A CBIR System Research and Implementation

Abstract. The content-based image retrieval technology CBIR meet the actual needs of users， it is a key part of CBIR about the appropriate image feature extraction and similarity measure. It is used in this paper that the image similarity is calculated by the histogram intersection algorithm. For this questions is studied about the color histogram can not report the space information， this paper uses the method of the color matrix to compensate for this defect and improve the color histogram. Moreover， a simulation is achieved in the face retrieval system.

Keywords： Face retrieval； CBIR； Face Detection； Picture Retrieval.

1. Introduction

Welcome to the 2010 International Workshop on Vocational and Technical Education. The Conference is a primary international forum for scientists and technicians working on topics relating to Vocational and Technical Education. It will provide an excellent environment for participants to meet fellow academic professionals， to enhance communication， exchange and cooperation of most recent research， education and application on relevant fields. It will bring you new like-minded researchers， fresh idea. It also provides a friendly platform for academic and application professionals from crossing fields to communication together.

The face recognition technology uses a computer to analyze face images， and to extract the identification information to recognize the identity or discriminat a pending state. It involves knowledge of many disciplines of pattern recognition， image processing， computer vision， physiology， psychology.It is one of the hot currently. However， the recognition results are often affected by the facial expressions， posture and illumination changes. Content-based image retrieval method is by a given BMP file picture to find the similar picture with the face in the picture folder. According to the human face characteristics， it can be judged to determine whether they contain the picture of the person so as to achieve the purpose of retrieval.

2. What is the Face retrieval

The Face retrieval is to find a face in a large picture library according to the contents of face characteristics. The retrieval process contains the retrieved or similar face pictures or video segment. In recent years， the face detection has been made leaps and bounds. Face as one of the most important visual objects has an important position in the images and video， computer vision， pattern recognition， multimedia research fields.

3. Face retrieval

CBIR technology （Content-based image retrieval） uses the image content directly to find the basic visual features within the image， such as color， texture， shape. Indexing can be made for similarity matching based on these characteristics， and ultimately find to satisfy the query conditions related images. The basic principles of CBIR is the first image in the image database and extract the features of the corresponding feature library， in the retrieval of images， using the same method to extract the characteristics of samples required， and the characteristics of the feature library similarity matching， the matching results back to the image library and then returned to the user， and finally arranged according to similarity of size.

As mentioned above， the CBIR technology can make use of color， texture， shape and other characteristics of the image within the image retrieval， in which color is the most widely used visual features in the image retrieval， objects and scenes and images contained in a very close relationship and compared with other visual features， less affected by rotation， translation and scale change， has a high robustness. The color histogram is the most common method of expressing color features， it describes the proportion of different colors in the entire image to calculate the image color distribution， ie the probability of occurrence of various colors， and use the histogram intersection algorithm calculate the similarity of the images and do not care about in which the spatial location of each color， so it does not identify the image of the object. The histogram is more suitable to describe the images that are difficult to object recognition and does not require consideration of the color space location. Image matching method， how to use the value of said image similarity in color， this is the image similarity measure.

In pattern recognition， the features of the similarity measure commonly used distance method， ie， the spatial distance of the feature vector to represent. ， By setting the distance between the histogram in the histogram retrieval， when their distance is less than a given threshold， that is considered to meet the search results， the distance metric， commonly used Euclidean distance， the horse distance， the second distance. Color histogram does not reflect the spatial information， you can use the color matrix to compensate for this， the principle is to first split the image to the appropriate block， and then use the color histogram extraction for each sub-block of local color feature.

4. Face retrieval software

Face retrieval is nothing more than to retrieve the picture containing the face， to implement this feature， you must have two steps： face detection and image retrieval.

4.1 Face Detection

Face detection is a key technology in the face recognition process. The face detection is to determine all face in the input image （if present） the location， size， position and orientation process. The face detection problem contains very broad， from different angles can have a variety of classification methods， such as： image sources can be divided into static images and dynamic images （video sequence）， Face Detection； can be divided according to the color information color images and grayscale images， face detection； can be divided into simple background and the complex background human face detection， the complexity of the image background.

Face detection method based on gray-scale features， such as the knowledge model-based methods， methods based on statistical models. Prior knowledge of the human face as a combination of organ characteristics the knowledge model-based approach to detect based on the geometric position relationship between the characteristics of organs such as eyes， eyebrows， mouth， nose and face. Statistical model-based approach is to use a lot of "face" and non-face training samples to construct classifier， and then use the classifier for face detection.

The system uses a statistical model-based approach to face detection， it is a total of two steps， first step is to train the classifier， the second step is to use the trained classifier for face detection.

4.1.1 Training the classifier

Trained the system call OpenCV Haar classifiers， although there is no classifier training， but it is necessary to do some introductory principles and process of the classifier training.

Face detection classifier based on haar features using Adaboost learning algorithm training， Adaboost learning algorithm is an iterative method， most taxonomic significance into a strong classifier from a large number of weak classifiers， select the purpose of order to improve the detection accuracy. First， the sample （about hundreds of pieces of sample images） haar features of classifier training， a cascade of boosted classifier. Training samples into the sample of positive examples and negative samples， positive sample means to be detected the target sample （such as human faces or cars， etc.）， and negative samples refers to any other picture， all sample images have been normalized to the same size size （for example， 20 * 20）.

Classifier training is finished， it can be applied to the detection of the input image in the region of interest （the same as the training sample size）. The detection of the classifier output to the target area （car or face） is 1， otherwise output 0. In order to detect the entire image， you can move the search window in the image to detect each position to determine the possible target. In order to search for a different size of the object， the classifier is designed to change the size， so that more effective than changing the size of the image to be seized. Therefore， in order to detect the object of unknown size in the image， the scanner typically need to scan the image several times with different ratio of the size of the search window.

Classifiers in a "cascading" refers to the final classifier is composed by several simple classifier cascade. Image detection， inspection window in turn by each level of the classifier， so that most of the candidate region in the front layers of detection were excluded from all through the detection area of each level classifier is the target area. Boosting technical support classifier there are four kinds： Discrete on Adaboost， Real on Adaboost， Gentle on Adaboost， Logitboost. "Boosted" refers to the cascade classifier of each layer can be selected from a boosting algorithm （the weight of votes）， and the basis of classification of self-training.

4.1.2 use the classifier to detect

Broadly go through three stages for the use of the classifier is loaded from the file classifier， the classifier for testing， the tests are complete， the release of the classifier. These three stages， corresponding to the three function： cvLoadHaarClassifierCascade， the， cvHaarDetectObjects cvReleaseHaarClassifierCascade. cvLoadHaarClassifierCascade used for training good cascade classifier； cvHaarDetectObjects the load from the file using the cascade classifier for the training of a target object in the image to find a rectangular region that contains the target object， and will return to these areas as a sequence of matrix box； cvReleaseHaarClassifierCascade used to release a cascade of dynamic memory.

4.2 Image Retrieval

Face images， the next step is to retrieve from a large number of face images in a similar picture with the face in the picture， implementation steps are as follows：

First collected a face image from the camera， the image size to 256X256 normalized， and divided into 4X4 block， and then calculate the color histogram of the 16 regions， the last to be identified to the client through the camera to capture face images Concurrent to the server， the server receives the Euclidean distance of color histograms of face images in the color histogram of human face images of the face image is calculated to be identified with people face database， arranged in descending order， if similar less than a certain threshold （statistical threshold value of 0.007） is considered to be to identify a face in the face and the face database for the same person， to achieve matching.

Face image retrieval server-side interface is as follows：

Fig. 1： Face image retrieval system.

5. Summary

Face as one of the most important visual objects in the images and video， computer vision， pattern recognition， multimedia technology plays an important fifth. Meet the actual needs of the user's content-based image retrieval technology CBIR， image feature extraction method is a key part of the CBIR， In this paper， the color histogram does not reflect the spatial information， the use of the color matrix to compensate for this defect， the principle is the first image segmentation for the appropriate block， and then use the color histogram extraction for each sub-block of local color features. On this basis， the face detection system， the system uses a statistical model-based approach to face detection， and achieved convincing results. But the current system is not perfect， especially for face detection light requirements study， a match quality is not high， further research is needed to improve.

References

[1] Fuzhen Huang， Janbo Su. Face Detection. Shanghai Jiaotong University Press， 2006.

[2] Yinghui Wang. Face Recognition - Principles， methods and techniques. Science Press， 2010