111 results about "Region proposal" patented technology

Presented are systems and methods that provide a unified end-to-end detection pipeline for object detection that achieves impressive performance in detecting very small and highly overlapped objects in face and car images. Various embodiments of the present disclosure provide for an accurate and efficient one-stage FCN-based object detector that may be optimized end-to-end during training. Certain embodiments train the object detector on a single scale using jitter-augmentation integrated landmark localization information through joint multi-task learning to improve the performance and accuracy of end-to-end object detection. Various embodiments apply hard negative mining techniques during training to bootstrap detection performance. The presented are systems and methods are highly suitable for situations where region proposal generation methods may fail, and they outperform many existing sliding window fashion FCN detection frameworks when detecting objects at small scales and under heavy occlusion conditions.

Systems and methods for cuboid detection and keypoint localization in images are disclosed. In one aspect, a deep cuboid detector can be used for simultaneous cuboid detection and keypoint localization in monocular images. The deep cuboid detector can include a plurality of convolutional layers and non-convolutional layers of a trained convolution neural network for determining a convolutional feature map from an input image. A region proposal network of the deep cuboid detector can determine a bounding box surrounding a cuboid in the image using the convolutional feature map. The pooling layer and regressor layers of the deep cuboid detector can implement iterative feature pooling for determining a refined bounding box and a parameterized representation of the cuboid.

The invention discloses a fruit and vegetable detection method based on deep learning. The method comprises the following steps that: S1: firstly, preprocessing data, and carrying out manual calibration on an original picture in advance to obtain a segmentation tag, wherein the calibration means the coordinates of the left upper angular point and the right lower angular point of a target frame in the original picture, and the tag is used for judging whether a target in each calibration frame is a fruit and vegetable and determining the category of the fruit and vegetable; S2: secondly, training the data, taking the original picture and the picture tag as a training set of a deep learning neural network, and combining with a RPN (Region Proposal Network) and a Fast R-CNN to train the data to obtain a final fruit and vegetable detection model; and S3: finally, testing test data, calling a final fruit and vegetable detection model and a test program, carrying out fruit and vegetable detection on a test picture, and analyzing a final fruit and vegetable detection model effect through the observation of a test result.

Systems and method are provided for controlling an autonomous vehicle. A camera configured to capture an image, and a controller can execute an autonomous driving system (ADS) that classify that image. The ADS comprises a classification system for classifying objects in an environment within a driveable area of the autonomous vehicle. The classification system comprises a processor configured to execute a region proposal generator module and an image classification module. The region proposal generator module generates a set of bounding box region proposals for the image. The bounding box region proposals are selected areas of the image that include objects to be classified. The image classification module classifies, via a neural network executed by the processor, the objects from the image that are within one of the bounding box region proposals

The invention provides a real-time target detection method based on a region convolutional neural network. The real-time target detection method mainly comprises an input image, a target detection system, alternative optimization learning and sharing, and classifier classification and detection. The real-time target detection method comprises the steps of: regarding an image of any size as input, inputting a plurality of regions of interest (RoIs) while inputting the image, proposing a detection region by means of a region proposal network (RPN), utilizing the proposed detection region by an R-CNN detector, sharing all spatial positions by means of complete connection layers, learning shared characteristics by adopting alternative training optimization, and carrying out classification detection by using the classifier. According to the real-time target detection method, the RPNs are used for generating region proposals, and the network parameters are reduced by using shared weights, thus the region proposing step costs almost nothing; and the region proposal network (RPN) and the region convolutional neural network (R-CNN) share two network between a convolutional layer, thereby the cost is significantly reduced, the detection speed is fast, and the efficiency is high.

The invention discloses a pedestrian detection method. Multiple times of convolution and pooling are performed on an input image through the pedestrian detection method based on a convolutional neuralnetwork; the features of the original image are extracted so as to obtain the corresponding feature graph of the original image; the corresponding feature graph after zooming of the original image isapproximately calculated through the image feature pyramid rules; a candidate window is generated through a region proposal network RPN; a candidate proposal window is further selected and summarizedaccording to the pedestrian size distribution in the candidate window; the corresponding weight of different scales of pedestrian targets on different scales of images is trained by using the training data having the tag; and the classifier network is trained. The summarized candidate window is solved, and the confidence obtained through the classifier and the set threshold are compared and finalpedestrian detection judgment is performed. Heavy calculation amount of obtaining the feature graph through image zooming calculation can be avoided by application of the image feature pyramid, and detection is performed on different feature graphs by using the weighing mode of different weights so that misjudgment and leak detection caused by single feature graph detection can be effectively avoided.

The invention provides a text detection and localization method in a natural scene based on deep learning. The size of anchor and the regression mode in an RPN (region proposal network) based on Faster-R CNN are changed according to the characteristic information of text. An RNN network layer is added to analyze image context information. A text detection network capable of detecting texts is constructed. In addition, the size of anchor is set through clustering. In particular, cascaded training is carried out through mining difficult samples, which can reduce the false detection rate of texts. In the aspect of test, a cascaded test method is employed. Finally, accurate and efficient text localization is realized.

The invention provides a method and device for counting the passenger flow volume. The method comprises the steps of detecting width and height information of target frames in a real-time monitoring image of a region to be counted through the Faster R-CNN (Faster Region with Convolutional Neural Network); acquiring a width range and a height range of a preset number of target frames, and determining a size filtering interval by combining a preset empirical coefficient; filtering target candidate frames detected by an RPN (Region Proposal Network) in the real-time monitoring image of the regionto be detected according to the size filtering interval in the subsequent passenger flow volume counting, acquiring target candidate frames located in the size filtering interval, and inputting the acquired target candidate frames into a Fast R-CNN (Fast Region with Convolutional Neural Network) to obtain pedestrian target frames; tracking each pedestrian target frame in the region to be countedto form a tracking trajectory of each pedestrian target frame; and performing counting when the tracking trajectories meet a counting trigger condition. The method and device provided by the inventioncan greatly improve the accuracy of target detection and can be applicable to wider and more complex application scenarios.

The invention discloses a faster R-CNN-based hand posture detection and identifying method. The method includes the following steps: S1. using a Faster R-CNN network, setting parameters in a network for a hand posture identifying application; S2. making a marker for a hand posture sample, making the marker sample as input of the Faster R-CNN network, outputting effective non-linear features from a shared convolutional layer and inputting the effective non-linear features to a region proposal network RPN and the Fast R-CNN; the RPN which has the disturbance overlap rate algorithm acquiring a region proposal of a hand posture, and feedbacking the region proposal to the Fast R-CNN; and S3. the Fast R-CNN outputting a hand posture position and a hand posture type through a classification layer and a frame regression layer. The invention also provides a Faster R-CNN-based hand posture detection and identifying method which has increased robustness and higher accuracy.

The invention discloses a fast multi-label picture retrieval system and a realization method. The method comprises the following steps: deploying an RPN (Region Proposal Network) for extracting region proposals in a convolutional neural network, extracting region proposal information of pictures, and performing ROI pooling calculation on the region proposal information; after pooling, building a multi-label classification loss function through a fully connected layer according to multi-label information to train the convolutional neural network, and building a weighted three-dimensional loss function to train the convolutional neural network; extracting the hash code of each picture from a picture candidate set through the convolutional neural network after multi-task learning, saving the hash codes to a database, and comparing the hash codes with the hash codes in the database, thus completing picture retrieval. The whole network is trained through multi-task learning of classification and hashing, and therefore, the accuracy of retrieval is ensured. Moreover, the similarity is measured using Hamming distance in the process of retrieval, and the efficiency of retrieval is improved greatly.

A voxel feature learning network receives a raw point cloud and converts the point cloud into a sparse 4D tensor comprising three-dimensional coordinates (e.g. X, Y, and Z) for each voxel of a plurality of voxels and a fourth voxel feature dimension for each non-empty voxel. In some embodiments, convolutional mid layers further transform the 4D tensor into a high-dimensional volumetric representation of the point cloud. In some embodiments, a region proposal network identifies 3D bounding boxes of objects in the point cloud based on the high-dimensional volumetric representation. In some embodiments, the feature learning network and the region proposal network are trained end-to-end using training data comprising known ground truth bounding boxes, without requiring human intervention.

The invention discloses an automatic identification method of a thyroid tumor ultrasound image based on faster r-cnn. The method comprises the following steps: performing data enhancement on a markedthyroid tumor ultrasound image, and increasing the number and scale of training samples; performing feature extraction on an image data set by using a resnet-50 network model; generating a proposal window (proposals) by using a region proposal network RPN, and mapping the proposal window onto a feature map to generate a region proposal box; then causing each RoI to generate a feature map with a fixed size through RoI pooling; and finally performing joint training on a classification probability and border regression by using softmax Loss and softmax L1 Loss. By adoption of the method disclosedby the invention, the tumor ultrasound image does not need to be manually segmented, end-to-end network training can be achieved, and the identification rate is improved by data enhancement.

Various image processing may benefit from the application deep convolutional neural networks. For example, a deep multi-task learning framework may assist face detection, for example when combined with landmark localization, pose estimation, and gender recognition. An apparatus can include a first module of at least three modules configured to generate class independent region proposals to provide a region. The apparatus can also include a second module of the at least three modules configured to classify the region as face or non-face using a multi-task analysis. The apparatus can further include a third module configured to perform post-processing on the classified region.

The invention discloses a micro-operating system target detection method, including the steps of using a depth residual error convolutional neural network to perform characteristic extraction on a sample image to obtain sample characteristic graphs; using a region proposal network to perform convolution operation on the sample characteristic graphs to obtain sample target candidate boxes; using anon-line difficult sample mining method to screen the sample target candidate boxes to obtain new sample target candidate boxes, and using the sample characteristic graphs and the new sample target candidate boxes as training samples of a fully connected classification network to complete training of the fully connected classification network; and applying the depth residual error convolutional neural network to an image to be identified to obtain characteristic graphs, combined with the region proposal network, obtaining target candidate boxes, and through an area-of-interest pooling layer and the fully connected classification network that is trained, obtaining a target identification result. The micro-operating system target detection method provided by the invention is applied to target detection in the micro-operating system, can effectively position and identify each object, and at the same time, ensures requirements for an accuracy rate and real-time performance.

Provided is a dense small commodity rapid detection recognition method based on target detection. The method includes: processing an acquired commodity picture through a python or matlab programming language; establishing a convolutional neural network model including a convolution layer, a pooling layer and a total connection layer, wherein the model comprises convolution and RoIpooling; performing multi-task combined training on classification of the convolutional neural network model and a region proposal network model and bounding box regression through a training sample, performing learning update on a convolution kernel parameter in the convolutional neural network model through reverse propagation, and determining a hyper-parameter of the convolutional neural network model through averification set until a loss function reaches a target set value; and testing the recognition precision of the trained convolutional neural network model through the test set. According to the method, high-efficiency and accurate statistics of the number and the distribution of the commodities can be realized, the work efficiency of commodity suppliers and shopping mall management personnel canbe greatly improved, the manpower cost is reduced, and the commercial value is high.

One embodiment facilitates recognizing parts of a vehicle. A convolution module is configured to generate a convolution feature map of a vehicle image. A region proposal module is configured to determine, based on the convolution feature map, one or more proposed regions, wherein a respective proposed region corresponds to a target of a respective vehicle part. A classification module is configured to determine a class and a bounding box of a vehicle part corresponding to a proposed region based on a feature of the proposed region. A conditional random field module is configured to optimize classes and bounding boxes of the vehicle parts based on correlated features of the corresponding proposed regions. A reporting module is configured to generate a result which indicates a list including an insurance claim item and corresponding damages based on the optimized classes and bounding boxes of the vehicle parts.

A processing device for performing visual recognition using deep learning attributes and method for performing the same are described. In one embodiment, a processing device comprises: an interface to receive an input image; and a recognition unit coupled to the interface and operable to perform visual object recognition on the input image, where the recognition unit has an extractor to extract region proposals from the input image, a convolutional neural network (CNN) to compute features for each extracted region proposal, the CNN being operable to create a soft-max layer output, a cross region pooling unit operable to perform pooling of the soft-max layer output to create a set of attributes of the input image, and an image classifier operable to perform image classification based on the attributes of the input image.

The invention provides a method for automatically generating a thumbnail by use of a deep neutral network. The main content comprises the followings: data set training, boundary frame prediction, image and thumbnail size pair inputting, model training, quick automatic thumbnail generation FATG implementation. The processes are as follows: first training a data base established by use of internet pictures, inputting an original picture and a target thumbnail size, predicting a boundary frame position to determine a region containing important information in combination with a RPN (region proposal network) and R-FCN (region-based full convolution network by use of a FATG model, and producing a boundary frame with aspect ratio equal to the target thumbnail size, extending until to touch the boundary, namely, generating the thumbnail with the required size. By use of the automatic generating provided by the invention, the problems that the generation of the thumbnail is time-consuming and the important information is easy to lose are solved, the thumbnail under the required size can be acquired faster and more accurate, and the picture thumbnail is more precise and real-time to facilitate the daily browsing and image uploading of the people.

The invention discloses an unsupervised domain adaptive target detection method based on center alignment and relation significance, and the method comprises the steps: in a training stage, generatinga corresponding target region proposal for images of a source domain and a target domain through a detector; performing relation modeling on the target area proposal and the category center, and updating the category center and the target area proposal; shortening the distance of each category between the target domain and the source domain by utilizing the category center obtained by updating, so that the distances between different categories of the target domain are increased by means of source domain information; and after the training is finished, directly carrying out classification detection on the target domain image. According to the method, the category center does not need to be independently calculated, and the category center and the target area proposal are put into the graph to be updated together, so that the model can be trained end to end; when the category centers are aligned, the inter-category difference of the target domain can be expanded while the distributiondifference of the source domain and the target domain is reduced, so that the target domain is effectively classified.

The present invention provides a vision object tracking method based on hierarchical convolution. The method mainly comprises the content consisting of hierarchical convolution, correlation filters, translation estimation from rough to fine, region proposal and model updating. The method comprises the processes of: employing hierarchical features in a convolutional layer, and employing bilinear interpolation to regulate each feature map to a larger fixed dimension; performing normalization of the cycle version of input features to a soft target score generated by a Gaussian function; searchingthe maximum value of a target object on a response map; giving a related response map set; performing hierarchical deduction of each layer of target translation; calculating one confidence coefficient score of each proposal; keeping long-term memory of the target appearance; and finally, performing minimization of output errors to update an optimal filter. The vision object tracking method basedon hierarchical convolution mitigates sampling fuzziness, reduces tracking drift, reduces errors caused by reasons such as illumination change, shielding, background hybridization, sudden movement andtarget drift out of a visual field, and improves identification accuracy and robustness.

The invention discloses a face detection method based on a multi-scale cascade dense connection neural network, belongs to the fields of image processing and computer vision, and is applied to intelligent systems of face recognition, face expression recognition, driver fatigue detection and the like. The face detection method comprises a construction method of a region proposal network and a construction method of a multi-stage densely connected convolutional network model and particularly comprises the steps that face pictures annotated with face bounding box information are collected to forma training data set meeting the input conditions of various sub-networks; the cascade densely connected neural network with the high generalization capacity is constructed; the various sub-networks are trained by means of the training data set separately, and an overall network model is obtained; and finally multi-pose faces in the pictures are detected by means of the overall network model. According to the face detection method, by introducing the dense connection mode into the network, the network can fully extract face feature information, and then the accurate rate of face detection in multiple poses is increased.

Systems and methods for cuboid detection and keypoint localization in images are disclosed. In one aspect, a deep cuboid detector can be used for simultaneous cuboid detection and keypoint localization in monocular images. The deep cuboid detector can include a plurality of convolutional layers and non-convolutional layers of a trained convolution neural network for determining a convolutional feature map from an input image. A region proposal network of the deep cuboid detector can determine a bounding box surrounding a cuboid in the image using the convolutional feature map. The pooling layer and regressor layers of the deep cuboid detector can implement iterative feature pooling for determining a refined bounding box and a parameterized representation of the cuboid.

The invention discloses a detecting and tracking method, device and equipment of a target object in a video. The method comprises a step of inputting a continuous video frame into a convolutional neural network obtained by training in advance, wherein the convolutional neural network at least comprises a set number of shared convolutional layers and region proposal network layers, a step of extracting features of the continuous video frames by using the shared convolutional layers to obtain feature mapping maps corresponding to different video frames, a step of determines a target area associated with the target object according to the feature mapping maps by using the region proposal network layers, and a step of detecting the position and running trajectory of the target object in the continuous video frames based on the target area. According to the method, the detection and tracking are unified by using a convolutional neural network model, the amount of calculation is reduced, theproblem of detecting multiple poses and multiple angles of view of the target can be solved, the target detection rate is improved, and the false detection rate is reduced.

The invention provides a method and system for detecting and segmenting a vehicle in an aerial image. The method comprises the steps: extracting a feature map of a remote sensing image through a deepconvolutional neural network, and constructing a feature pyramid according to different scales of the feature map; performing adaptive feature fusion on the multilayer features in the original featurepyramid network, and outputting a new multi-scale feature map; extracting regions of interest corresponding to different scales from the multi-scale features by using a region proposal network; and sending the region of interest into a three-head network based on an attention mechanism for classification, bounding box regression and mask segmentation to obtain a classification result, a horizontal bounding box regression result and a mask segmentation result. According to the invention, the method can remarkably improve the vehicle detection and segmentation effects, and is higher in precision and robustness.

The invention discloses a pedestrian early warning method and system for freeway entrance and exit. The method comprises the steps of collecting a pedestrian data set, making sample tag files as training samples, and extracting features of pedestrians from the training samples by using a convolutional neural network method; according to the extracted pedestrian features, iteratively training a region proposal network and a target detection network to obtain a pedestrian detection network model; and shooting road condition videos in real time, transmitting the videos to a trained pedestrian detection network model for obtaining target probabilities and target boxes of the pedestrians, immediately giving an alarm and prompting pedestrian position information. The quantity and the geographicpositions of the pedestrians are determined while real-time pedestrian detection is performed; and under the remote condition, the detection speed is high and the accuracy is high, so that the freewayemergency event processing efficiency is improved, the introduction of additional errors can be avoided, and the processing accuracy is ensured.

Plant species can be distinguished mainly on the basis of the identification of plant leaf characteristics. Nevertheless, most identification systems show poor performance when small targets, including plant leaves and the like, are detected under a complex background. In order to improve the identification ability of plant leaves in a complex environment. The invention puts forward a plant leaf identification method based on deep learning. Inception V2 with BN (Batch Normalization) is used for replacing a convolutional neural layer in a Faster RCNN (Region Convolutional Neural Network) to provide multi-scale image features for an RPN (Region Proposal Network). In addition, an original image is firstly segmented into an appointed size according to a grid, and the segmented images are loaded to a network which is put forward in sequence. Through the accurate classification of Softmax and bounding box regressor, and the segmented imaged with identification tags are spliced to obtain a final image. An experiment result identifies leaves under the complex background. The method has higher identification accuracy than the Faster RCNN.

The invention belongs to the technical field of ship information detection, and discloses an SAR ship detection system and method based on a deep neural network, and the system comprises a fusion feature extraction module which is used for extracting features from an SAR image, and fully fusing the features from the bottom to the top and from the top to the bottom; a region proposal module which is used for classifying SAR image ships and backgrounds by taking the fusion features provided by the FEEN as input, and generating coarse candidate windows containing ship target positions; and a finedetection module which is used for taking the features provided by the FEEN and the coarse anchor frame provided by the RPN as input, refining the coarse anchor frame, and carrying out finer ship detection to obtain a final detection result. The detection method provided by the invention has good performance in multi-scale ship and small target ship detection under SAR complex backgrounds (far coast and near coast), and high ship detection precision is obtained.

An RPN (Region Proposal Network)-based optic disk positioning method comprises the steps of extracting overall characteristics of a fundus oculi image based on a deep convolutional neural network; preliminarily detecting an optic disk region based on an RPN network; and performing position finishing on an optic disk candidate region based on the deep convolutional neural network. According to theRPN-based optic disk positioning method, based on a deep learning method, the deep convolutional neural network is constructed to automatically position the optic disk, the accurate, rapid and robustpositioning of the optic disk can be implemented to assist in diagnosis treatment of a fundus oculi disease.