7127results about How to "Improve generalization ability" patented technology

The present invention discloses a multi-task learning convolutional neural network (CNN)-based face attribute analysis method. According to the method, based on a convolutional neural network, a multi-task learning method is adopted to carry out age estimation, gender identification and race classification on a face image simultaneously. In a traditional processing method, when face multi-attribute analysis is carried out, a plurality of times of calculation are required, and as a result, time can be wasted, and the generalization ability of a model is decreased. According to the method of the invention, three single-task networks are trained separately; the weight of a network with the lowest convergence speed is adopted to initialize the shared part of a multi-task network, and the independent parts of the multi-task network are initialized randomly; and the multi-task network is trained, so that a multi-task convolutional neural network (CNN) model can be obtained; and the trained multi-task convolutional neural network (CNN) model is adopted to carry out age, gender and race analysis on an inputted face image simultaneously, and therefore, time can be saved, and accuracy is high.

The invention discloses an unmanned aerial vehicle patrol detection image power small component identification method and system based on Faster R-CNN. The method comprises the following steps: carrying out pre-training on a ZFnet model, and extracting a feature graph of an unmanned aerial vehicle patrol detection image; training an RPN region proposed network model obtained through initialization to obtain a region extraction network, generating a candidate region frame on the feature graph of the image by utilizing the region extraction network, and carrying out feature extraction on the candidate region frame to extract position features and in-depth features of a target; carrying out training on a Faster R-CNN detection network obtained after initialization by utilizing the position features and in-depth features of the target and the feature graph to obtain a power small component detection model; and carrying out actual power small component identification detection by utilizing the power small component detection model. The beneficial effects are that Faster R-CNN is utilized to realize identification and positioning of a plurality of types of power small components, so that identification speed of about 80 ms per picture and 92.7% accuracy can be achieved.

The invention discloses an image semantic segmentation method based on deep learning. The method comprises four parts of data set processing, deep semantic segmentation network construction, deep semantic segmentation network training and parameter learning, and semantic segmentation on a test image. The RGB image and the gray level image of the input image are used as the input of the network model, the edge information of the gray level image is fully utilized, and the richness degree of input characteristics is effectively increased; a convolutional neural network and a bidirectional threshold recursion unit are combined, and more context dependency relationships and global feature information are captured on the basis of learning image local features; coordinate information is added tothe feature map through the first coordinate channel module and the second coordinate channel module, the coordinate features of the model are enriched, the generalization ability of the model is improved, and a semantic segmentation result with high resolution and accurate boundary is generated.

The invention discloses an image quality testing method based on a deep convolutional neural network. The image quality testing method based on the deep convolutional neural network comprises the following steps of firstly, establishing a sample set, establishing a deep convolutional neural network model, training the deep convolutional neural network model under different initial conditions, connecting the optimal deep convolutional neural network models obtained through the multiple times of training in parallel to obtain an image quality testing system, and using the obtained image quality testing system for testing images to be tested. According to the image quality testing method based on the deep convolutional neural network, characteristic learning is carried out by simulating the learning process of human brains, the problem that characteristics are difficult to select in an existing image quality testing method is solved, the contingency of predicting results is reduced, integration is high, generalization ability is strong, and testing effects are good.

The invention provides a method for detecting an X-ray mammary gland lesion image based on feature pyramid network under transfer learning. The method comprises the steps: 1, establishing a source domain data set and a target domain data set; 2, establishing a deformable convolution residual network layer by a deformable convolution and extended residual network module; 3, establishing a multi-scale feature extraction sub-network based on a feature pyramid structure through a feature map up-sampling and feature fusion method in combination with a deformable convolution residual network layer;4, establishing a deformable pooling sub-network sensitive to the focus position; 5, establishing a post-processing network layer to optimize a prediction result and a loss function; and 6, migratingthe training model to a small sample molybdenum target X-ray mammary gland focus detection task so as to improve the detection precision of the network model on the focus in the small sample image. According to the method, a transfer learning strategy is combined to realize focus image processing in a small sample medical image.

The invention discloses a zero sample image classification method based on combination of a variational autocoder and an adversarial network. Samples of a known category are input during model training; category mapping of samples of a training set serves as a condition for guidance; the network is subjected to back propagation of optimization parameters through five loss functions of reconstruction loss, generation loss, discrimination loss, divergence loss and classification loss; pseudo-samples of a corresponding unknown category are generated through guidance of category mapping of the unknown category; and a pseudo-sample training classifier is used for testing on the samples of the unknown category. The high-quality samples beneficial to image classification are generated through theguidance of the category mapping, so that the problem of lack of the training samples of the unknown category in a zero sample scene is solved; and zero sample learning is converted into supervised learning in traditional machine learning, so that the classification accuracy of traditional zero sample learning is improved, the classification accuracy is obviously improved in generalized zero sample learning, and an idea for efficiently generating the samples to improve the classification accuracy is provided for the zero sample learning.

The invention discloses a method and a system for enhancing features by the aid of bidirectional long-term and short-term memory recurrent neural networks. The method includes steps of training inputted features with noise and corresponding features without noise by the aid of bidirectional long-term and short-term memory recurrent neural network models; enhancing the features by the aid of the bidirectional long-term and short-term memory recurrent neural network models. The method and the system have the advantages that the long-term and short-term memory recurrent neural network models can be built in two directions by the aid of the inputted features, and accordingly models of contexts of current frames can be effectively built; long-term and short-term memory cells and certain control variables are introduced into the neural networks, and accordingly the models can be built in the dependence on the long contexts; excellent enhancement effects can be realized by the models owing to long-term dependence when the features are enhanced by the aid of the models, and accordingly the performance of voice recognition systems and audio event classification systems can be improved.

The invention discloses an automotive exhaust emission data fusion system. The automotive exhaust emission data fusion system comprises a roadside air pollutant concentration estimation module, a roadside air pollutant concentration prediction module, a city global atmospheric environment prediction module, an automotive exhaust emission factor estimation module and an automotive exhaust emission feather analysis module, wherein the five modules are used for respectively realizing different data analysis functions, and the different functions can be realized by virtue of the different modules; the modules can be independently used, or two or more modules can be combined for use, so as to realize the storage, analysis and fusion of automotive exhaust telemetering data, automotive attributes, driving working stations, detection time and meteorological condition data; and by combining with a vehicle-mounted diagnosis system database, a portable emission test system database, a vehicle inspection station offline database, a traffic information database and a geographic information database, automotive exhaust telemetering data is analyzed, and the highest discriminatory key indexes and statistical data are acquired, so that effective supports are provided for the formulation of relevant decisions of government departments.

The invention discloses a deep learning-based image high-density population counting method. The method comprises the following steps of S1, establishing a depth complementation convolutional neural network by utilizing a deep learning framework caffe; S2, performing image data enhancement on an image according to operations of angle rotation, image multi-scale zooming, image mirroring and image pyramid zooming; S3, performing Gaussian kernel fussy normalization processing on the enhanced image data to obtain a real crowd density graph, outputting an estimated density graph and the real density graph by the network, and performing continuous iterative training optimization on the whole network structure according to a loss function; and S4, inputting a crowd picture and a tag picture to the network for training, and performing continuous iterative optimization to obtain a trained network model finally. According to the method, the end-to-end convolutional neural network is designed; a picture is given and input, and the estimated density graph corresponding to the picture is output, so that an estimated crowd number is obtained; and by outputting the density graph, more useful information is reserved.

The invention discloses an image classification method based on a confrontation network generated through feature recalibration. The image classification method based on the confrontation network generated through feature recalibration is suitable for the field of machine learning and comprises the steps that to-be-classified image data are input into a confrontation network model for network training; a generator and a discriminator which are constituted by a convolutional network are constructed; random noise is initialized and input into the generator; the random noise is subjected to multilevel deconvolution operation in the generator through the convolutional network, and finally, generated samples are obtained; the generated samples and authentic samples are input into the discriminator; and the input samples are subjected to convolution and pooling operation in the discriminator through the convolutional network, thus a feature graph is obtained, a compressed and activated SENetmodule is imported into an intermediate layer of the convolutional network to calibrate the feature graph, thus the calibrated feature graph is obtained, global average pooling is used, and finally,image data classification is output. The SENet module is imported into the intermediate layer of the discriminator, the importance degree of each feature channel is automatically learned, useful features relevant to a task are extracted, features irrelevant to the task are restrained, and thus semi-supervised learning performance is improved.

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 embodiment of the invention discloses a face recognition method and a device. The method comprises the steps of extracting the Haar feature of a current to-be-recognized face image, and detecting the human face area of the to-be-recognized face image by adopting an ADaBoost classifier so as to obtain a face region image; performing the multi-scale feature extraction on the face region image by utilizing a convolution neural network model, and obtaining a feature vector of the face region image; inputting the feature vector, a pre-built legal face database and a preset user similarity threshold value into a multi-task learning model pre-constructed according to a Softmax loss function and a Triplet loss function, and judging whether the to-be-recognized face image is a legal user or not according to the output value of the multi-task learning model. The extracted feature is good in robustness and good in generalization ability. Therefore, not only the face recognition rate improved, but also the accuracy of face recognition is improved. The safety of identity authentication is improved.

The invention relates to a special audio event layered and generalized identification method based on a combination of an SVM (Support Vector Machine) and a GMM (Gaussian Mixture Model), and belongs to the technical field of a computer and audio event identification. The special audio event layered and generalized identification method comprises the following steps of: firstly, obtaining an audio characteristic vector file of a training sample; secondly, respectively carrying out model training on a great quantity of audio characteristic vector files (of the training samples) with various types by using a GMM method and an SVM method, so as to obtain the GMM model with generalization capability and an SVM classifier, and complete offline training; and finally, carrying out layered identification on the audio characteristic vector files to be identified by using the GMM model and the SVM classifier. With the adoption of the method provided by the invention, the problems that the conventional special audio event identification is low in identification efficiency on a continuous audio stream, very short in continuing time, high in audio event false dismissal probability can be solved. The method can be applied to searching a special audio and monitoring a network audio based on contents.

The invention relates to an LSTM (Long Short-Term Memory)-RNN model-based air pollutant concentration forecast method. The method comprises the steps of monitoring and collecting air pollutant concentration data in a relatively long time; preprocessing historical data to construct training, verification and test sample data of a to-be-trained LSTM-RNN model; obtaining a pre-trained LSTM-RNN model through the training sample data, then performing fine adjustment and training through the constructed verification and test sample data to obtain LSTM-RNN model parameters, improving model precision by further correcting model parameters, and taking a corrected LSTM-RNN model as an air pollutant concentration forecast model; and finally taking the preprocessed air pollutant concentration data of a target city in the relatively long time as input data of the LSTM-RNN model, and performing model output to obtain a forecast result of air pollutant concentration at a current moment or at a moment in the future.

The invention relates to a convolutional neural network based crowd density distribution heat map generation method, which comprises the steps of dividing a crowd picture set into a training sample set and a test sample set, performing crowd label image segmentation by using convolutional neural network architecture, and carrying out number regression by using a convolutional neural network; correcting a density map through a multi-scale template operation, generating a crowd density distribution heat map according to the corrected density map and the regression number, and completing crowd density distribution estimation. According to the invention, deep characteristics of an image are extracted by using a powerful learning ability of the full convolutional neural network so as to perform accurate crowd segmentation, and low efficiency and blindness of density calculation of a traditional method for full image characteristics are overcome; a crowd near-far perspective effect is overcome to a certain extent through multi-scale template correction; and mapping is performed in allusion to the estimated number of people, lateral comparison can be performed on heat maps of different cameras, the method is applicable to various crowd scenes, and the crowd density distribution heat map can be acquired in real time.

The invention discloses an orchard pedestrian detection method based on a YOLOv3 algorithm. The method specifically comprises the steps of collecting pedestrian images in an orchard; Preprocessing thecollected images, and constructing a standard pedestrian detection data set; Placing a training set into the modified Darknet-53 network structure to extract pedestrian features; Generating a predicted pedestrian boundary frame by using a K-means clustering method, performing category prediction by using a binary cross entropy loss function, and performing multi-scale fusion prediction by using asimilar FPN network; Finally, removing redundant prediction boundary frames through the Soft-NMS, and outputting final prediction boundaries and categories. The pedestrian detection accuracy is high,the real-time performance is good, the robustness of a training model to a complex background is enhanced by aiming at data augmentation methods such as Random Eraging proposed in an orchard environment, and through the adopted Soft-NMS algorithm, the recall rate of detection can be increased, and the introduced group normalization Group Normalizations can reduce the requirements of a trained model on hardware.

The invention discloses a deeper layer network constructing method used for gender identification or age estimation based on human face. The method includes a step (101) dividing all training pictures into a plurality of groups; (102) extracting high layer features of a group of pictures based on a convolution neural network and thereby obtaining a first matrix composed of the high layer feature vectors, and extracting low layer and global features of the same group of the training images based on an artificial neural network and thereby obtaining a second matrix composed of the low layer feature vectors, obtaining a group of gender identification or age estimation results based on the extract first matrix, the second matrix and the defined judgment formula, wherein the values of a first weight matrix W1, a second weight matrix w2, an offset matrix b and an adjusting weight beta in the defined judgment formula are updated by utilizing an error back propagation algorithm and the final values of the parameters are obtained and the network construction is completed. Judgment of age and gender of a human face is performed based on the judgment formula determined according to the values of the parameters when the network construction is completed.

The present invention discloses a workflow scheduling method based on depth enhancement learning. The method comprises the following steps of A) collecting M task execution workflow directed acyclic graphs (DAG) in an actual execution environment as a sample pool; B) carrying out the Markov decision process (MDP) modeling on each workflow DAG to generate a task state set S; C) according to a training method DQN of a neural network, using the task state sets S generated by the M workflow DAG and the corresponding known action sets A as the input to substitute into a deep neural network formula, and obtaining the value of a neural network parameter matrix. According to the present invention, and by the above method, the defects that the execution time of the workflow scheduling method is long and the generalization performance is poor under a current distributed environment, are solved, the time efficiency of an algorithm is guaranteed acceleratedly, at the same time, the generalization performance of the algorithm itself is increased, and a scheduling machine can learn a scheduling strategy autonomously according to the actual scene characteristics.

The invention provides a business risk assessment method and device, and a risk control system. The method comprises steps that to-be-assessed business data is received, and a risk behavior rule set matched with the business data is identified according to a preset risk rule engine; for multiple rules of the rule set, a risk value of each corresponding rule is assessed according to a pre-established risk assessment model; the risk values of the corresponding rules are accumulated, the accumulated risk value sum is compared with a total risk score threshold of each preset risk grade, and a risk grade of the business data is determined. The method is advantaged in that generalization capability and processing efficiency of a risk system can be improved.

The invention belongs to the field of deep learning and intelligent vehicle road detection, and specifically relates to a vehicle detection method based on improved YOLOv3. The method comprises the following steps: redesigning a convolutional neural network structure between a Darknet layer and three yolo layers, and designing a YOLO-TN network by referring to an idea of TridentNet weight sharing;carrying out model pruning on the YOLO-TN convolutional neural network; constructing a vehicle detection data set, and labeling vehicle position information in the data set; respectively training vehicle detection models based on the YOLO-TN and the YOLOv3, completing a vehicle detection task, and comparing detection results of the YOLO-TN and the YOLOv3. The method has high average precision while guaranteeing real-time detection, and is lower in omission ratio and more accurate in positioning when used for detecting distant vehicles and small-scale targets.