CN105426447A - Relevance feedback method based on transfinite learning machine - Google Patents

Abstract

The invention relates to a relevance feedback method based on a transfinite learning machine. The relevance feedback method comprises the following steps: inputting an inquiry image; retrieving the image to obtain a retrieval result, and marking a result by a user; respectively extracting an SIFT characteristic, a Colour characteristic and an LBP characteristic from the marked image; training three basis classifiers by utilizing the three kinds of characteristics; respectively putting the image in a retrieval image library in the three basis classifiers, voting according to a prediction result, and automatically marking each unmarked image; training and updating the classifiers again; classifying the image in the image library; and returning a result. The relevance feedback method disclosed by the invention is established on the basis of the transfinite learning machine; human-computer interaction is carried out by introducing the inquiry intention of human beings; learning data are enriched by effectively utilizing the unmarked image in the image library; therefore, the image feedback precision is greatly increased; furthermore, the processing speed is controlled well; expression of the image in a computer accords with understanding of human beings to image semantics well; and thus, the relevance feedback method has a good feedback effect.

Description

A kind of related feedback method based on the learning machine that transfinites

Technical field

The invention belongs to image processing field, relate to the related feedback method in image retrieval, be specifically related to a kind of related feedback method based on the learning machine that transfinites.

Background technology

Current social has stepped into the main huge data age based on multimedia information data, wherein outstanding with digital image information data again.Compared with other multimedia data, the content in view data is abundanter, expresses more directly perceived, inevitable and become the form that in daily life, Information Sharing is very main.In the face of increasing image information data, effectively excavate the dark voluminous amount of information hidden in view data, thus can in large-scale image data base fast and find out the image information of user's actual needs exactly.This trend has become the Main Topics of the association area such as computer vision and multimedia data information retrieval gradually.But as the huge spread also existed between the low-level image feature of image and the semanteme of high level, this content-based image data information retrieval still can not obtain satisfied result.Relevant feedback improves the effective way of CBIR effect, and one is comprehensively fed back and can improve retrieval precision greatly.But, data relevant feedback are absolutely necessary, but use artificial come free-hand mark mass data be an intolerable lengthy procedure.Due to the specific demand to relevant feedback, if markd and very rare data can be made full use of, and a large amount of unlabelled data, this is a kind of desirable mode utilizing data.

In recent years, being applied in different field of the learning machine that transfinites is widely used, but is seldom used to relevant feedback.Because the learning machine that transfinites has very good nicety of grading and processing time, and precision is high, speed is evaluate the vital factor of relevant feedback performance soon simultaneously.

The present invention proposes one based on the image retrieval of the learning machine relevant feedback scheme that transfinites with Gaussian kernel, to overcoming above-mentioned time and the restriction of precision, this method uses three Weak Classifier cascades, to form powerful sorter, be used for learning different features to extract a small amount of handmarking's data characteristics, then use it to carry out data untaggeds a large amount of in automatic terrestrial reference note image data base.Can see from experiment, adopt the method for Gaussian kernel, have higher nicety of grading, the processing time also declines to a great extent to some extent simultaneously.Meanwhile, it use the data of a small amount of handmarking as training data, and employ unlabelled data while combining the framework of coorinated training.Therefore, the characteristic of image retrieval how is effectively utilized to be an emphasis in image retrieval research field.

Summary of the invention

At these several years recently, along with the image data volume explosion type of digital form rapidly increases, this needs an efficient and effective method, allows user to be searched for by so a large amount of set.Therefore under such conditions, CBIR technology becomes more and more popular, and in this many decades, many systems are by this trend development.Therefore in the present invention, propose a related feedback method based on the learning machine that transfinites, the method can feature organically in combining image retrieval and relevant feedback, and applies these features to build feedback method.

Based on a related feedback method for the learning machine that transfinites, it is characterized in that comprising the following steps:

Step 1, input one secondary query image;

Step 2, retrieves image, obtains result for retrieval, allows user mark result;

Step 3, extracts SIFT feature respectively to labeled image, Color feature, and LBP feature;

Step 4, utilizes three kinds of features training, three fundamental classifier;

Step 5, puts into the image of retrieval picture library respectively in the middle of three base sorters, voting, carrying out automatic mark to each secondary unmarked picture according to predicting the outcome;

Step 6, re-training upgrades sorter;

Step 7, classifies to picture library picture;

Step 8, returns results.

Method of the present invention has the following advantages:

The present invention is based upon transfinites on the basis of learning machine, by introducing the query intention of the mankind, carry out man-machine interaction, effectively utilize unmarked picture library image to enrich learning data, the precision of image feedback can be made greatly to improve, and processing speed is well controlled, makes image expression in a computer more meet the understanding of the mankind to image, semantic, make the present invention have good feedback effects.

Accompanying drawing explanation

Fig. 1 is the overview flow chart inventing described method.

Fig. 2 is the process flow diagram of voting process in this method.

Fig. 3 be in this method automatic mark without mark data process flow diagram.

Embodiment

Below in conjunction with embodiment, the present invention is described further.

The flow process of the related feedback method based on the learning machine that transfinites of the present invention as shown in Figure 1, comprises the following steps:

Step 1, input one secondary query image;

Step 2, retrieves image, obtains result for retrieval, allows user mark result;

Step 2.1, in searching system, what user needed mark is the result of primary retrieval or the result after once feeding back.Picture after user is labeled is as data set L, and user does not have labeled picture as data set U, and now user does not have labeled data set to be whole picture library data set D-L.Positive example P and negative routine N, therefore L=P ∪ N is comprised in labeled data centralization.

Step 3, extracts SIFT feature respectively to labeled image, Color feature, and LBP feature;

Step 4, utilizes three kinds of features training, three fundamental classifier;

Step 4.1, by the enlightenment of coorinated training algorithm, respectively to having mark training set L and extracting three kinds of features without mark training set U, three base sorters are trained respectively by these three kinds of features, the otherness of data to set of classifiers can be strengthened like this, the uncertainty that effective reduction is brought by single data Expressive Features, strengthens prediction effect with this.

Step 5, puts into the image of retrieval picture library respectively in the middle of three base sorters, voting, carrying out automatic mark to each secondary unmarked picture according to predicting the outcome;

Step 5.1, puts into whole data untagged collection U the set of classifiers be made up of three base sorters and predicts, voting mechanism as shown in Figure 2, provides data untagged after the prediction of each base sorter and concentrates corresponding i ^thpredictive marker, predictive marker is designated as Y _in, predict that all predictive marker collection ψ drawn are according to three kinds of different characteristics:

$\mathrm{\Ψ}={\left[\begin{array}{ccc}{Y}_{11}& Y_{12}& Y_{13}\\ .& & .\\ .& & .\\ .& \mathrm{...}& .\\ Y_{M1}& Y_{M2}& Y_{M3}\end{array}\right]}_{M\×3}.$

Step 6, re-training upgrades sorter, automatic mark data untagged collection;

Step 6.1, after poll closing, obtains the L_temp of all data untaggeds with predictive marker result, carrys out re-training set of classifiers with L_temp, promote the classification capacity of sorter device group.After updated set of classifiers, then data untagged collection U is put into set of classifiers carry out prediction classification, as shown in Figure 3.

Step 6.2, as a result, obtains with positive example mark and two group data sets with negative example mark.Data set with positive example mark is denoted as P*, and the data set with negative example mark is denoted as N*.

Step 7, returns results.

Step 8, treats that user determines whether to carry out lower side feedback operation, if it is skip to step 2.1.

Claims (2)

1. based on a related feedback method for the learning machine that transfinites, it is characterized in that: the method comprises the following steps:

Step 1, input one secondary query image;

Step 2, retrieves image, obtains result for retrieval, allows user mark result;

Step 3, extracts SIFT feature respectively to labeled image, Color feature, and LBP feature;

Step 4, utilizes three kinds of features training, three fundamental classifier;

Step 5, puts into the image of retrieval picture library respectively in the middle of three base sorters, voting, carrying out automatic mark to each secondary unmarked picture according to predicting the outcome;

Step 6, re-training upgrades sorter;

Step 7, classifies to picture library picture;

Step 8, returns results.

2. a kind of related feedback method based on the learning machine that transfinites according to claim 1, is characterized in that: comprise the following steps:

Step 1, input one secondary query image;

Step 2, retrieves image, obtains result for retrieval, allows user mark result;

Step 2.1, in searching system, what user needed mark is the result of primary retrieval or the result after once feeding back; Picture after user is labeled is as data set L, and user does not have labeled picture as data set U, and now user does not have labeled data set to be whole picture library data set D-L; Positive example P and negative routine N, therefore L=P ∪ N is comprised in labeled data centralization;

Step 3, extracts SIFT feature respectively to labeled image, Color feature, and LBP feature;

Step 4, utilizes three kinds of features training, three fundamental classifier;

Step 4.1, by the enlightenment of coorinated training algorithm, respectively to having mark training set L and extracting three kinds of features without mark training set U, three base sorters are trained respectively by these three kinds of features, the otherness of data to set of classifiers can be strengthened like this, the uncertainty that effective reduction is brought by single data Expressive Features, strengthens prediction effect with this;

Step 5, puts into the image of retrieval picture library respectively in the middle of three base sorters, voting, carrying out automatic mark to each secondary unmarked picture according to predicting the outcome;

Step 5.1, puts into whole data untagged collection U the set of classifiers be made up of three base sorters and predicts, provides data untagged and concentrate corresponding i after the prediction of each base sorter ^thpredictive marker, predictive marker is designated as Y _in, predict that all predictive marker collection ψ drawn are according to three kinds of different characteristics:

$\mathrm{\Ψ}={\left[\begin{array}{ccc}{Y}_{11}& Y_{12}& Y_{13}\\ \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}\\ \\ ...\end{array}& \begin{array}{c}.\\ .\\ .\end{array}\\ Y_{M1}& Y_{M2}& Y_{M3}\end{array}\right]}_{M\×3};$

Step 6, re-training upgrades sorter, automatic mark data untagged collection;

Step 6.1, after poll closing, obtains the L_temp of all data untaggeds with predictive marker result, carrys out re-training set of classifiers with L_temp, promote the classification capacity of sorter device group; After updated set of classifiers, then data untagged collection U is put into set of classifiers carry out prediction classification;

Step 6.2, as a result, obtains with positive example mark and two group data sets with negative example mark; Data set with positive example mark is denoted as P*, and the data set with negative example mark is denoted as N*;

Step 7, returns results;

Step 8, treats that user determines whether to carry out lower side feedback operation, if it is skip to step 2.1.