Research On Face Detection Algorithm Based On Multi-scale Feature Extraction

Face detection was initially the core aspect of automatic face recognition and its fundamental purpose is to determine whether an image contains a face target,and determine its location,size and other information.In recent years,the development and application of machine vision and pattern recognition algorithms have led to the widespread concern of researchers as an independent subject.But so far,there is no algorithm can solve all the problems of face detection.Due to the randomness of the complex structure and the condition of the scene,it is very important to study the stable detection frame in the complex external environment,both in theory and in practice.In general,the face detection system includes three parts:image data collection,feature extraction and classification detection.The focus of the researchers is how to better obtain one or more stable features to express the face.The human eye is often viewed from two main aspects:the local and global,and the multi-scale analysis theory in the field of computer vision is based on these two aspects.The basic concept of multi-scale analysis is to introduce the scale parameters to the image,according to the scale transformation of the human eye visual system,from fine to coarse description of the essential characteristics of the object.The application of multi-scale technology in feature extraction is of great significance to promote the development of face detection algorithm.According to different levels of perspectives,the face will show a variety of structures and forms.The characteristics of the different faces of the same face show different information.The large scale feature is fuzzy,but contains the whole contour information.The overall outline of the small scale feature is not obvious,but contains the details of the information.Therefore,the most intuitive approach is to obtain a scale transformation framework,which can extract the feature information of the face under the dynamic scale change.By fusing the facial features under different scales,we can get better classification effect and improve the accuracy of the detection rate.Based on the previous research results,this paper proposes a new face detection algorithm on the basis of the advantages and disadvantages of the existing facial feature extraction algorithms and the multi-scale analysis theory.The specific contents are as follows:1 The multi-scale theory shown that the characteristics of different scales describe face based on different size descriptors.In fact,for the whole face image,each region is correlated with each pixel,including the feature information of the face.In this paper,we divide the image into three sub-blocks of different scales,extract the HOG characteristics for each sub-block,and weight the sub-blocks according to the difference.The weighted HOG characteristics at different scales are concatenated into a vector,and the dimensionality reduction is done by the Principal Component Analysis(PCA).Finally,the support vector machine is trained as the feature of the dimension vector after the dimensionality reduction to obtain the classification detection model.Compared with other detection algorithms on two open face databases,the proposed method has obvious improvements with the improved detection rate in the same false detection rate.2 Based on improved LBP with the multi-scale feature analysis theory,this paper proposes a face detection algorithm based on multi-scale-direction LBP and SVD.With extending the feature-level feature extraction to the region-level,this paper defines two different scales of the window to traverse the image,and calculate the center of the window pixel or area LBP value.Considering the directionality of face features,LBP eigenvalues are extracted for each pixel in four directions.Through the feature extraction of multi-scale-direction LBP,eight feature images can be obtained for each image,the singular value of the feature image is extracted to get the eigenvector of the samples,and the classification model is constructed to the face detection.The results of Yale B and CMU databases can be proved better than other four face detection algorithms.The detection method proposed in this chapter is better and has strong robustness to changes in illumination and attitude.3 As the proportion of the face in the image is small,the method of searching the face region by traversing the whole image does not satisfy the real-time requirement.Thus we use skin color detection to obtain the approximate range of face targets,followed by initial screening of the possible face regions based on the geometry of the face geometry.After expanding the possible area through the screening to the DOG scale space,the characteristics of each region can be dynamically extracted.By adjusting the scale parameters,the images are analyzed from global to local and its intrinsic properties can be mining.In order to fuse the characteristics of different scales,applying an effective way of weighting is also very necessary.The weighted approach adopted proposed here takes into account the difference among different scales while representing the feature,and fills all the feature images according to the scale parameters of each layer.Finally,the accurate location information will be got by the second filter of SVM.Compared with other detection algorithms on the FDDB face database,the proposed method has obvious improvement while maintaining the low false detection rate with an improved detection rate.