open circuit, hole, excessive copper and scratch; the requirement on real-time detection in the production process is met; the system is reasonable in structure, convenient to operate, and high in automation degree; an artificial intelligent method replaces a traditional algorithm to realize automatic detection; the system is of a noncontact type, and is high in detection accuracy, fast, high in interference resistance, and convenient to connect to an ERP quality detection module in a communication manner.

The present invention discloses an automatic ship detection method for inland river bridge areas, the method comprises: 1) collecting traffic surveillance videos of real inland river bridge areas andobtaining original sampled image data; 2) performing manual label to the original sampled image data and forming a label image; 3) using depth convolution neural network to learn characteristics of the label image, classifying the characteristics learned, and obtaining a trained deep learning model; 4) using the trained deep convolution neural network model to detect ships in real scene, framing out position of the ships with rectangular frames, and indicating classification of the ships. The method has good adaptability to different weather, different lighting conditions and nighttime dark conditions, and can achieve relatively high accuracy, and has processing speed reaching real-time requirements.

The invention provides a vehicle detection method based on multi-feature fusion. The vehicle detection method based on the multi-feature fusion comprises the steps of step 1, establishing one or a plurality of monitoring areas (ROI (region of interest)) in an input image; step 2, extracting Harris angular points and longitudinal gradient information in the ROI on the basis of a grayscale map of an image to be detected; step 3, computing a geometry size of a detection window at a certain position inside the ROI according to a circumscribed trapezoid of an obtained monitoring area; step 4, counting an angular point number of the detection window and computing the distribution symmetry of the angular points; step 5, extracting structured horizontal line information on the basis of a longitudinal gradient in the detection window; step 6, utilizing a cascaded strong classifier to judge whether a current detection window is a vehicle area, if so, reserving the current detection window, and if not, dismissing the current detection window; step 7, moving the detection widow according to a certain step size, switching to the step 3, and repeating the steps from the step 3 to the step 7; and step 8, reducing the probability of the repeated detection of the same target through detection result combination.

The invention relates to an artificial mark detection method applied to augmented reality. The method comprises the specific steps of S1: acquiring a frame image, roughly sampling the frame image, performing slant grid scanning, and performing detection to obtain edge pixels of the frame image; S2: based on an RANSAC algorithm, performing detection to obtain edge line segments in the frame image; S3: fusing the edge line segments; S4: extending and screening the edge line segments; and S5: detecting edge corner points of a quadrangle and constructing the quadrangle according to the edge corner points of the quadrangle. According to the method, the frame image is preprocessed before calculation, rough grid sampling is performed, and each grid region is subjected to edge detection, so that the program operation time is greatly shortened, the detection speed is increased, the real-time property is good, and the real-time detection requirement is met; and by adopting an edge-based detection method, line segment testing is performed at first and then a quadrilateral frame of a mark is reconstructed according to the line segments obtained by the line segment testing, so that the method has very good robustness for the illumination change and the occlusion condition.

The invention relates to a television station logo detection and recognition method, and belongs to the technical field of image detection and recognition. The method comprises the following steps: (1) collecting station logo picture samples, carrying out labeling on corresponding station logo positions of pictures, and carrying out size normalization processing on all the images of object detection; (2) setting parameters of a learning rate, an iteration number and the like of TB-Net network training; (3) sending the normalized images into a TB-Net network for training, extracting primary features and abstract features of the images, sending the same into a position regression layer and a classification layer, and obtaining a best station logo detection model after training; and (4) loading the station logo detection model and network by a test program, sending a to-be-detected picture into the network, and outputting a result of whether the image contains a station logo and a position where the station logo is located. According to the method, station logos in video pictures can be more accurately detected, and accuracy of station logo recognition can be improved.

The invention provides a product appearance defect detection method based on fuzzy relaxation constraint multi-core learning, which belongs to the product quality detection field of machine vision. Firstly, some features of the real-time image are extracted. Then the fuzzy constraint theory is used to quantitatively analyze the mapping relationship between the characteristics and the evaluation indexes. A multi-core learning model is established to classify the appearance defects, and the fuzzy relaxation boundary of each kernel function weight is delineated by combining the mapping quantization relation. The fuzzy relaxation constraint (FRC) method is used to determine the weights of the multi-kernel model and the fuzzy range of the weights. At last, the multi-core learning model with different kinds of defects is obtained by calculating the weights of defects, and the defect detection results are obtained by using the multi-core learning model. The invention adopts the multi-featurefusion multi-core learning classification detection method, which makes the detection range wider, combines the fuzzy relaxation constraint to adapt to different detection requirements, can meet the real-time detection, and has high detection accuracy.

The embodiment of the invention discloses a machine learning method and device, a privacy code determination method and device and electronic equipment, the machine learning method includes: obtainingsample data containing positive and negative samples in batch, the positive sample data contains a privacy code file, and the negative sample data does not contain a privacy code file; screening outa plurality of first code files from the sample data based on similarity measurement parameters of the code files, and screening out a plurality of second code files from a preset code library with known privacy tags of the code files; determining a target parameter corresponding to the sample data based on the plurality of first code files and the plurality of second code files; and training a target model by taking the target parameters corresponding to the sample data and the labels of the sample data as input, the target model being used for determining similarity measurement parameters ofa batch of code files to be detected and private code files.

The invention discloses a night infrared pedestrian detection method and system based on improved YOLOv3, and the method comprises the steps: collecting a plurality of infrared pedestrian detection data sets through a plurality of infrared cameras, carrying out the pixel value contrast enhancement of the data sets, increasing the pedestrian pixel values, and reducing the background pixel values; a night infrared pedestrian detection network model YOLOv3-SAB is improved and constructed based on YOLOv3, a stem down-sampling module and asymmetric convolution are introduced to improve the network feature extraction capability and feature expression capability, calculation parameters in a bottleneck residual error reduction model are introduced, and the pedestrian detection speed of the model is improved; a specific prior aiming frame is generated through clustering by using a mean value clustering algorithm, and the model target positioning precision is improved; cIoU is used as a YOLOv3-SAB network bounding box regression loss function, so that model convergence is accelerated, and the accuracy of a prediction box is improved; a YOLOv3-SAB network is trained to generate a night infrared pedestrian detection model; and performing real-time infrared pedestrian detection at night by using the night infrared pedestrian detection model. According to the invention, the detection precision and the detection speed of pedestrians at night are effectively improved.

The invention discloses a resin lens defect detection method based on a convolutional neural network. The method comprises the following steps: acquiring an image of a resin lens with defects; carrying out gradient sharpening and median filtering on the image to enhance detail information of the image; marking the position and category of the defect, and making a data set; constructing a convolutional neural network model, wherein the network model is based on an improved DarkNet53 as a backbone network, a multi-scale feature fused neck network and a multi-classification regression network as a detection head; inputting the data set into the network model for training; and inputting a to-be-detected picture into the trained network model for detection, and finally outputting information such as a defect category, a defect position and confidence, namely a final detection result. According to the invention, online detection of the defects of the resin lens can be realized, the generalization performance is good, and the detection efficiency and the detection precision are improved.

The invention discloses an infrared video moving small target real-time detection method based on space-time tensor decomposition, and belongs to the field of video processing and target detection. Each input video frame image is partitioned, the partitioning results of several adjacent frames of images are fully utilized, the three-dimensional matrix tensor is constructed, only the memory space of the key tensor in one three-dimensional matrix tensor is reserved, the memory allocation and release processes are omitted, each frame of target detection result picture is deleted, and memory management is optimized. The video frame required for constructing the space-time image block tensor for the first time is directly partitioned according to the size of the image block, so that the situation that the image blocks with overlapped information are merged into the process of constructing the three-dimensional matrix tensor is avoided, and the initialization process of constructing the space-time tensor is further optimized. The two-dimensional tensor of the target image is obtained through tensor decomposition. And according to the two-dimensional tensor of the target image obtained through tensor decomposition, the infrared small target is detected through a threshold segmentation method, namely, the real-time detection of the infrared video moving small target is achieved based on space-time tensor decomposition.

The invention discloses a method for judging DC series arc fault based on a frequency domain sage specific value, and belongs to the technical field of the DC arc fault detection. The method comprisesthe following steps: computing amplitude sum of each of 12 frequency bands in a DC power supply loop current spectrum, forming six groups of specific value relations as characteristic values Rarc1-Rarc4, S_Rarc1, S_Rarc2, and finding out the maximum amplitude in 51 points from the 75th to the 125th in the loop current spectrum, and computing the specific value of the 50 times of the maximum valueto other 50 points, and serving the specific value as the characteristic value E. when a threshold condition is satisfied twice or more in quartic data collection cycles of seven groups of characteristic values, the DC series arc fault in the DC power supply system can be confirmed. The method disclosed by the invention can be applied to the DC series arc fault detection in the DC power supply system.

The invention discloses a quick morning and night fog detection method, device and equipment based on terrain constraint and deep learning and a medium. The method comprises the following steps: step 1, acquiring DEM data of a research area and H8/AHI data at morning and night moments; step 2, preprocessing the data obtained in the step 1: carrying out wave band cutting and useless wave band elimination on the H8/AHI data, and then combining the reserved wave band data and DEM data according to channels to obtain fused AHI-DEM data; and step 3, inputting the AHI-DEM data into a pre-trained morning fog detection model based on deep learning, and outputting to obtain a morning fog detection result of the research area. The terrestrial fog detection method can accurately, efficiently and quickly realize terrestrial fog detection at morning and evening.

The invention relates to a battery defect detection method based on a lightweight neural network. The method comprises the following steps: constructing a defect detection model based on the lightweight neural network, wherein the defect detection model takes ShuffleNetV2 as a backbone network, a refined cross-stage local mechanism is fused into a ShuffleNetV2 network, a refined cross-stage local ShuffleNetV2 network is obtained, and a parallel grouping attention module is further fused into the refined cross-stage local ShuffleNetV2 network; and refining an output feature map of the cross-stage local ShuffleNetV2 network, inputting the output feature map into a recommended region in a region recommendation network after the output feature map passes through a fusion module guided by low-layer features, and classifying and regressing the recommended region to obtain a defect category and a defect position. A parallel grouping attention module realizes shallow-layer and deep-layer feature fusion, and a fusion module guided by the low-layer features expands the receptive field; and the method solves the problem that the lightweight network detection effect is not ideal.