Method and system for searching images by images based on deep learning

A deep learning and image technology, applied in the field of image search based on deep learning, can solve problems such as difficult control of high-level features, and achieve the effect of meeting the needs of fast analysis, fast search results, and reducing impact

Active Publication Date: 2015-06-03
武汉众智数字技术有限公司
6 Cites 51 Cited by

AI-Extracted Technical Summary

Problems solved by technology

The deep network described in the document "ImageNet Classification with Deep Convolutional Neural Networks" solves the problem of feature extraction to a certain extent, but because the high-l...
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Method used

Synthesize the classification feature that the first part produces, the image self-encoding feature that the second part produces and the self-defined feature that the third part produces, these features are carried out further ...
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Abstract

The invention relates to the technical field of image searching and provides a method for searching images by images based on deep learning. The method comprises the following steps: calculating image category features, and performing classification feature extraction on input images by using a trained deep convolutional neural network; calculating image coding features, and performing coding feature extraction on the input images by using a trained deep learning automatic coding algorithm; compacting mixed feature codes, integrating the classification features and the image own coding features, and coding the integrated features by a deep learning automatic coding algorithm; calculating image similarity according to the features, and ranking and outputting the image similarity. According to the method disclosed by the invention, advanced features are generated by the deep convolutional neural network, so the similarity of the results of searching images by images in image category is guaranteed; low-level image coding features are generated by using the automatic coding algorithm, so the similarity of the images in content is guaranteed; according to the mixed own coding feature method, the classification features and the image own coding features are further fused, so that the dimensionality is reduced, and the search result is carried out more quickly and more stably.

Application Domain

Biological neural network modelsSpecial data processing applications

Technology Topic

Feature codeImage based +3

Image

  • Method and system for searching images by images based on deep learning
  • Method and system for searching images by images based on deep learning

Examples

  • Experimental program(3)

Example Embodiment

[0024] Example 1:
[0025] Embodiment 1 of the present invention provides a system for searching images based on deep learning, such as figure 1 As shown, the image input platform 10, the integrated access gateway 20, the intelligent management server 30, and the intelligent analysis server 40 are included. The image input platform 10, the integrated access gateway 20, the intelligent management server 30 and the intelligent analysis server 40 are connected in sequence, specific:
[0026] The image input platform 10 is used for image entry, image transmission, image storage and image preprocessing; the integrated access gateway 20 is used for statistical access of the image input platform to the intelligent management server; the intelligent management The server 30 is used to manage and analyze resources; the intelligent analysis server 40 is a functional entity for image search and is composed of multiple image analysis units, each of which can independently complete the analysis of an image input platform.

Example Embodiment

[0027] Example 2:
[0028] Embodiment 2 of the present invention provides a method for searching images based on deep learning, which is characterized in that it includes:
[0029] In step 201, the image category features are calculated, and the trained deep convolutional neural network is used to extract the classification features of the input image;
[0030] In step 202, the image self-encoding feature is calculated, and the trained deep learning automatic encoding algorithm is used to extract the encoding feature from the input image;
[0031] In step 203, hybrid feature encoding compression, combining the classification features and image self-encoding features, and encoding these features through a deep learning automatic encoding algorithm;
[0032] In step 204, the image similarity is calculated according to the features, and the output is sorted.
[0033] This embodiment uses a deep convolutional neural network to generate high-level features to help analyze image categories to ensure that the results of image search are similar in image categories; and use automatic encoding algorithms to generate low-level image encoding features to ensure that the image is in content It is similar to human senses as much as possible; hybrid self-encoding feature method: further fusion of classification features and image self-encoding features, reducing the dimensionality and reducing the impact of redundant features on the retrieval results. Make the search results faster and more stable, while being able to meet the needs of rapid analysis.
[0034] In combination with this embodiment, there is a preferred solution, where the hybrid feature encoding compression also includes custom features, and the custom features include color features, shape features, and/or texture features, then step 203 is specifically executed It is: synthesize the classification features, image self-encoding features and custom features, and encode these features through a deep learning automatic encoding algorithm.
[0035] And, before the step 203, it also includes step 205, such as figure 2 As shown, specifically:
[0036] In step 205, the custom feature is calculated.
[0037] With reference to this embodiment, preferably, the calculation of image similarity according to features and the sorted output specifically include:
[0038] Calculate the geometric distance between the image input by the user and the hybrid encoding feature of each other image in the database, and sort the geometric distance from small to large, and output the sorting result.
[0039] With reference to this embodiment, preferably, the deep convolutional neural network is composed of a convolutional layer and a fully connected layer, and the network layer and the layer include the pooling method, the dropout method and/or the dropconnect method in deep learning.
[0040] With reference to this embodiment, preferably, the automatic coding algorithm for deep learning includes:
[0041] Any one of autoencoder, sparse autoencoder, stack autoencoder, and noise reduction autoencoder.
[0042] In combination with this embodiment, preferably, the comprehensive feature compression method is specifically:
[0043] Self-encoder, sparse self-encoder, stacked auto-encoder, noise reduction auto-encoder, any one of component analysis.
[0044] With reference to this embodiment, preferably, the distance between features in the calculation of image similarity is specifically:
[0045] Any one of Mahalanobis distance, Euclidean distance, and chessboard distance.

Example Embodiment

[0046] Example 3:
[0047] Embodiment 3 of the present invention combines actual cases to provide specific implementation methods for the implementation of Embodiment 1 and Embodiment 2. Specifically, it includes the five parts of computing image category features, computing image self-encoding features, computing custom features, hybrid feature encoding compression, and computing image similarity and sorting output as described in Embodiment 2.
[0048] Part 1: Calculate image category features
[0049] The algorithm for calculating image category features uses deep convolutional neural networks, such as the ``ImageNet Classification with Deep Convolutional Neural Networks'' algorithm described in the article. The network is composed of 5 convolutional layers and 3 fully connected layers. The image passes through the convolutional layer and The fully connected layer finally obtains the method of high-level image features, which are mainly used for image classification.
[0050] The training steps of deep convolutional neural network are as follows:
[0051] The deep convolutional network is trained using the ImgNet data training set, the training sample size is 1 million labeled images, and the classification category is 1000 categories. The network parameters and network structure used are the same as those in the paper ``ImageNet Classification with Deep Convolutional Neural Networks'' the same.
[0052] The implementation steps of deep convolutional neural network are as follows:
[0053] The image passes through the deep convolutional neural network, and extracts the 1000-dimensional node data of the third fully connected layer as the category feature.
[0054] Part 2: Calculate the image self-encoding feature
[0055] Pass the input image through 3-5 coding layers. Use any coding layer from the third layer to the fifth layer as the image self-encoding feature.
[0056] The training of the deep learning automatic coding algorithm uses 100,000 pictures. The category includes common unlabeled pictures of people, cars, things, etc., taking a 3-layer self-encoding network as an example, the network structure is that the input image is scaled to a size of 32*32, and the number of output nodes of the first-layer self-encoder is 500. The number of nodes in the second layer is 200, the third layer is 100, and the 100-dimensional coding feature output by the third layer is used as the similarity feature.
[0057] Part 3: Calculate other custom features
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

PUM

no PUM

Description & Claims & Application Information

We can also present the details of the Description, Claims and Application information to help users get a comprehensive understanding of the technical details of the patent, such as background art, summary of invention, brief description of drawings, description of embodiments, and other original content. On the other hand, users can also determine the specific scope of protection of the technology through the list of claims; as well as understand the changes in the life cycle of the technology with the presentation of the patent timeline. Login to view more.
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

Similar technology patents

Generating a movement signal of a part of the human or animal body

ActiveUS20180353140A1Reduce complexityReduce dimensionalityMagnetic measurementsDiagnostic recording/measuringReceiver coilVIT signals
Owner:SIEMENS HEALTHCARE GMBH

Naive Bayes classifier-based dermatosis image color feature extraction method

InactiveCN106557771AImprove accuracyReduce dimensionalityRecognition of medical/anatomical patternsColor modelDisease
Owner:UNIV OF ELECTRONIC SCI & TECH OF CHINA

Fraud node identification method based on graph embedded representation and recurrent neural network

PendingCN112395466AReduce labor costsReduce dimensionalityWeb data indexingOther databases indexingMachine learningNetwork model
Owner:SHANGHAI JIAO TONG UNIV

Micro-blog short text classification method based on lexical chain feature extension and LDA (latent Dirichlet allocation) model

ActiveCN108920482AImprove classification performanceReduce dimensionalityNatural language data processingSpecial data processing applicationsExtension methodLexicon
Owner:ZHEJIANG UNIV OF TECH

Power system dynamic safety assessment method based on relation exploration and regression tree

PendingCN110264116AReduce dimensionalityLighten the computational burdenCharacter and pattern recognitionResourcesAssessment methodsTime dynamics
Owner:CHINA THREE GORGES UNIV

Classification and recommendation of technical efficacy words

  • Reduce dimensionality
  • Reduce the impact

People also interested in

Feature extractionDeep learningConvolutional neural network
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic.

© 2023 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap