72 results about "Fisher vector" patented technology

In an image classification method, a feature vector representing an input image is generated by unsupervised operations including extracting local descriptors from patches distributed over the input image, and a classification value for the input image is generated by applying a neural network (NN) to the feature vector. Extracting the feature vector may include encoding the local descriptors extracted from each patch using a generative model, such as Fisher vector encoding, aggregating the encoded local descriptors to form a vector, projecting the vector into a space of lower dimensionality, for example using Principal Component Analysis (PCA), and normalizing the feature vector of lower dimensionality to produce the feature vector representing the input image. A set of mid-level features representing the input image may be generated as the output of an intermediate layer of the NN.

The invention relates to an automatic recognition method of the standard section of the fetus face of an ultrasound image. The method includes the following steps that the original ultrasound image is preprocessed; features of the preprocessed ultrasound image are extracted through RootSIFI; the extracted features are transformed through a Gaussian mixture model, Fisher vector encoding is performed on the transformed Gaussian mixture model, and values of a vector encoding histogram are obtained; normalization is performed on the histogram to obtain feature vectors; the standard sector is learned and recognized according to the feature vectors through an SVM classifier based on a random dual coordination mechanism. The invention further relates to an automatic recognition system of the standard section of the fetus face of the ultrasound image. Through the method and system, the accuracy of the diagnosis result obtained by a doctor can be improved, workloads can be reduced, and diagnosis time can be shortened.

The invention discloses an improved method for learning discriminative segments in fine-grained identification. The method comprises the following steps of: extracting segments with discrimination properties in an original image: obtaining a feature map by the original image through a convolutional pooling layer in a convolutional neural network, considering a vector of each space fixed position as a detector of a corresponding position segment in the original image, assuming that a detector, a discriminative region of which has highest response, in the original image is learnt, carrying out convolutional operation on the detector and the feature map so as to obtain a new response map, and selecting a position with a maximum value in the new response map so as to obtain the segments with discriminative properties; and learning features of the segments with the discriminative properties and carrying out classification by using the features: obtaining a local saliency map according to the segments with the discriminative properties, and encoding the local saliency map by using a space weighted Fisher vector. The method is capable of learning discriminative features more suitable forfine-grained identification tasks, and decreasing the interferences of background information in the discriminative segments so as to improve the classification precision.

The invention discloses a cross-media Chinese herbal medicine plant image searching method based on deep learning. The method comprises the following steps that 1, on the basis of OCR, text structural processing is performed, and plant classification description characters are extracted from books such as 'plant taxonomy'; 2, a Chinese word segmentation tool is used for performing preprocessing including word segmentation and stop word removal on all plant classification description characters; 3, a word2vec algorithm is used for generating word vectors according to a description text; 4, Fisher Vector is used for encoding the description text; 5, a convolutional neural network is used for performing training on an image set, and the network structure is made to converge to the best state; 6, the full last but one connection layer output of the convolutional neural network is extracted as an image feature vector; 7, textual features and image features are fused; 8, a linear kernel SVM classifier is used for performing model training; 9, during user retrieval, images can be input, plant texts can be described, and then a final image retrieval result is obtained through the steps 4, 6, 7 and 8.

A cross-media search method using a VGG convolutional neural network (VGG net) to extract image features. The 4096-dimensional feature of a seventh fully-connected layer (fc7) in the VGG net, after processing by a ReLU activation function, serves as image features. A Fisher Vector based on Word2vec is utilized to extract text features. Semantic matching is performed on heterogeneous images and the text features by means of logistic regression. A correlation between the two heterogeneous features, which are images and text, is found by means of semantic matching based on logistic regression, and thus cross-media search is achieved. The feature extraction method can effectively indicate deep semantics of image and text, improve cross-media search accuracy, and thus greatly improve the cross-media search effect.

The invention discloses a spatial Fisher vector based image classification method which mainly solves the problems of 'foreground and background visual ambiguities' of images and lack of feature point space distribution information in image description vectors in the prior art. The method includes the implementation steps: (1) extracting 'scale-invariant feature transformation' feature points of all images; (2) in a feature point space of images in a training set, clustering the feature points by the aid of a Gaussian hybrid clustering algorithm to obtain a code book; (3) utilizing gradient vectors and Cholesky components for generating a Fisher vector of each image; (4) performing 2X2 space region division on each image, and making statistics on number of feature points and coordinates of each cell block; (5) utilizing a Fisher vector of each cell block for stitching to generate a space Fisher vector of each image; and (6) utilizing a support vector mechanism for constructing a classification hyperplane to achieve image classification. The method has the advantages that image information can be described accurately, accuracy rate of image classification is improved, and the method can be used for large-scale image classification and retrieval system construction.

The invention discloses a figure behavior identification method based on random projection and Fisher vectors. The method employs a random projection theorem method to replace a principal component analysis method for characteristic dimension reduction, for the purpose of solving the problems of large time consumption, indeterminate reservation of principle components and the like. A random projection theorem indicates that through a compression measurement matrix, original signals with a sparse property can be projected to a certain low-dimension subspace, and the point distance between a vector after mapping and an original high-dimension characteristic vector maintains basically unchanged, i.e., data distortion is not generated in a whole compression process. Besides, different from hard division of a BoW model, the method provided by the invention employs a GMM-Fisher vector hybrid model for soft division of locus characteristic vectors, is integrated with the characteristics of a Fisher nucleus generation mode and a discrimination mode, can calculate the occurrence frequency of each characteristic descriptor, can also describe the probability distribution conditions of these characteristic descriptors in the perspective of statistics, enriches characteristic expression of behavior motion and also improves the behavior identification efficiency.

The invention provides an iris image classification method. The method comprises steps that a sample iris image is processed at a construction stage of an iris texture primitive, and the iris texture primitive is acquired; an iris classifier is constructed based on the iris texture primitive and a support vector machine at a construction stage of the iris classifier; target iris images are classified through utilizing the iris classifier at a discrimination stage. The method is advantaged in that iris image classification can be effectively accomplished, high effectiveness and security of iris identification are improved; characteristics acquired through depth learning are utilized to replace characteristics acquired through traditional manual design to extract the iris texture primitive, high precision, high robustness and high reliability are realized, and the method is suitable for iris image classification of multiple application demands such as in vivo detection, race recognition and sex identification; a system safety problem and a large-scale data retrieval problem existing in an iris system productization process are effectively solved.

The invention discloses an RGB-D object identification method which comprises the following steps of acquiring a gray scale image which is generated by a color image, and a surface normal vector thatis generated by a depth image, using the color image, the gray scale image, the depth image and the surface normal vector as multi-data mode information; respectively extracting high-layer characteristics in the color image, the gray scale image and the surface normal vector through a convolutional-recurrent neural network; extracting the high-layer characteristic of the depth image by means of aconvolutional-Fisher vector-recurrent neural network; and performing characteristic fusion on the plurality of high-layer characteristics, obtaining a total characteristic of the object, and inputtingthe total characteristic of the object into a characteristic classifier for realizing an object identification task. According to the RGB-D object identification method, a plurality of data modes arecombined; more accurate RGB-D object characteristics are extracted; and furthermore object identification accuracy is improved.

The invention discloses a model identification method based on video Fisher vector descriptors. The method comprises the following steps: first of all, tacking a vehicle for training, tracking an image of each model in a video, and extracting an SIFT characteristic of each frame of the image of the model; then, performing Fisher vector coding calculation on all the SIFT characteristics of the model image; then performing PCA dimension reduction on obtained Fisher vector descriptors; then performing binarization on the descriptors after the dimension reduction to obtain the video Fisher vector descriptors of the model; performing SVM training on all the obtained descriptors to obtain an identification system with N model types; and for a vehicle video for testing, extracting the video Fisher vector descriptors of the vehicle video, introducing the video Fisher vector descriptors into a well trained SVM classifier for the testing, and identifying the model of the test vehicle video.

The invention discloses an automatic detection method of corn tassel traits. The method comprises the following steps: performing target detection on an acquired farmland corn bottom view image at first to generate a tassel candidate frame to acquire a tassel potential area; subsequently, performing feature description and target detection on the tassel by utilizing a multi-view image characteristic and a Fisher vector coding method so as to confirm an affiliated area of the tassel; further completing the segmentation on tassel fine forms by utilizing semantic segmentation based on a detection result; finally establishing a mapping relation of the image characteristic and seven biomasses, such as length trait, width trait, perimeter trait, diameter trait, tassel color trait, branch quantity trait and total tassel quantity trait which have the physical significance. According to the method, the growth state of the corn tassel can be continuously monitored in real time, so that the detection result is high in correction rate, and the method has the great significance in corn reproductive growth research, corn genomics and genetics research and yield estimation.

The invention discloses a foundation cloud picture classification method based on a heterogeneous feature fusion network. The foundation cloud picture classification method comprises the following steps: (1) preprocessing a plurality of foundation cloud pictures with noise; (2) features of the processed foundation cloud images are extracted respectively, and a manual feature extraction method is combined with Fisher vector coding to obtain a feature vector corresponding to each foundation cloud image; (3) inputting the output of the step (2) into a four-layer full connection layer network, wherein the output of the network is recorded as fc; and (4) after the training set is amplified, training a convolutional neural network model, fusing fc with the deep semantic feature fg obtained by the last pooling layer, and obtaining a classification probability corresponding to each class through a full connection layer. According to the method, the generalization ability of a foundation cloudpicture classification and recognition task can be remarkably improved, the robustness of the model is high, visual information is combined from multiple angles, the cloud shape can be accurately positioned even if noise is manually added, and a good recognition result is obtained.

The invention relates to the technical field of medical lesion image recognition, in particular to a medical lesion image feature expression method based on region division and a Fisher vector. The method uses an expanded lesion region as a region of interest (ROI), divides the ROI into N sub regions according to the gray values of all pixels in the ROI; in each sub region, extracts small image blocks as local feature descriptors; aggregates the local feature set of each sub region into a vector by using a Fisher vector (FV) algorithm, and then connects the obtained N vectors end to end to obtain the feature expression of a medical lesion image. The feature expression method of the present invention utilizes the region information and space position information around the lesion, and uses the FV algorithm which is more effective than a conventional bag-of-word model to make the constructed feature expression more discriminative, thereby contributing to improving the accuracy of clinical adjuvant diagnosis.

The invention provides a mobile visual search framework based on a CRBM and Fisher network and relates to the image retrieval of a mobile terminal. The method comprises the following steps of 1) constructing and training a continuous restricted Boltzmann machine network; 2) constructing and training a Fisher layer network. According to the invention, the sub-space feature information of the localfeature nature of non-Gaussian distribution is found by adopting the nonlinear dimensionality reduction algorithm CRBM in the aggregation global compact binary feature algorithm. Meanwhile, more efficient global features are obtained through adopting a fisher-based network structure aggregation fisher vector. A compact self-adaptive feature is obtained by adopting the scalar quantization algorithmand the bit self-adaptation algorithm. Therefore, the length of the transmitted image feature information can be selected according to different self-adaptive selections of the network bandwidth of amobile terminal. During the retrieval stage, global features are roughly matched to obtain a candidate set, and the accurate matching of the geometric consistency test is conducted by adopting localfeatures. Therefore, the mobile visual search framework can be adapted to large-scale image retrieval tasks.

The invention discloses a compact visual descriptor deep neural network generation model in visual retrieval, and relates to image retrieval. The model is implemented by the following steps of: constructing a Fisher layer network; constructing a grouping and secondary classification module; training a loss function on the basis of a maximum boundary condition; for image library images and a queryimage, firstly extracting local features of the images, carrying out aggregation and binary embedding on the local features of the images by using a trained network structure so as to obtain binary codes of the images, carrying out matching in an image library according to the binary code of the query image, returning images with high similarities as a matched candidate set, carrying out geometricconsistency inspection on the candidate set by using the local features so as to carry out accurate matching, and returning a final query result. According to the model, the flexible Fisher network is used to aggregate local features of images so as to generate more efficient global feature Fisher vectors, and the grouping and secondary classification module is used to carry out binary coding onthe Fisher vectors so as to obtain compact global binary features.