Face Retrieval In The Wild Based On HOG And LBP Features

In recent years,with the wide-spreading and development of monitoring equipments and smart phones,content-based image retrieval has become a practical research direction,which attaches great importance to industry and academia.In the 1970 s,image retrieval technology mainly based on the keyword image retrieval technology,which need manually annotate the text keywords to describe the image features.This technique relies on artificial annotation of the image,consumes more manpower and accompanied by subjective,inaccurate issues.Especially in the face image retrieval,keyword-based text retrieval for images performs poorly.Since 1990 s,Content-based image retrieval(CBIR)has begun to become popular.However,in recent years,due to the rapid growth of the image generated by cameras and information systems,image retrieval based on visual content has also become an important research field.Because of the environment influence,low resolution and multi-pose,constitute the main influence factors of the extraction of facial features.At present,natural light scene,low resolution,multi-pose and other mixed conditions of the study is still a challenging problem.In this paper,the multi-local face feature fusion method is used to extract the feature of the face image and has obtained the expression of the feature of the face image.Then,the representation of the face image in the dataset are clustered by clustering method,and the performance of all the clusters is evaluated.According to the clustering performance,we select the different clustering retrieval process,and design an accurate and fast face retrieval algorithm.The main contributions of this paper are as follows:1.Collecting a complete dataset.The face image dataset provides sufficient conditions for verifying the face retrieval algorithm,and a complete dataset will verify the performance of the algorithm more comprehensively.Among many human face image retrieval experiments,FERET,BioID,LFW,ORL and other image sets are commonly used.But most face images in these dataset are on the front and with specific light,which does not adapt to practical application.In this paper,we collect and organize a more natural dataset,which includes college surveillance video that records the moving situation of the staff,students and other personnel in my college.In the video,the monitor is in natural light.The processing of the college surveillance video: the monitoring of video changes in each frame for skin detection,oval detection,face extraction,and storage.Specifically,when there is a change between the foreground image and the background frame image obtained in the next frame,it is generally considered that there will be one or more person thus a change has happened.If the degree of change reaches a certain threshold,we think that the figure appears.The paper uses the current image frame minus the background image to obtain the part of the change,and this part will be processed by skin detection,and oval detection,Then we use the smallest rectangular to describe the irregular contour,and save each smallest rectangular for each contour face.This is the most likely to exclude the natural environment of the non-face area.2.Image preprocessing by using light compensation.The process of data collection has been used for lighting compensation.However,the light compensation preprocessing is also used before the face image feature extraction.In the process of using the image pixels in the first five percent of the maximum value of the pixel as the compensation factor,on the basis of each pixel value to compensate,so as to achieve the purpose of weakening the impact of light.3.Local feature extraction and fusion.In this paper,considering the disadvantage of retrieval effectiveness only by using face search the HOG and LBP algorithms are used to extract the feature and obtain the face representation respectively.Finally,the natural feature in the wild based on the HOG and LBP features of the corresponding image is obtained according to the weight ratio of 1: 1.4.Clustering and its performance evaluation.Due to the impact of posture and size and other objects lead to more alienated objects,resulting in higher false detection rate of face retrieval.The local feature extraction and fusion are based on the clustering of human face description.In this paper,the clustering of face feature description is carried out on the basis of local feature extraction and fusion,so as to guarantee the distance between the same objects is as compact as possible,and the distance between classes is alienated.At the same time,the recall rate retrieved during the retrieval process is guaranteed,and the missed rate is reduced.In clustering,this paper chooses to fully believe in clustering.Finally,for non-clustering,fully believe clustering method and dubious the comparability between clustering method and dubious clustering method.The results show that the dubious clustering method is better than the others.Using clustering can improve the recall rate of similar faces,at the same time,it also ensures the accuracy of similar face retrieval.In this paper,the face image in the dataset is retrieved by using the image preprocessing based on light compensation,local feature extraction,local feature fusion and clustering.First,the illumination compensation weakens the impact of light.The fusion features make the features more compact and accurate.The clustering of features improves the recall rate and accuracy rate of face retrieval.The above research result has far-reaching significance and important practical value for the work certification,the suspect tracking and many other practical applications.