183 results about "Non maximum suppression" patented technology

The subject disclosure presents systems and computer-implemented methods for automatic immune cell detection that is of assistance in clinical immune profile studies. The automatic immune cell detection method involves retrieving a plurality of image channels from a multi-channel image such as an RGB image or biologically meaningful unmixed image. A cell detector is trained to identify the immune cells by a convolutional neural network in one or multiple image channels. Further, the automatic immune cell detection algorithm involves utilizing a non-maximum suppression algorithm to obtain the immune cell coordinates from a probability map of immune cell presence possibility generated from the convolutional neural network classifier.

The invention discloses a pedestrian and vehicle detection method and system based on improved YOLOv3. According to the method, an improved YOLOv3 network based on Darknet-33 is adopted as a main network to extract features; the cross-layer fusion and reuse of multi-scale features in the backbone network are carried out by adopting a transmittable feature map scale reduction method; and then a feature pyramid network is constructed by adopting a scale amplification method. In the training stage, a K-means clustering method is used for clustering the training set, and the cross-to-parallel ratio of a prediction frame to a real frame is used as a similarity standard to select a priori frame; and then the BBox regression and the multi-label classification are performed according to the loss function. And in the detection stage, for all the detection frames, a non-maximum suppression method is adopted to remove redundant detection frames according to confidence scores and IOU values, and an optimal target object is predicted. According to the method, a feature extraction network Darknet-33 of feature map scale reduction fusion is adopted, a feature pyramid is constructed through feature map scale amplification migration fusion, and a priori frame is selected through clustering, so that the speed and precision of the pedestrian and vehicle detection can be improved.

The invention discloses a pedestrian detection method based on an end-to-end convolutional neural network in order to solve the problem that the existing pedestrian detection algorithm has the disadvantages of low detection precision, complex algorithm and difficult multi-module fusion. A novel end-to-end convolutional neural network is adopted, a training sample set with marks is constructed, and end-to-end training is performed to get a convolutional neural network model capable of predicting a pedestrian candidate box and the confidence of the corresponding box. During test, a test picture is input into a trained model, and a corresponding pedestrian detection box and the confidence thereof can be obtained. Finally, non-maximum suppression and threshold screening are performed to get an optimal pedestrian area. The invention has two advantages compared with previous inventions. First, through end-to-end training and testing, the whole model is very easy to train and test. Second, pedestrian scale and proportion problems are solved by constructing a candidate box regression network, the pyramid technology adopted in previous inventions is not needed, and a lot of computing resources are saved.

The invention relates to a pedestrian recognition and tracking method based on an accelerated area Convolutional Neural Network. Firstly, training and testing data set are preprocessed according to the requirements through a robot with an infrared camera to acquire a training dataset and a testing dataset at night, and then, actual target position labeling is conducted on all training and testing photos and is recorded to a sample file; then, the accelerated area Convolutional Neural Network is constructed, the accelerated area Convolutional Neural Network is trained by using the training dataset, and the final probability belonging to a pedestrian area and a bounding box of the area are calculated out from network output by the usage of a non-maximum suppression algorithm; the accuracy of the network is tested by the usage of the testing dataset, and a network model consistent with the requirements is obtained; photos collected by the robot at night are input to an accelerated area Convolutional Neural Network model, and the probability belonging to the pedestrian area and the bounding box of the area are online output by a model in real time. According to the pedestrian detection and tracking method based on the accelerated area Convolutional Neural Network, a pedestrian in an infrared image can be effectively recognized, and real-time tracking for a pedestrian target in an infrared video can be achieved.

The invention relates to an aluminum material surface defect detection algorithm based on deep learning, and the algorithm comprises the steps: (1) employing a camera to shoot the surface of an aluminum material, obtaining a related data set, employing a labelImg tool to label an image, and obtaining label information; (2) dividing the image into a training set and a test set, and performing dataenhancement on the training set; (3) inputting a defective image, a non-defective image and label information of the defective image into the network at the same time every time to carry out model training; and (4) inputting the test image into the trained aluminum material surface defect detection model, and obtaining the position and the corresponding category of the defect. According to the method, a defective image and a non-defective image can be effectively utilized; the generalization ability and the detection precision of the model are improved, the detection performance is further improved by fully utilizing context information around the candidate region, the detection performance of dense small defects can be improved by utilizing a soft non-maximum suppression algorithm, and the method is an efficient aluminum material surface defect detection algorithm.

The invention belongs to the technical field of defect detection, and discloses a defect target detection method based on an attention mechanism, which comprises the following steps: marking various defects of all pictures in an original data set to obtain a standard training data set with marks; Obtaining a training label according to the standard training data set, determining a loss function, obtaining a network model, and performing training by using a reverse conduction method to obtain a defect regression detection network model based on an attention mechanism and having enhanced defectpart weight; Performing classification prediction and regression prediction on the to-be-detected picture by utilizing the defect regression detection network model; Carrying out non-maximum suppression processing on the predicted defect bounding box and filtering the defect bounding box to obtain an output result; According to the method provided by the invention, the weight of the defect area isimproved through an attention mechanism, so that the defect detection precision is improved; The industrial product surface defect classification and regression detection method can be applied to other types of surface defect detection frameworks to improve the detection precision, and is high in universality.

The invention discloses a generative adversarial network improved CT medical image pulmonary nodule detection method. The method includes: 1), acquiring a section of a pulmonary CT image; 2), separating according to image morphological properties to acquire a ROI pulmonary parenchyma area; 3), acquiring different suspected pulmonary nodule candidate sets according to a connected domain formed by abinarized image; 4), building a model of an assistant classifier generative adversarial network to generate positive samples overcome the circumstance that positive-negative sample number is unbalanced; 5), building a convolution neural network to classify suspected pulmonary nodule parts to acquire pulmonary nodule areas; 6), using a non-maximum suppression algorithm to acquire a final area of pulmonary nodule. By the method, efficient processing performance of a computer can be fully utilized, certain expandability is provided, and data processing efficiency is improved; through a convolution neural network algorithm, classifying accuracy is improved, CT image data processing performance is improved, and pulmonary nodule images can be built and analyzed more efficiently.

The invention discloses a non-segmented character positioning and identification method based on deep learning. The non-segmented character positioning and identification method comprises steps of constructing a deep convolution neural network, wherein the deep convolution neural network comprises a universal convolution network, a candidate positioning network and a classification identification network, constructing a target function in order to realize global end-to-end training of a whole network, adopting an artificially-calibrated training set and a progressive-combined training mode to perform training on the network, using a network obtained through training to extract possible areas of a plurality of characters in a test image and a classified identification result when applied to a test, and performing non-maximum suppression on a result obtained by the network and performing post-processing on score threshold determination to obtain a final detection result. The non-segmented character positioning and identification method is simple, does not need to perform character segmentation pre-processing, can be compatible with a plurality of character forms, has a strong capability of resisting background interference and can be used as a general character detection method.

The invention discloses a steel rail tread defect recognition method based on combination of a gray image and a depth image. The method comprises the steps of establishing a steel rail tread image data set by utilizing a registered gray image and depth image pair of a steel rail tread, dividing the data set into a training sample set and a test sample set, and preprocessing the images of the dataset; performing defect recognition by adopting a special convolutional neural network structure, wherein a front end of the network is provided with two branch structures, which can extract features from the gray image and the depth image of the steel rail tread respectively; fusing feature information of the gray image and the depth image through feature graph connection, and then outputting preliminary prediction results by adopting a prediction module; and finally, screening the preliminary defect prediction results by adopting a non-maximum suppression method to obtain a final steel rail defect recognition result. The method combines two-dimensional and three-dimensional features of the steel rail tread at the same time, has the capability of distinguishing real defects and pseudo-defects, can reduce the misjudgment rate and the missing detection rate, and is especially suitable for defect recognition in a complex environment.

Systems and methods are provided for detecting edges in an image. In an example method, the image is smoothed using a filter. A gradient magnitude and angle are determined for each pixel of the image. A non-maximum suppression is performed on the resulting image data. A double threshold with an upper and a lower threshold are applied to the resulting image data to determine the edges in the image. The upper and the lower thresholds are calculated automatically using a distribution of intensity values of the image. In example systems, an image is acquired by a camera, or other image acquisition units. An image processing unit is configured to detect the edges in the image using an adaptive threshold edge detection method.

The present invention discloses a scene text detection method based on an end-to-end full convolutional neural network, which is used for the problem of finding a multi-directional text position in animage of a natural scene. The method specifically comprises the following steps: obtaining a plurality of image data sets for training scene text detection, and defining an algorithm target; carryingout feature learning on the image by using a full convolution feature extraction network; predicting an affine transformation matrix in an instance level for each sample point on the feature map, andcarrying out feature expression on the text according to the predicted affine transformation deformation sampling grid; classifying feature vectors of a candidate text, and carrying out coordinate regression and affine transformation regression to jointly optimize the model; using the learning framework to detect the precise position of the text; and carrying out non-maximum suppression on the bounding box set output by the network to obtain a final text detection result. The method disclosed by the present invention is used for scene text detection of real image data, and has a better effectand robustness for multi-directional, multi-scale, multi-lingual, shape distortion and other complicated situations.

Described is a system for multi-object detection and recognition in cluttered scenes. The system receives an image patch containing multiple objects of interest as input. The system evaluates a likelihood of existence of an object of interest in each sub-window of a set of overlapping sub-windows. A confidence map having confidence values corresponding to the sub-windows is generated. A non-maxima suppression technique is applied to the confidence map to eliminate sub-windows having confidence values below a local maximum confidence value. A global maximum confidence value is determined for a sub-window corresponding to a location of an instance of an object of interest in the image patch. The sub-window corresponding to the location of the instance of the object of interest is removed from the confidence map. The system iterates until a predetermined stopping criteria is met. Finally, detection information related to multiple instances of the object of interest is output.

A method of extracting linear features from an image, the method including the steps of: (a) applying a non maximum suppression filter to the image for different angles of response to produce a series o filtered image responses; (b) combining the filtered image responses into a combined image having extracted linear features.

The invention discloses a method for performing multi-visual-angle face detection by means of integral channel features. The method is mainly characterized in that three LUV color channels in ten ACF channels are improved for obtaining a gray scale single channel, thereby forming an eight-channel characteristic and realizing quicker feature extraction; four-stage Adaboost cascaded classifier training is performed on the extracted features, thereby forming a cascaded strong classifier which comprises 4096 weak classifiers; and image detection is performed by means of the cascaded classifier and a quick feature pyramid method for quickly and accurately detecting faces. According to the method of the invention, detecting blocks are acquired by means of successive sliding of a sliding window on a characteristic pyramid according to a step length; the detecting blocks are classified by means of the trained Adaboost classifier; overlapped window elimination is performed on the detecting blocks which comprise the faces through a non-maximum suppression method; a final face detection window is kept and detection precision is improved.

The invention provides a linguistic text detecting and positioning system, and a linguistic text detecting and positioning method based on the system. The system includes a feature extraction network for extracting bottom-layer CNN features out of a to-be-detected image to obtain a plurality of different feature maps; at least three area suggestion networks for conducting the dichotomy and the boundary frame regression of the above different feature maps respectively and generating text candidate areas according to text-contained feature maps; a transition region for connecting a plurality of text candidate areas and generating an area convolution feature map according to the above text-contained feature maps and the text candidate areas; and a text detection network for generating the text area boundary frame bias information according to the above area convolution feature map, conducting the non-maximum suppression on the bias information and the filtering operation on non-reasonable areas, and generating the predicted text area boundary frame coordinate information in an image coordinate space.

The invention discloses a wheat field weed detection method based on deep learning, and the method comprises the steps: collecting a large number of wheat and wheat field main weed pictures at different growth stages, building a data set, and dividing the data set into a training set and a test set; inputting the training set into a preset convolutional neural network model for training through atransfer learning method to obtain a crop weed classification recognizer, and testing the crop weed classification recognizer by using a test set to obtain a classification recognition result so as toperform fine adjustment; generating a large number of interest domains with different sizes on the to-be-detected picture by adopting a sliding window method, and inputting each interest domain intoa crop weed classification recognizer for classification and recognition to obtain a corresponding prediction category and a correct probability; and screening out an interest domain corresponding tothe local maximum correct probability of each type from all interest domains by applying a non-maximum suppression algorithm, and outputting a classification and positioning prediction result. According to the method, crops and weeds can be quickly and accurately identified and positioned, and the requirement for data is low.

The invention discloses a pedestrian detection method in unmanned driving based on an improved YOLOv2. Firstly, a YOLOv2 network is trained by a KITTI dataset, and a training model is obtained. Then,the video is captured by a car camera, each frame in the video is used as the input of the YOLOv2 network, and the training model is imported into the YOLOv2 network. Then the network is run and the location information and probability of pedestrians are obtained. At last, after suggestion boxes which can not contain the pedestrian target area are screened out, the final pedestrian detection boxesare obtained by using the non-maximum suppression. The method can be effectively applied to pedestrian detection in unmanned driving.

The invention discloses a driving scene vehicle detection method based on an SSD neural network, comprising the following steps: constructing a data set and dividing the data set into a training set and a test set; based on Caffe's deep learning framework, using a SqueezeNet as a feature extraction network; selecting and merging six convolution layers of SqueezeNet network to be detected; after merging the six convolution layers of SqueezeNet network to be detected, adding a position regression layer and a class confidence discrimination layer to complete the construction of the training network model; obtaining a network pre-training model by initializing the training network model; using the network pre-training model, and obtaining the final training model by using the DSD method to carry out multiple rounds of training on the produced data set; capturing the forward image and inputting the image into the final training model, then using the non-maximum suppression algorithm to remove the redundant detection box so as to the detection results. The invention can quickly and accurately detect the vehicle target in front of the vehicle, and is a powerful measure for improving the environment perception ability of the intelligent driving vehicle.

The invention relates to the field of robot vision, pattern recognition and machine learning, in particular to an underwater robot target detection method, a system and related equipment, aiming at improving the robustness of target detection technology to underwater target occlusion, deformation and illumination changes. The object detection method of the invention comprises the following steps:obtaining an original image to be detected; normalizing the pixel value of the original image to be detected, and obtaining the image to be detected after preprocessing; the preprocessed image being input into the target detection network for detection, and the bounding frame of the region of interest and the probability of belonging to each target class being obtained; according to the bounding box of ROI and the probability of belonging to each target class, the improved non-maximum suppression algorithm being used to obtain the bounding box and the class of the target object, wherein, a deformable convolution neural network is used to extract feature map in the target detection network, and the candidate region method is used to detect the target. The detection method of the invention improves the detection precision under the condition of guaranteeing the speed.

The invention discloses a single-stage video behavior detection method. The method comprises the following steps: constructing a multi-scale behavior fragment regression network based on a convolutional neural network at a training stage; taking a training video and a frame-level real behavior label as inputs, and training the multi-scale behavior fragment regression network with an end-to-end optimization method of multi-task learning to obtain a well-trained multi-scale behavior fragment regression network model; when a new video is input at a use stage, generating an input frame sequence with the same length as the training video by a time dimension sliding window, and predicting the behavior category of the input frame sequence and a corresponding time position with the trained multi-scale behavior fragment regression network mode; and processing a prediction result by non-maximum suppression to produce a final behavior detection result. Through adoption of the method, the detection performance and detection efficiency can be improved.

The invention discloses a human face detection method and a human face detection system. The human face detection method comprises the following steps of establishing a convolutional neural network framework, wherein the convolutional neural network framework at least comprises a candidate region generation network, a correction network and a multi-information output network; connecting the convolutional neural network framework with a data preparation module; running the candidate region generation network, and generating multiple first human face candidate region frames; running the correction network, screening the first human face candidate region frames, performing position correction on the residual first human face candidate region frames, and performing non-maximum suppression; andrunning the multi-information output network, screening the first human face candidate region frames, performing the position correction on the residual first human face candidate region frames, performing the non-maximum suppression, and outputting human face feature points and human face poses corresponding to the residual first human face candidate region frames and the first human face candidate region frames.