45 results about "Attribute learning" patented technology

The invention discloses a pedestrian recognition method based on combination of depth learning and property learning. According to the invention, a convolution neural network containing five implicit strata is constructed. Network training is performed by an anti-convolution method and a concept of property learning is combined. Preferred features obtained from the convolution neural network are input to property classifiers, so that the posterior probability of the property of a sample is obtained. Then by combining with a property class mapping relation, the posterior probability of the class is obtained, so that the class of the sample can be judged. The method is good in detection recognition performance and intrinsic features of an image can be extracted. Besides, since the property has better semantic expression performance than low-stratum features and due to the insensitivity to light and view angles, the algorithm has a good recognition effect.

The invention relates to the technical field of mode recognition, and in particular to an attribute learning and Riemannian manifold combined pedestrian re-recognition method. According to the method,a model with stronger generalization ability is disclosed through researching representations and semantic attributes of pedestrians; and through importing pedestrian attribute labels, using a convolutional neural network model and using multi-target loss functions, the model not only needs to predict pedestrian IDs, but also needs to predict various correct pedestrian attributes. Pedestrian pictures are converted into output tensors of a convolution layer through the trained convolutional neural network, so as to respectively calculate a covariance descriptor of each pedestrian picture. Thecovariance descriptors are utilized to carry out feature fusion so as to eliminate feature redundancy, so that measurements of Riemannian manifolds where the descriptors are located are researched andthen more correct similarity calculation is realized.

The invention discloses a road surface garbage intelligent identification and sensing implementation device. The device comprises a camera, a vehicle-mounted industrial personal computer, a vehicle-mounted direct current power supply, and a controller. The camera is used for acquiring pavement garbage images in real time. The vehicle-mounted industrial personal computer identifies an image transmitted by the camera and transmits identification information to the controller through the CAN interface, and the controller comprises a chassis integrated controller and an upper integrated controllerand is used for receiving an instruction of the industrial personal computer, outputting a digital I/O signal to a corresponding actuator and controlling the actuator to act. The invention also discloses a road surface garbage intelligent identification and perception method, which can realize road characteristic information detection, identification and judgment, road surface garbage coverage rate detection and identification and road surface garbage attribute learning and classification, and forms operation characteristic learning and identification big data.

The invention provides a face super-resolution reconstruction system based on combined multi-task learning. The face super-resolution reconstruction system comprises an acquisition module, a first extraction module, a reconstruction module, a second extraction module and a training module. The method comprises the steps of obtaining shared representation of face features among related tasks through a joint training method for face multi-attribute learning tasks; then, demonstrating the feasibility of the perception loss in the aspect of improving the reconstruction effect of the human face semantic information; and finally, enhancing the face attribute data set, screening out data without related attribute tags, re-extracting the attributes of the feature points by using a face key point detection algorithm, and carrying out combined multi-task learning on the basis to generate a super-resolution result with a more real visual perception effect.

The invention provides an image analysis method based on a machine learning algorithm. Main contents comprise attribute learning, scale-based attribute metering and consistent visualization in a cluster mode. Attribute characteristics of an object of an image are provided according to depth learning of a data set carried out by a classifier. A monitoring learning method is utilized, picture notes are learned by a training machine, an object attribute classifier is established, test performance standards are improved to a maximum degree, machine classification is carried out, operation to a bigger database is further carried out, and the classifier is made to be an effective tool for picture analysis. The method is advantaged in that two novel vision algorithms are utilized, relevant attribute clusters are detected through image clustering, according to the identified attributed clusters, pictures and object clusters are further identified, the object attributes are provided according to machine learning, and the method further contributes to new design in the industrial field and innovation solution schemes in the engineering field.

The invention relates to a defective needle flaw detection method of a circular knitting machine. With a machine learning method, detection on a defective needle flaw of the fabric can be realized; pretreatment of a flaw image is carried out by using an adaptive image enhancement technique and an area of interest is segmented; and feature information of a fabric flaw is extracted effectively according to a multi-scale pyramid method with consistent phases. In terms of fabric flaw classification and identification, a fabric defective needle flaw is detected successfully by using hypergraph regularization attribute learning. Therefore, the fabric flaw detection speed and efficiency can be improved substantially.

The invention discloses a pedestrian attribute recognition system and method based on multi-layer feature learning, the system comprises a feature bottom-to-top extraction module, a bottom-to-top feature fusion module, a feature prediction module, a multi-layer prediction fusion module and a test module, and the method comprises the following specific steps: processing pictures layer by layer frombottom to top to obtain multi-layer features; fusing the features of the adjacent layers layer by layer from top to bottom, compressing the channel by the feature map obtained by the higher layer, carrying out feature fusion and channel dimension reduction on the compressed channel and the feature map sampled by the upper layer, and outputting the feature of the current layer; obtaining preliminary prediction results of different levels through a maximum pooling layer and a full connection layer according to the fused features and the extracted uppermost features; overlapping the preliminaryprediction results of different levels, and correspondingly endowing each attribute predicted by each level with a weight value to obtain a final prediction result; and extracting a prediction resultcorresponding to the picture, and calculating a result of each index. According to the method, a group of specific weights are learned for each attribute according to the predicted values obtained bythe fused features, so that each attribute can better utilize multi-layer features to obtain a better recognition effect.

The invention provides an indoor map semantic identification method based on WiFi channel state information and belongs to the technical field of internet of things. The method comprises the steps that a signal transmitting device and a signal receiving device are arranged in a room needing semantic identification; CSI (channel state information) data are collected; the collected data are processed and collected through a PEM algorithm, and the data are recorded; the number of people in the indoor environment can be sensed in real time, the data are recorded, and the size of a time block is calculated; an attribute learning type decision tree classification is carried out, and a decision tree classifier is generated; a decision tree classification result is used as a room identification result to achieve a room semantic identifier. The method in the invention can effectively solve the problems that in the prior art, the application range is narrow, a room semantic identifier cannot bechanged in real time, the equipment energy consumption is high, and secrets are easy to be leaked.

The invention discloses a vehicle re-identification method based on multi-task joint discriminant learning. According to the method, a plurality of tasks are jointly learned through a multi-branch network to obtain fine-grained discriminative features of a vehicle. A network obtains a network output feature vector through two branches of attribute learning and ID learning, meanwhile, a metric learning task and an ID learning task are used for constraining the feature vector, and more robust features are obtained through joint learning of the four tasks, wherein the ID learning uses an ArcFL loss function different from other methods, and the metric learning uses a Trihard loss function different from other methods. Through the proposal of an innovative network structure and the improvementof a loss function, the precision of vehicle re-identification and retrieval is significantly improved. The method is realized based on a large vehicle data set of a road monitoring scene, and can beeffectively applied to a vehicle search task.

The invention discloses an attribute learning-based finger vein recognition method and device. The method comprises the following steps of: obtaining a finger vein image library and carrying out dataaugmentation to form a training set; and aiming at characteristics of finger veins, designing interpretable attributes, and respectively establishing corresponding attribute learning modules for the interpretable attributes in the training set so as to obtain values of the interpretable attributes. In order to furthermore improve feature discrimination performance, a discriminative assistance branch-based convolutional neural network is used as a discriminative attribute learning model, and a new Inter-intra loss function is designed, so that the discrimination performance of the discriminative attribute learning model is strengthened; and to-be-recognized images are obtained, and attribute features of the to-be-recognized images are determined according to all the attribute learning models so as to carry out finger vein recognition. Compared with traditional features, the method provided by the invention is capable of improving the interpretability, recognition precision and recognition efficiency of existing features.

Described is an audio classification system for classifying audio signals. In operation, the system extracts salient patches from an intensity spectrogram of an audio signal. Thereafter, multi-scale global average pooling (GAP) features are extracted for all salient patches. The GAP features are clustered, with each cluster becoming a key attribute. A test audio signal can then be mapped onto a histogram of key attributes. Based on the histogram, the test audio signal can then be classified as a sound class, allowing for operation of a device based on the classification of the sound class.

The invention provides a human face aesthetic prediction method based on the combination of the biologically inspired computation and the deep attribute learning, and relates to a human face aestheticprediction method. According to the method, firstly, the eye movement information of a user is acquired when the user observes a human face image through an eye view monitoring system. After that, the aesthetic space area when the user looks at a human face is further extracted. The aesthetic space area is divided into a plurality of characteristic spaces through clustering. After that, a human face aesthetic detector is trained through the supervised learning method of the convolution neural network, wherein the obtained human face aesthetic detector can be used for preprocessing a frontal face image and then obtaining the aesthetic level of the face image. The aesthetic space area when the user observes the image is extracted by collecting the middle-layer attribute characteristic information of the image. The aesthetic characteristic area for determining the face aesthetic level is fully verified through an obtained human face aesthetic model. Meanwhile, compared with other human face aesthetic evaluation methods, the accuracy of the method is greatly improved.

The invention relates to a weak supervision target detection method based on image attribute representation learning, and belongs to the technical field of image processing. The method sequentially comprises the steps of label text description data processing, image feature extraction and target suggestion box extraction, label text feature construction, text and image feature fusion-based pseudo ground-truth mining, image attribute learning and prediction module, target classification and target suggestion box regression. According to the method, the interpretability of target classification is improved through image attribute learning, the mined pseudo group-truth is more accurate by utilizing text feature and image feature fusion, and the detection capability of the weak supervision model is improved.

The invention discloses a small sample target recognition method based on hybrid attribute learning. The method comprises at first, semantic attributes are defined and auxiliary samples are obtained;the target training samples are inputted into the feature sub-network for obtaining hidden attributes, and the hidden attributes of the target training samples are outputted; the auxiliary sample is input to a feature subnetwork for obtaining semantic attributes, which is used to obtain the semantic attributes of the auxiliary sample; the extracted features are dictionary-learned to obtain the encoding of the features, wherein the encoded data including semantic attributes and hidden attributes are input into a multi-classifier for classification training to obtain a classification network, which is used for subsequent classification of the target. This method overcomes the shortcomings of the other methods which only have good interpretability and poor discriminating ability, or only havegood discriminating ability and poor interpreting ability, and synthesizes the advantages of semantic attributes and hidden attributes, and gives attention to both interpretability and discriminatingability.

The invention discloses a data knowledge dual-drive modulation intelligent identification method. The method mainly solves the problems that an existing modulation identification method is low in classification accuracy under the low signal-to-noise ratio, depends on a large number of training samples, and high-order modulation modes are likely to be confused in the identification process. The method comprises the following steps of: acquiring spectrum data; constructing corresponding attribute vector tags according to different modulation modes; constructing and pre-training an attribute learning model according to the attribute labels of different modulation modes; constructing and pre-training a modulation mode identification visual model; a feature space conversion model is constructed, and a data knowledge dual-drive modulation mode intelligent identification framework is constructed in combination with a visual model and an attribute learning model; migrating parameters of the pre-training visual model and the pre-training attribute learning model, and retraining the conversion model; and judging whether network training is finished, and outputting a classification result. According to the invention, the identification accuracy under a low signal-to-noise ratio is obviously improved; and confusion between high-order modulation modes is reduced.

Techniques for generating an image are disclosed. In some embodiments, a received input image is transformed to generate an output image using a machine learning based framework that is trained on a constrained set of images. The generated output image comprises an attribute learned by the machine learning based framework from the set of images.

A transformer fault diagnosis method based on weighting and selective naive Bayes is characterized in that an attribute selection method based on x2 statistics removes a part of redundant attributes and constructs an attribute learning classifier better for a classification result, and comprises the following steps: 1) collecting historical fault data of a main transformer, including attribute data and fault types, discretizing conditional attribute data; wherein the fault type is a decision attribute, and dividing the data into a training set and a test set; 2) selecting an optimal reduction subset RAS by using an attribute selection method based on x2 statistics; 3) learning prior probability: calculating prior probabilities of all decision attributes and conditional probabilities of attributes in RAS by the training set, and respectively storing results into a CP table and a CPT table; 4) establishing a weight table of the attribute data by using a correlation probability method; calculating all weights of the attributes in the RAS table under different categories, and storing the weights in a weight table AW; using the test set to test model performance; and accessing model accuracy according to the actual category of the test data.

The invention discloses a pedestrian recognition method based on a local feature perception image-text cross-modal model and a model training method, and belongs to the technical field of mode recognition. The pedestrian recognition method comprises the following steps: acquiring image-text data of pedestrians, and inputting the image-text data of the pedestrian into a pre-trained local feature perception image-text cross-modal model for feature extraction, and outputting a pedestrian recognition result. The local feature perception image-text cross-modal model comprises a visual feature extraction module and a text feature extraction module, PCB local feature learning is introduced to visual feature extraction, a multi-branch convolution structure is introduced to text feature extraction, and image-text local features can be efficiently extracted without introducing semantic segmentation, attribute learning and the like. Cross-modal matching is carried out on three levels of shallow features, local features and global features, and image-text feature distribution is gradually pulled in. The method is simple in structure and high in accuracy, and application of the image-text cross-modal pedestrian retrieval field in actual scenes can be promoted.