Human Perception Mechanism Based Image Enhancement Research And Its Applications

Human being put forward higher and higher requirements on the quality and application of visual image with the papid progress of science and technology.It is urgent to apply image enhancement technology and methods to improve image quality due to the fact that image quality cannot meet the application demand for the limitation of imaging facilities.Some image enhancement methods at home and abroad have been read and researched extensively.Some image enhancement limits at present have been studied intensively such as easy loss of detail information,difficulty in controlling the degree of enhancement,easy over-enhancement or under-enhancement,and difficulty in satisfying practical application and so on.Some innovative methods have been proposed as follows.A sub-image histogram equalization based image enhancement method has been developed.The histogram of the input image is segmented recursively into several sections based on the mean and variance of the brightness component of the image.The histogram of each section is modified and equalized respectively to overcome the loss of image details caused by the dominance of high frequency information.The processed image brightness component is normalized to solve the saturation problem of intensity and interference caused by non-uniform illumination.A cascade image fusion strategy has been used to enhance the image.We can know that the proposed method can not only enhance image contrast,but also maintain image brightness and texture detail,according to the analysis of experimental results.An adaptive image enhancement method based on information entropy is proposed.The input image histogram is recursively segmented and the sub-histogram is dynamically redistributed according to the information entropy of the image histogram.It is can be used to avoid gray-level merging which leads to the loss of image texture details.Adaptive probability density function adjustment strategy is adopted.The image enhancement is adjusted adaptively to prevent local over-enhancement and local noise amplification.The image has excellent performance in the wide dynamic range of image intensity grade distribution after that.The experimental results show that the method as mentioned above can not only enhance image contrast effectively,but also maintain image details by comparisons with some state-of-the arts based on HE.An image enhancement method has been proposed based on adaptive gamma correction and nonlinear transformation.An exponential-logarithmic function strategy is adopted to assign different weights to the light and dark areas of the image according to human visual perception mechanism.The image brightness component is processed to avoid over-enhancement and under-enhancement.A nonlinear normalization transformation strategy is used to adjust the dynamic range of image gray value to broaden further the dynamic range of image brightness.The experimental results show that the developed method can not only suppress effectively the tendency of the curve of image cumulative distribution function changing too fast,but also avoid effectively the problem of image detail information loss.An image enhancement method has been developed based on both derived graph and Retinex mechanism.A shallow image enhancement method is used to get the derived image under weak illumination at first.Deep convolutional neural network is selected to train and learn light components.An end-to-end mapping relationship is gotten by training and learning the brightness component between low-light image and normal light image.Depth enhancement network learning based image enhancement is realized to overcome the problems of low overall brightness and unclear details in dark areas under weak illumination.Experimental results show that the enhanced image has rich texture details with better contrast and better visual effect by comparisons.An identity authentication under complex lighting environment method has been proposed.A light invariant extraction method has been developed to extract some facial illumination invariance features based on Lambert light reflection model,and it is insensitive to complex illumination changes.Meanwhile,both fast mean filtering and non-nonlinear normalization transformation strategies are adopted to further enhance facial texture details.The identity authentication is realized based on machine learning classification method.Moreover,Log-gabor filter is used to get Log-Gabor characteristic images with different scales and directions.Some LBP shallow features of facial images are extracted.Deep Belief Network(DBN)is employed to automatically extract facial depth texture features from bottom up.Classification and recognition based on facial depth features are carried out.Experimental results show that the proposed method has excellent recognition performance under complex lighting conditions with different illuminations,expressions,poses and occlusions.