31results about How to "Solve the problem of insufficient training samples" patented technology

The invention discloses a medical image organ recognition method. The method comprises steps: a to-be-processed medical image is acquired, the medical image is segmented into a plurality of two-dimensional images in X-axis, Y-axis and Z-axis directions respectively, and a detection window is set according to the size of a target organ; the detection window is used to carry out traversing detection on the two-dimensional images respectively according to a set detection step length, and detection results in the X-axis, Y-axis and Z-axis directions are acquired; and result fusion is carried out on the detection results, pixel points which are detected to be positive in the X-axis, Y-axis and Z-axis directions are kept, and a target organ boundary is determined. The medical image organ recognition method of the invention can quickly and accurately recognize the target organ area, the target organ boundary is determined and the adaptive ability is strong. In addition, the invention also provides a medical image organ segmentation method.

The invention discloses a CNN model, a CNN training method and a vein identification method based on CNN. The CNN model comprises multiple convolution layers, a full connection stair layer and a SoftMax layer. In a CNN training process, a database is firstly expanded and multiple biological feature databases including similar features are combined to carry out a training of a model; the full connection layer and the SoftMax layer serve as a multi-classification classifier together; a multi-classification neural network is trained so that the neural network is allowed to learn features capable of identifying vein features; and after the training is finished, the previous layer of the full connection layer is output to serve as feature, and similarity of a pair of images is measured by calculating cosine distance of the features. According to the invention, multi-mode biological feature data is fused and used for training a network, so a problem of insufficient training samples is solved and retrieval speed can be greatly improved in a super large identity identification database.

The invention discloses a sea cucumber autonomous identification and grabbing method based on deep learning and binocular positioning. The sea cucumber autonomous identification and grabbing method comprises the following steps: performing underwater sea cucumber identification and positioning based on deep learning; acquiring sea cucumber spatial positioning information by utilizing binocular stereoscopic vision; and performing sea cucumber grabbing by using a PID control method. According to the invention, the GAN model is used to learn the characteristics of underwater sea cucumbers, and the generation network is used to generate sea cucumber samples, thereby effectively solving the problem of sea cucumber training sample insufficiency. According to the invention, mean filtering, medianfiltering and Wiener filtering are combined into a design filtering operator, so that the influence of non-uniform light, high turbidity, low visibility and the like on the image is solved. Accordingto the method, the convolutional neural network is utilized to learn and conclude the existing data, the sea cucumbers are accurately and quickly detected and two-dimensionally positioned, and a powerful guarantee is provided for subsequent spatial three-dimensional positioning and grabbing of the sea cucumbers. High-precision camera internal and external parameters are obtained, and accurate grabbing of the manipulator is guaranteed.

The invention provides a vehicle automatic driving virtual environment construction method. The vehicle automatic driving virtual environment construction method comprises the following steps: step 1,modeling and generating a virtual scene; 2, virtual data and signals are collected; Step 3, automatically generating a special type marker in the virtual scene; 4, map elements and resources in the virtual scene are dynamically loaded based on the special type markers; 5, according to the virtual data and the signals, labeling the interested target in the virtual scene; Step 6, based on a dynamicmodel and the methods in the step 1, the step 2 and the step 3, simulating and carrying out an automatic driving system boundary test in a virtual scene; And step 7, testing an identification resultof the automatic driving system in the virtual scene based on the methods of the step 1, the step 2 and the step 5. The virtual target object and the special type marker are generated through programsoftware to replace the scene of the real world and the acquisition process of the marker, and expensive acquisition equipment is prevented from being used to reduce the cost.

The invention discloses an abnormal flow alarm log detection method and device, equipment and a medium, and the method comprises the steps: respectively generating a first flow alarm log and a secondflow alarm log corresponding to an original flow package according to a first preset alarm rule and a second preset alarm rule, wherein the precision of the first preset alarm rule is higher than thatof the second preset alarm rule; labeling the first flow alarm log, and labeling the second flow alarm log according to the label in the first flow alarm log; taking the second flow alarm log with the label as sample data, and training a preset flow alarm log classification model by utilizing the sample data; and utilizing the trained flow alarm log classification model to classify the obtained to-be-detected flow alarm logs so as to determine whether the to-be-detected flow alarm logs are abnormal flow alarm logs or not. Therefore, the detection accuracy can be improved, the false alarm rateis reduced, and the traffic threat detection capability is enhanced.

The invention discloses an underwater acoustic signal modulation mode inter-class identification method based on an improved dense neural network, and the method comprises the following steps: firstlyreceiving a to-be-identified underwater acoustic modulation signal, and extracting the features of the underwater acoustic modulation signal; performing dimension reduction and denoising on underwater acoustic modulation signal features by using a principal component analysis method; performing normalization and dimension change; based on the dense neural network, removing the pooling layer to obtain an improved dense neural network, and training the neural network; and inputting the processed underwater acoustic modulation signal features into the trained improved dense neural network, and finally completing modulation mode inter-class identification. Low-delay and high-accuracy inter-class recognition of the underwater acoustic signal modulation mode is finally realized, and the recognition method is high in anti-interference capability, low in calculation cost and high in recognition accuracy.

The invention discloses a three-dimensional face recognition method based on Bayesian multivariate distribution characteristic extraction. The method comprises three-dimensional data preprocessing, characteristic extraction and identification classification. The method has the advantages that the defect of large computation amount in the prior art is overcome; a three-dimensional face depth map is used for identifying, so that the computation amount can be reduced, and the identification efficiency is increased; the problem of insufficient training samples in single sample identification can be solved, and the training samples are added with a partitioning method; a characteristic extraction method based on Bayesian analysis is provided on the basis, so that the obtained characteristics have a minimum intra-class distance and a maximum inter-class distance, namely, optimal separability is realized; moreover, a classification method based on Mahalanobis distance is adopted, so that optimal identification classification is realized. As proved by experiment data, the method disclosed by the invention has a good three-dimensional face recognition result.

The invention provides an air conditioner user frequency modulation capability evaluation method based on a generative adversarial network. According to an air conditioner user frequency modulation mechanism, selecting measurement data of frequency modulation capability and strong correlation factors influencing the frequency modulation capability, and constructing a small sample training set, a small sample generation set and a test sample set of a frequency modulation capability evaluation model; improving a generator model of a generative adversarial network algorithm, training the improvedgenerative adversarial network model by using the small sample training set to obtain a trained improved generative adversarial network model, further generating a synthetic sample set by using the small sample generation set, and constructing a training set of a frequency modulation capability evaluation model; and constructing a multi-layer feedforward neural network model, training the model by using the training set of the frequency modulation capability evaluation model, and obtaining the trained multi-layer feedforward neural network model as an air conditioner user frequency modulationcapability evaluation model. The accuracy of the air conditioner user frequency modulation capability evaluation model is improved.

The invention discloses an audio event classification method based on stack-based sparse representation and a computer device. The method comprises the following steps: at the training stage, first, creating audio dictionaries of all kinds of audio events; then, constructing a large-scale dictionary by stacking the audio dictionaries of the all kinds of audio events; at the testing stage, extracting a sparse representation coefficient of a tested audio sample according to the large-scale dictionary constructed at the training stage, and mapping the sparse representation coefficient through a softmax function; and finally, constructing the confidence degree of the tested audio file on the all kinds of audio events according to the mapped coefficient, and carrying out classified judgment according to the magnitude of the confidence degree. According to the method, the large-scale dictionary is constructed through the stacking base innovatively, and then the sparse representation coefficient of the sample is obtained; the extracted sparse representation coefficient can well represent the audio event sample, the inter-class difference of the samples is increased, the intra-class difference is reduced, and the classification accuracy is improved.

The embodiment of the invention provides an enterprise financial service risk prediction method and device. The method comprises the steps that operation state information of a target enterprise whichis not authorized by financial service currently is acquired; inputting the operation state information of the target enterprise into a weak supervision scoring model for financial service risk prediction, and taking the output as a financial service risk prediction level of the target enterprise to determine whether to provide financial service for the target enterprise or not; wherein the weaksupervision scoring model is obtained by scoring a plurality of enterprises by applying a fusion model in advance, the fusion model is obtained based on a marking model and historical enterprise datawith unknown labels, and the marking model is obtained by training historical enterprise data with known labels in advance. According to the invention, the accuracy and the intelligent degree of the financial service risk prediction process of an enterprise without financial service authorization can be effectively improved, the enterprise financial service risk prediction efficiency can be effectively improved, and the operation reliability and safety of a financial institution can be improved.

The invention discloses an appearance defect detection method and device for an industrial product and a storage medium. The appearance defect detection method for the industrial product comprises thefollowing steps: acquiring an appearance image of the industrial product to be detected; based on the appearance images, performing appearance defect detection on the industrial product through a pre-trained appearance defect detection model, and training the appearance defect detection model through the following operations: obtaining multiple sample images of a sample industrial product with appearance defects; performing data enhancement on the plurality of sample images by adopting a preset data transformation rule to generate a first sample image set; based on the first sample image set,generating a second sample image set by using a preset DCGAN model; and training an appearance defect detection model by using the second sample image set.

The invention discloses an underwater acoustic signal modulation mode inter-class identification method based on an improved dense neural network. The method comprises the following steps: firstly receiving a to-be-identified underwater acoustic modulation signal, and extracting the features of the underwater acoustic modulation signal; carrying out dimension reduction and denoising on underwater acoustic modulation signal features by using a principal component analysis method; carrying out normalization and dimension change; based on the dense neural network, removing the pooling layer to obtain an improved dense neural network, and training the neural network; and inputting the processed underwater acoustic modulation signal features into the trained improved dense neural network, and finally completing modulation mode inter-class identification. Low-delay and high-accuracy inter-class recognition of the underwater acoustic signal modulation mode is finally realized, and the recognition method is high in anti-interference capability, low in calculation cost and high in recognition accuracy.

The embodiment of the invention provides an enterprise financial service risk prediction method and device, which can be used in the technical field of artificial intelligence, and the method comprises the steps that the operation state information of a target enterprise which is not authorized by a financial service at present is input into a financial service risk prediction model to obtain a financial service risk prediction level, whether to provide the financial service to the target enterprise is determined based on the financial service risk prediction level; the financial service risk prediction model is obtained after multiple enterprises are scored by applying a fusion model in advance, the fusion model is obtained based on a marking model and historical enterprise data with unknown labels, and the marking model is obtained through training based on historical enterprise data with known labels processed in a transfer learning mode and a resampling mode. The accuracy and reliability of the financial service risk prediction process of the target enterprise without financial service authorization of the target financial institution can be effectively improved, and the pertinence and effectiveness of the financial institution for providing the financial service for the enterprise can be improved.

The invention provides a construction method of a special marker in vehicle automatic driving virtual environment. The method comprises the following steps of: step I: determining the type of a modelneeding to generate a special type marker, and acquiring corresponding model parameters; step II: calling an application program interface corresponding to the special type marker model, and executinga generation algorithm, thereby obtaining a marker object description compatible with a virtual scene; and step III: outputting the marker object to a memory. In addition, the invention also providesa construction method of vehicle automatic driving virtual environment based on the marker. By adopting the road generation algorithm in the construction method of the special marker of the invention, special markers such as roads and the like can be vividly simulated, and data acquisition and road adjustment in a simulation environment are facilitated. In addition, a virtual target object and aspecial type marker are generated in the invention through a program software to replace an acquisition process of a scene and the marker of the real world, thereby avoiding using expensive acquisition equipment and reducing the cost.

The invention provides an automatic sample key point labeling method, device and system, and the method comprises the following steps: receiving a real-time image, and extracting SIFT features of thereal-time image; calculating a matching point according to the SIFT features of a target template image and the SIFT features of the real-time image; calculating a transmission transformation matrix from the target template image to the real-time image according to the matching point; calculating the positions of key point positions in the target template image in the real-time image according tothe transmission transformation matrix; and storing the real-time image and the corresponding key point positions as sample data. According to the method, an existing template matching technology is utilized, targets and corresponding key points under different backgrounds are matched in real time, and high-quality image-key point sample pairs are acquired in batches. Compared with a traditional manual sample labeling method, the method is higher in efficiency and better in universality.

The invention discloses a hyperspectral image classification method and system based on 3D CutMix-Transform. The method comprises the following steps: dividing hyperspectral data into a labeled training data set and a labeled verification data set; the 3D CutMix is pre-trained through the CNN; training a region-level teacher model and a sample-level teacher model by using enhanced data obtained by performing 3D CutMix on the training data set; and jointly training a student model by using the two teacher models and a small amount of labeled data sets. According to the method, 3D CutMix pre-training is carried out by using CNN, then data enhancement is carried out on an original label data set by using 3D CutMix, and respective self-supervision loss and mutual cross pseudo-supervision loss of two teacher models are optimized, so that the co-trained student models are better in robustness, and the training efficiency is improved. The generalization ability and accuracy of the model under the small sample in the existing hyperspectral image classification technology are enhanced and improved, and the method can be used for hyperspectral image classification.

The invention discloses a CNN model, a CNN training method and a CNN-based vein recognition method. The CNN model includes a plurality of convolutional layers, a fully connected layer and a SoftMax layer; The biometric database of the feature is used to train the model; the fully connected layer and the SoftMax layer are used as a multi-class classifier; a multi-class neural network is trained to learn the characteristics that can distinguish the vein category; after the training is completed, The output of the previous layer of the fully connected layer is used as a feature, and the similarity of a pair of images is measured by calculating the cosine distance of these features. The invention integrates the multimodal biometric database for training the network, solves the problem of insufficient training samples, and can greatly improve the retrieval speed in a super-large identity authentication database.

The invention discloses a production reality-oriented safety helmet detection method. The method comprises the following steps of 1, collecting and marking initial training data; 2, training the data set Cf obtained in the step 1 to obtain an FCOS target detection model Mf; 3, training the data set Cr obtained in the step 1 to obtain a Resnet18 target classification model Mr; and 4, inputting an original image, and carrying out production workshop image detection by using a head detection positioning-safety helmet reclassification two-step method. According to the robust safety helmet detection system suitable for the complex production scene, the labor cost is greatly reduced, and the practical application value of the method is improved; according to the method, FCOS and Resnet18 are subjected to cascade detection, so that the problem of insufficient training samples is effectively solved; according to the method, while the target position is quickly detected, the target features are accurately extracted, and accurate target classification is realized.

The invention discloses an automatic recommendation method for a signal reconstruction scheme. The method comprises the following steps: 1, constructing a single-target optimization model for signal reconstruction; 2, constructing a data set; 3, for the data set, using a deep recurrent neural network to train a recommendation model; 4, fitting a signal to reconstruct a corresponding tree structure and acquiring features; and 5, inputting features obtained after fitting into the model so as to recommend a signal reconstruction scheme. According to the method, automatic recommendation of the signal reconstruction scheme can be realized, and the recommendation accuracy and efficiency are improved.

The invention belongs to the field of rolling bearing fault diagnosis, provides an aero-engine rolling bearing fault diagnosis method based on twin network metric learning, combines a twin network with metric learning, and provides an aero-engine bearing fault diagnosis method based on a twin network structure and adopting a metric learning strategy. The original learning strategy of the twin network is changed, and the difference between samples is measured by adopting the distance of sample features in metric learning. Under the condition that the sample size is seriously insufficient, a multi-classification task with the insufficient sample size is converted into a plurality of dichotomy tasks by adopting a form of constructing a learning task, so that available model training tasks are greatly expanded. A powerful support is provided for the aero-engine bearing fault diagnosis problem under the condition that samples are lacked, and meanwhile the method has high engineering practical application value.

The invention discloses a method for identifying organs in a medical image, comprising: acquiring a medical image to be processed, splitting the medical image into several two-dimensional images in the directions of X, Y, and Z axes, and Set the size of the detection window; use the detection window to traverse and detect the two-dimensional image according to the set detection step, and obtain the detection results in the X, Y and Z axis directions; perform result fusion on the detection results , keep the pixel points that are detected as positive in the three directions of the X, Y and Z axes, so as to determine the boundary of the target organ. The method for identifying organs in medical images of the invention can accurately and quickly identify areas where target organs are evenly located, determine the boundaries of target organs, and has strong self-adaptive ability. In addition, the present invention also provides a method for segmenting organs in medical images.

The invention discloses a rib fracture image detection method based on small sample deep learning. The method comprises the following steps: step 1, marking a collected rib fracture CT image; 2, performing feature extraction by adopting a YOLOv3 model; 3, carrying out transfer learning on the extracted feature vectors; and 4, carrying out deep learning training. Compared with the prior art, the method has the positive effects that a YOLOv3 model is provided, small samples are processed by means of a transfer learning method, and a deep learning model for rib fracture image detection is established. According to the method, few rib CT images are used as data of an input layer for deep learning training, the obtained effective rate is high, and therefore a small sample deep learning model can be established, and a good basis is provided for diagnosis of doctors.

The invention discloses an audio event classification method based on stack-based sparse representation and a computer device. The method comprises the following steps: at the training stage, first, creating audio dictionaries of all kinds of audio events; then, constructing a large-scale dictionary by stacking the audio dictionaries of the all kinds of audio events; at the testing stage, extracting a sparse representation coefficient of a tested audio sample according to the large-scale dictionary constructed at the training stage, and mapping the sparse representation coefficient through a softmax function; and finally, constructing the confidence degree of the tested audio file on the all kinds of audio events according to the mapped coefficient, and carrying out classified judgment according to the magnitude of the confidence degree. According to the method, the large-scale dictionary is constructed through the stacking base innovatively, and then the sparse representation coefficient of the sample is obtained; the extracted sparse representation coefficient can well represent the audio event sample, the inter-class difference of the samples is increased, the intra-class difference is reduced, and the classification accuracy is improved.

A 3D face recognition method based on Bayesian multivariate distribution feature extraction, including 3D data preprocessing, feature extraction and recognition classification. The advantages of the present invention are: to overcome the shortcomings of the large amount of calculation in the prior art, the present invention uses a three-dimensional face depth map for recognition, which can reduce the amount of calculation and improve the recognition efficiency; and solve the problem of insufficient training samples in the single-sample recognition problem , using the block method to increase the training samples; on this basis, a feature extraction method based on Bayesian analysis is proposed, so that the obtained features have the smallest intra-class distance and the largest inter-class distance, that is, the best separable and use the classification method based on the Mahalanobis distance to obtain the optimal recognition classification. Experimental data proves that the method of the present invention has better three-dimensional face recognition results.

The invention discloses an under-screen fingerprint image generation method and device, and relates to the technical field of image processing. The generation method comprises the following steps: training a generative adversarial neural network for mutual style conversion between a non-under-screen fingerprint image and an under-screen fingerprint image; and performing image conversion by using the trained generative adversarial neural network, and converting the input non-under-screen fingerprint image into a generated under-screen fingerprint image. The generation device is sequentially provided with an acquisition unit, a conversion unit and an output unit; the acquisition unit is used for acquiring the non-under-screen fingerprint image; the output end of the acquisition unit is connected with the conversion unit, and the conversion unit is used for converting the acquired under-screen fingerprint image into the generated under-screen fingerprint image; and the output end of the conversion unit is connected with the output unit, and the output unit is used for outputting the generated under-screen fingerprint image. According to the method, a large number of existing non-under-screen fingerprint images are used for generating the under-screen fingerprint images, and the problems that in the fingerprint recognition field, the number of existing under-screen fingerprint image samples is insufficient, collection is difficult, and privacy is leaked are solved.

The present invention relates to a method for estimating air quality in urban areas based on collaborative training. The present invention makes full use of regional spatial features, such as traffic conditions and road network structures in the area, and is a semi-supervised learning method based on multi-classifier collaborative training. Model the feature vector of the region to learn multiple classifiers; then pruning these classifiers to form the final combined classifier; use the pruned combined classifier model to perform air monitoring for areas without air monitoring stations. quality rating estimates. This method can estimate the air quality level according to the spatial differences between the areas with air monitoring stations and the areas without air monitoring stations under the condition of limited air monitoring stations, and the estimation results are accurate.

The invention discloses a finger vein verification method and device based on a neural network and a storage medium. Finger vein verification is carried out by adopting a dual-channel network with only three convolution layers; the problem of insufficient training samples in finger vein recognition is effectively solved. According to the method, firstly, an original image is acquired, then the sensitivity to finger displacement in the original image is solved by extracting the minimum region of interest from the original image, and finally, the original image and the minimum region of interestare integrated by using a double-flow network, so that the problem of insufficient training samples in finger vein recognition is effectively solved, and the requirement of effective solution of a neural network is met.

The invention provides a face recognition method and apparatus, a computing device and a storage medium. The method comprises the steps of obtaining a to-be-recognized face image; a target image corresponding to the to-be-recognized face image is acquired through a conversion network, and the face deflection angle of the target image is smaller than or equal to a preset threshold value; and face recognition is carried out based on the target image, so that the accuracy of face recognition is effectively improved.

The invention discloses a power distribution network investment decision-making method based on deep transfer learning, and the method comprises the steps: collecting and screening data, describing the data distribution characteristics of a power grid through an edge distribution probability, representing the network relation characteristics of the power grid through a conditional distribution probability, and completing the characteristic transfer from a source domain power grid to a target power grid, so that adaptive learning under a small sample of power distribution network investment is realized, and finally an input-output nonlinear mapping model based on a target power distribution network is established to make a decision on power distribution network investment. According to the method, a power grid investment input-output association relationship is constructed through a deep learning network, a power grid investment decision problem is analyzed from the perspective of pure data, a transfer learning process is introduced, and data distribution characteristics and network relationship characteristics are migrated from other similar power grids through a small number of samples of the transfer learning process by utilizing the self-adaptive characteristic of the transfer learning process, so that the problem that training samples are insufficient in the association mining process of an existing data driving method is solved.

The invention proposes a method for evaluating the frequency modulation capability of an air conditioner user based on a generative confrontation network. According to the frequency regulation mechanism of air conditioner users, select the measurement data of the frequency regulation capability and the strong correlation factors affecting the frequency regulation capability, and construct a small sample training set, a small sample generation set and a test sample set for the frequency regulation capability evaluation model; improve the generator of the generative adversarial network algorithm model, and use the small sample training set to train the improved generative adversarial network model to obtain the improved generative adversarial network model after training, and further use the small sample generation set to generate a synthetic sample set, and construct the training set of the frequency modulation capability evaluation model; The multi-layer feedforward neural network model is trained by using the training set of the frequency modulation capability evaluation model, and the trained multi-layer feedforward neural network model is obtained as an air conditioner user frequency modulation capability evaluation model. The invention improves the accuracy of an air conditioner user's frequency modulation capability evaluation model.