187 results about "Small data sets" patented technology

The invention discloses a human behavior recognition method based on an attention mechanism and a 3D convolutional neural network. According to the human behavior recognition method, a 3D convolutional neural network is constructed; and the input layer of the 3D convolutional neural network includes two channels: an original grayscale image and an attention matrix. A 3D CNN model for recognizing ahuman behavior in a video is constructed; an attention mechanism is introduced; a distance between two frames is calculated to form an attention matrix; the attention matrix and an original human behavior video sequence form double channels inputted into the constructed 3D CNN and convolution operation is carried out to carry out vital feature extraction on a visual focus area. Meanwhile, the 3DCNN structure is optimized; a Dropout layer is randomly added to the network to freeze some connection weights of the network; the ReLU activation function is employed, so that the network sparsity isimproved; problems that computing load leap and gradient disappearing due to the dimension increasing and the layer number increasing are solved; overfitting under a small data set is prevented; and the network recognition accuracy is improved and the time losses are reduced.

The invention discloses a small data set face recognition algorithm based on machine vision, which comprises the following steps: 1) construction and pretreatment of the data set; 2) generating a virtual image for a category with insufficient images by using a generated antagonistic network; 3) carrying out data expansion on all kinds of images by using a data enhancement algorithm; 4) using convolution neural network to construct a model suitable for data; 5) training the design model; 6) using a camera to obtain a target image, importing the trained model, and carrying out face recognition.The algorithm of the invention can carry out face recognition with high accuracy for application scenes lacking face data.

The invention discloses an image classification method and an image classification device. The image classification method comprises the following steps: based on a big image data set, training an AlexNet model structure; migrating the five trained convolution layers to a small database to form a lower-level feature extraction layer, and constructing together with a residual network layer, a multiscale pooling layer, a feature layer and a softmax classifier to obtain a migration model structure, wherein the residual network layer includes two convolution layers; inputting small image data set into the migration model structure, upgrading parameters by adopting a batch gradient descending method, and training an image classification hybrid model; and classifying according to the image classification hybrid model to obtain a classification result. By migrating the pretrained convolution layers on the big data set to the small data set, increasing the multiscale pooling layer, and serially connecting the feature quantity output by the residual network layer and the multiscale pooling layer and inputting to the classifier, the feature quantity is increased and the overfitting problem can be relieved; and through the hybrid model trained based on a convolution nerve network and a migration learning, the image classification accuracy can be effectively improved.

The present invention relates to a system, method and computer program product that is a multi-dimensional data mining environment and that operable to apply a series of temporal and relative rules (i.e., STDMn0) and is further operable in at least one of the following ways: to incorporate a framework to support temporal abstractions and relative alignments to data (i.e., STDMn0); and to derive characteristics within the data (STDMn0). The present invention may incorporate data from multiple sources, and potentially multiple centres. The analysis and alignment of the data may involve both temporal dimensions and other dimensions (or relative aspects) of the data. The present invention may further be a data mining environment that is flexible enough to permit relatively open ended queries thereby enabling, for example, the detection of trends, including trends with new dimensions, or trends based on relatively small data sets.

The invention discloses a dam image crack detection method based on transfer learning. The method comprises the following steps of collecting the dam crack image, preprocessing the image data set to fill the data set through a generated antagonism network GAN; using a pre-training model MobileNet that does not retain a top layer and a full-connection layer to extract image features, splicing a Flatten layer after MobileNet, splicing the full connection layer whose activation function is ReLU after the Flatten layer, and splicing the Full Connection Layer whose activation function is Sigmoid asthe output layer; frozening the first K depth decomposable convolution structures in MobileNet, and fixing the relative weights of these K depth decomposable convolution structures; training the model and only updating the weights of the unfrozen network layer in the process of model training; using the trained model to detect the dam cracks in the image. The method of the invention solves the over-fitting problem under the condition of small data set, and improves the prediction performance and the running speed through the migration learning idea.

The invention discloses a pedestrian re-identification method based on transfer learning and feature fusion, comprising the following steps of acquiring the pedestrian data, carrying out initial training through a neural network, modifying the structure and combining the improved loss function to carry out re-training in a data set; carrying out the manual feature extraction and neural network feature extraction; after extracting features, fusing the two features to get high-Low grade features; using an XQDA algorithm to calculate the high-Low grade features, classifying and verifying to obtain the re-recognition results. The method of the invention adopts the cross entropy loss function and the triple loss function to restrict the whole network more strongly, and then extracts manual features and convolution network features to carry out feature fusion, and forms high-low features, covers different levels of pedestrian feature expression, achieves better recognition results, and reduces the training time in the form of fine-tuning, and has good generalization and portability for small data sets.

The embodiment of the invention discloses a high-resolution remote sensing image scene classification method for a small data set; and a three-scale three-channel end-to-end multi-scale convolution joint neural network model is established. The method comprises three steps: step one, random sub region extraction is carried out on a remote sensing image with an image size of N*N according three dimensions, wherein the dimensions of the sub regions of the image after extraction are N / 2*N / 2, N / 4*N / 4, and N / 8*N / 8 and inputs of a multi-channel convolution feature extractor; step two, a feature joint network based on multi-scale and multi-channel fusion is established and two-two fusion enhancement is carried out on features with different dimensions to realize a high-layer-feature joint enhancement expression; and step three, on the basis of a joint loss function, high-layer features of joint enhancement of the scene are classified. Therefore, high-precision classification of a small sample training set is realized; and the high-layer feature expression capability in the small sample data set and the computing speed are improved.

The invention discloses a scene identification method based on a mixed depth structure. A prior mixed depth identification frame is improved and applied to the task for scene identification, a depth self-encoder is adopted to automatically extract the local image block characteristic to replace the local characteristic extraction layer of a conventional mixed depth network to obtain an image block high-grade local characteristic; at the same time, spatial information is introduced to improve local characteristic coding layer of scene identification, and at the end, the scene is identified via depth discrimination network to improve the mixed depth mixed scene identification frame, so that the improved mixed depth scene is approximate to a convolutional neural network in the aspects of form and identification accuracy and is higher than a depth convolutional neural network in the aspect of calculation efficiency. In addition, the scene data is selectively expanded for the within-class difference and intra-class similarity of scene data, the construction robustness is high, and the method is suitable for the depth mixed scene identification model of a small data set.

The invention discloses a single classifier anomaly detection method based on a multilayer random neural network. The the method comprises: only inputting a training data set of a normal class; through multilayer ELM-AE autoencoder and decoding processing, input sample data obtaining a reconstructed characteristic value; inputting the reconstructed characteristic value into the last layer of ELM to obtain actual output; sorting the obtained distance error vectors of the actual output and output tags from large to small, and determining a threshold for separating a normal class from an abnormalclass according to a set threshold parameter; and finally, inputting the test data into the multi-layer random neural network single classification abnormity detection model, and testing the recognition effect of the model. According to the method, main information is extracted more quickly and efficiently, dimensionality reduction is carried out, and then recognition and classification are carried out. And the speed is higher, the accuracy is higher, and the generalization performance is better. The method is not only suitable for small data sets, but also suitable for high-dimensional largedata sets, and has universality. And the method has important significance for practical application in future.

The invention provides an image high-level semantic recognition method based on multi-feature fusion of a deep network. A global color histogram extracts the color features of an image, the LBP algorithm extracts the texture feature of the image, the deep object network extracts the object features of the image and the deep emotion network extracts the deep emotion features of the image to be fused to identify the composite emotion of the image and the main object contained, Finally, for the input image, the network model can generate descriptive phrases with high-level semantic information, including emotional semantics and object semantics. The invention starts from the aspect of depth learning for small data sets. In this invention, a new method of data expansion is proposed, which combines pre-extracted low-level features such as color and texture statistics, and proposes a multi-feature fusion method to identify the emotion of images and the high-level semantic information of objects. This method improves the accuracy of the experimental results and makes the selection of experimental images more scientific.

The invention discloses a seismic facies pickup method. According to the method, a multi-task deep convolutional neural network model is constructed; a weighted classification loss function is defined; a classification and regression combined loss function is designed; training, verifying and testing are carried out on the model by utilizing an acquired large-scale three-directional station waveform data set of a station network in a certain region, and the seismic phase is detected, recognized and accurately picked up when arriving at time, then a transfer learning and data enhancement methodis adopted, the model is applied to training, verification and testing of a small-scale data set of a transformer network in a target area, and seismic phase picking-up in the target area is achieved. According to the method, the multi-task convolutional neural network is utilized to simultaneously realize detection identification and time-of-arrival estimation of the seismic facies, accurate seismic facies pickup of a target area with only a small data set is realized by utilizing transfer learning, and the problem of seismic facies pickup is solved.

The invention discloses an urban road scene semantic segmentation method based on deep learning. The method comprises the following steps: 1) collecting an image at the front end of a vehicle; 2) input data expansion of the annotated image and the original image: randomly cutting, splicing or adding different types of noise to the image, transforming the image through an image affine matrix, and finally maintaining the original resolution of the image through filling, cutting and other transformation to obtain a data set; 3) training a network by using the image after data expansion and the annotation image, wherein the residual U-net network comprises a down-sampling part, a bridge part, an up-sampling part and a classification part; and 4) modifying the time interval T of the acquisitionmodule, inputting subsequently obtained images into the trained deep learning model, outputting predicted semantic segmentation images, and returning different gray levels in the images to the processor. A small data set is used, too fast gradient descent can be prevented, and it can be ensured that an over-fitting problem does not occur during training.

The invention discloses a migration convolution neural network method for face expression recognition, which comprises the following steps: S1, acquiring a face expression image data set, dividing thedata set into a training set, a verification set and a test set; 2, cascade that migration network and the convolution neural network to construct a migration convolution neural network model; The input of MCN model is facial expression image data, and the output is facial expression category. 3, train that migration convolution neural network model by using a training set, and optimize the trained migration convolution neural network model by using a verification set; S4, the facial expression recognition accuracy of the test set is tested by the optimized migration convolution neural network model. The invention utilizes the migration convolution neural network to recognize the facial expression, and solves the problems that the small data set cannot be converged and over-fitted on a large number of networks.

The invention discloses a Hadoop framework-based short-term load prediction method for a distributed BP (Back Propagation) neural network. The method specifically comprises the steps of obtaining an initial load data set; dividing the load data set into small data sets and storing the small data sets in data nodes of a distributed file system; initializing BP neural network parameters and uploading a parameter set into the distributed file system; training the BP neural network according to a current load sample, and obtaining correction values of a weight and a threshold of the BP neural network in the current data set; performing statistics on sum of weight and threshold parameters of all layers and between the layers of the network according to a key value of a key value pair; judging whether the convergence precision or the maximum iterative frequency is reached or not in a current iterative task, and if yes, establishing a distributed BP neural network model, or otherwise, performing correction of the weight and threshold parameters of the network; and inputting prediction day data and obtaining load power data of a prediction day. According to the method, the load prediction speed is increased and the requirements of load prediction precision are met.

In various embodiments, methods and systems can provide highly efficient order processing suitable for high volume order processing businesses. A batch of orders can be preprocessed to create smaller data sets that group orders for the same item together. The smaller data sets can then be analyzed and processed in parallel without causing database locks. Thus, the time required to complete the order fulfillment cycle may be reduced, which may in turn, increase one or more of operation efficiency, customer satisfaction, sales volume and profits.

The invention discloses a parallelization method of convolutional neural networks in a fuzzy region under a big-data environment. The parallelization method comprises the following steps: firstly, constructing the convolutional neural networks in the fuzzy region, putting a given target assumption region and object identification into the same network, carrying out convolutional calculation, and updating the weight of the whole network in a training process; and secondly, dividing an input log data set into a plurality of small data sets, introducing multiple workflows to pass through the convolutional neural networks in the fuzzy region in parallel for convolution and pooling, and independently training each small data set by virtue of gradient descent. By virtue of the parallelization method, a network structure and parameters are optimized, and relatively good analysis performance and precision are realized; furthermore, the number of FR-CNN obfuscation layers is adjusted aiming at different log data sets, so that the extracted features can well reflect the characters of oil-gas reservoirs, and the fuzzification problem of the log data can be solved; and the parallel training and execution of FR-CNN are carried out by virtue of multiple GPUs, so that the efficiency of the FR-CNN is improved.

The invention relates to a face recognition method based on the convolutional neural network. The problems that an existing face recognition method for a small data set is low in accuracy and poor inself-adaptation and needs manual parameter regulation are solved. The method comprises the steps of 1, dividing face images needing to be recognized into training samples, testing samples and verification samples, and reading data to be trained; 2, performing normalization processing; 3, constructing a network structure of the convolutional neural network; 4, placing the training samples into thenetwork structure constructed in the step 3 for training, wherein the training process includes forward broadcasting and reverse broadcasting; 5, saving model parameters trained in the step 4; 6, adopting the trained model parameters for detecting the testing samples obtained after normalization processing in the step 2, and obtaining a model detection result. The method is used for the field of pattern recognition.

The invention discloses a small data set finger vein recognition method based on a neural network. The method comprises the following steps: constructing and preprocessing a data set; carrying out data expansion on all types of images by utilizing a data enhancement algorithm; utilizing a convolutional neural network to construct a model suitable for data; training the designed model; acquiring atarget image by using acquisition equipment for preprocessing, and outputting a finger vein feature vector of the preprocessed image through model output; and comparing with a stored finger vein feature vector template to determine the category of the input finger vein. According to the method, through combination of a neural network and data enhancement, features with robustness are extracted; and through Euclidean distance comparison between different image output feature vectors, different types of finger vein images are distinguished according to the size of the distance, so that finger vein recognition can obtain high accuracy on the premise of a small data set.

Surrogate hashing is described, including running a hashing algorithm against a portion of a file to generate a hash value, determining whether the hash value is substantially similar to a stored hash value associated with another portion of another file, the portion and the another portion being standardized, wherein at least one of the portion and the another portion is an extremely small dataset, and identifying a location of the another file if the hash value is substantially similar to the stored hash value associated with the another portion of the another file.

The invention discloses a product surface defect detection method, electronic equipment and a readable storage medium. The detection method comprises the following steps: S1, acquiring an image of a to-be-detected product; S2, inputting the image of the to-be-tested product into a first processing module and a second processing module for parallel operation, and if the first processing module andthe second processing module both determine that the to-be-detected product is a good product, determining that the to-be-detected product is a good product, or if at least one of the first processingmodule and the second processing module determines that the to-be-detected product is a suspected defective product, entering the step S3; and S3, rechecking the suspected defective product. According to the detection method provided by the invention, the detection efficiency is improved, the probability of misjudging defective products as good products is reduced, and meanwhile, the problems ofdichotomy of defective products and good products under a small data set and accurate positioning of defect points are further solved.

The invention discloses a large-scale malicious domain detection system and method based on self-feedback learning and relates to the technical field of computer network security. For the deficiency of an existing detection technology on mass data processing and detection model updating, a malicious domain real-time detection system applicable to large-scale data is designed and realized. A methodof extracting a small data set for verification and updating is provided innovatively. The online learning efficiency is improved. Core algorithms comprise an algorithm of detecting malicious domainsbased on a support vector machine (SVM) in mass real-time domain detection, an online learning algorithm fSVM based on the self-feedback learning and an automatic calibration algorithm. Through theoretical demonstration and experimental verification, according to the algorithms provided by the invention, when the newly-presented malicious domains are copied with, the response can be carried out timely, and the excellent operation efficiency is achieved. According to the system and the method, the further analysis of the detected malicious domains is also realized. The system and the method play an enlightening role in perceiving malicious domain related threat intelligence.

The invention provides a comparison detector and a building method as well as a cervical cancer cell detection method. The method applying the comparison detector to detect cervical cancer cell comprises the following steps that firstly, a class reference sample is selected by applying a t-SNE (t-distributedstochastic neighbor embedding) visualization method, and then the features of the class reference sample and a target detection image are extracted by applying a feature extraction network which is composed of a basic convolution network and a pyramid convolution network in the comparison detector; the extracted feature of the reference sample is processed by applying a reference sample processing module so as to obtain prototype expression of each class; a recommended region is obtained by applying a region recommending module, and features are obtained from a corresponding feature pyramid; and finally, the class of the recommended region is obtained by comparing the pyramid feature of the recommended region and the prototype expression of class, and fine tuning is conducted on the rectangular frame by utilizing the features of a candidate region. The method builds a target detection network which detects targets on small data sets, and can relieve the overfitting problem of the target detection network on the small data sets.

A system and method for coordinating online and offline schedules. The system allows a user to streamline calendar information from various online and offline sources into one distribution channel. The system and method provides an online interface with numerous outlets to other calendar systems, without the need for a license for each additional calendar system. In general, the system is capable of maintaining varying capacity and is scalable. As such, the system is able to manage huge variation and user populations in a small data set. Information can be standardized to be entered into the system only once and subsequently be distributed to various types of programs and platforms.

The invention relates to a pressing method for using three-dimensional seismic data to process linear noise of the three-dimensional seismic data, belonging to the oil geophysical exploration field. The concrete steps are that: high density seismic data is collected and recorded; the data is divided into data groups with smallest data sets; rank ordering is carried out on all paths of each data group according to the shot-geophone distance and the azimuth, thus forming a three dimensional denoising smallest data set; the linear noise speed is analyzed according to the formed frequency wave-number spectrum (FK spectrum) or the original seismic data so as to eliminate the linear noise; pressing on the existing linear noise is repeated until the noise is eliminated. The pressing method is applicable to different seismic data collected by observation systems, in particular to non-cross banding and irregular bunchy orthogonality observation system, thus pressing the linear noise of the original data well.

The invention discloses a convolutional neural network automatic segmentation method and system for a mammary molybdenum target data set, which can obviously reduce model parameters and improve the practicability while ensuring the precision of a deep learning model on a mammary molybdenum target small data set. The method comprises the following steps: pre-training a convolutional neural big network on a mammary molybdenum target big data set; performing model compression on the trained convolutional neural large network by adopting attention transfer and knowledge distillation methods to obtain a convolutional neural small network; and carrying out fine tuning on the convolutional neural small network on the mammary molybdenum target small data set.

The invention discloses a plant disease and insect pest identification method and system based on deep learning. The plant disease and insect pest identification method comprises the steps: (1) receiving an original training image which comprises various types of plant disease and insect pest image data; (2) preprocessing based on an image preprocessing mode selected by a user for the training image; (3) inputting the preprocessed image into a deep learning model according to disease and pest categories, wherein the deep learning model comprises a convolution layer, a residual attention mechanism model and a full connection layer which are connected in sequence; and visualizing the output of the residual attention mechanism model to serve as a new training image, and returning to the step(2); (4) training a deep learning model based on all the training images; and (5) identifying diseases and insect pests. According to the invention, targeted enhancement and expansion processing are carried out on the small data set by using a method of combining data enhancement and an attention mechanism, so that the plant disease and insect pest identification accuracy can be effectively improved.

The invention provides a natural scene horizontal character detection method based on a deep convolutional neural network. According to the method, deep optimization is carried out on the basis of a TextBoxes network model, a new text prediction convolution group is added, and the network depth is expanded, so that the feature learning of the network for a small data set is more sufficient, and the feature information of a plurality of convolution layers is fully utilized to carry out fusion learning under certain model complexity. After feature learning is performed on original picture data through convolution layers with different receptive fields, a text prediction layer is utilized to return to the position of a textbox and predict a text category. According to the detection method, the influence of factors such as background complexity of a natural scene and insufficient features of a small data set on character detection is effectively solved. Experimental verification is carriedout under a Caffe platform, and results show that the model can effectively improve the recall rate and comprehensive evaluation indexes of natural scene level character detection under a small dataset.

An iterative approach to vector median filtering wherein the resulting median vector need not be a member of the original data set. The iterative vector median filtering allows for fast convergence for complex computations and an output which is approximate to the mean, particularly for small data sets. In addition, a method and system for registering and matching 2.5 normal maps is provided. Registration of two maps is performed by optimally aligning their normals through 2-D warping in the image plane in conjunction with a 3-D rotation of the normals. Once aligned, the average dot-product serves as a matching metric for automatic target recognition (ATR).

A character that includes a character element is represented and / or displayed by receiving a character element code that specifies the character element and skeleton point data that represent a position of the character element, providing a character element generating function corresponding to the character element code, and generating the shape of the character element using the character element generating function with the skeleton point data as arguments therefor.

The present invention provides a logging facies identification and analysis method based on fuzzy depth learning in a big data environment. The method comprises: constructing a fuzzy area convolution nerve network, putting a given target hypothesis area and target identification into the same network to share the convolution calculation, and employing a training process to update the whole network weight; dividing input logging data to a plurality of small data sets, performing convolution and pooling operation steps of each small data set through the fuzzy area convolution nerve network; and employing classified features to construct a logging facies-sedimentary facies knowledge base which is configured to establish a corresponding knowledge base including the sedimentary facies, the sedimentary subfacies and the sedimentary microfacies based on the unambiguous logging data and the sedimentary facies fusion method to support the association analysis of the sedimentary facies and the logging facies so as to establish the logging facies-sedimentary facies knowledge base and determine the corresponding relation of the current logging data and the sedimentary facies.