Color Face Image Reconstruction Based On Superpixel Representation In The Quaternion Space

With the rapid development of intelligent monitoring system,surveillance cameras have been used widely in various fields,including security system.However,suffering from the low-definition of the monitoring equipment and the long distance between the target face and the camera,the monitoring equipment is almost impossible to provide useful recognition information.In addition,the quality of the monitoring image will be deteriorated by the poor environmental factors in the real monitoring scene.The obtained target face image features are too fuzzy or contain noise interference,which can not be recognized by human eyes or the machine vision.In order to obtain enough facial feature details for recognition,face super-resolution technology generates high-resolution face images from low-resolution face images to meet the actual needs.Face hallucination has become a research hot spot,due to this field has high application value in the current social background.In recent years,face hallucination method based on learning has been well applied in gray images.However,most of the existing methods either have no flexibility to any pattern shape,or ignore the inherent color information.In order to solve these problems,taking the widely used color face image as the research object,this paper proposes two new learning models.The main research contents are summarized as follows:(1)For low resolution color face images,a color face hallucination method of superpixel-guided locality quaternion representation is proposed in this paper.The model uses superpixels to segment the image into atomic regions with perceptual significance,instead of most of the existing grid segmentation methods.Rather than handling the squared patches with fixed size,the proposed method handles superpixels with adaptive shapes segmented from the face images according to semantic contents,which can well preserve the face spatial features.In addition,in order to fully explore the abundant information between different color channels in the original color image,mapped to the quaternion space,the image is transformed into an orthogonal feature space,and the coefficients are encoded in the quaternion domain.The quaternion coding can not only capture the real topology of face image manifold,but also retain the internal correlation between different color channels.Experiments show that the color face images generated by this method have better performance than several advanced methods.(2)For low resolution color face image affected by noise,a robust color face image super-resolution method based on weighted superpixel local quaternion representation is proposed.The model determines the contribution of pixels in the reconstruction process according to the degree of noise pollution.The core idea of the model is to subtly adjust the contribution of each element by introducing an appropriate reweighting strategy.Severely damaged pixels will lead to large coding residuals.By reducing the contribution of noise elements,the influence of noise can be suppressed,and the role of information dominated pixels will be highlighted in data representation.The proposed method obtains the weight vector through adaptive learning,which can meet the model requirements and data patterns,so as to represent the noisy data well.Experiments show that this method has achieved more outstanding results under the condition of noise.Our researches have important application value in the field of intelligent security system and computer vision.