Face Image Illumination Processing Algorithm Research Based On Retinex Theory

Face recognition has important theoretical value and broad application prospects, however, illumination variation under unconstrained environment will lead to the performance degradation of face recognition system seriously. To deal with the illumination variation problem, domestic and foreign researchers have proposed a variety of face image illumination processing algorithms, among which processing algorithms based on signal analysis have become a mainstream method for engineering application owing to their fast computing speed and requirement for no prior knowledge. Recent years, Retinex algorithms belonging to the above-mentioned algorithms have been introduced and present a new approach to solve illumination variation in face recognition for practical application.In order to improve the accuracy, robustness and practicability of illumination variant face recognition; this dissertation is focused on improving the effects and efficiency of Retinex algorithms in processing face images. The main work and innovation of this dissertation includes:First, a Retinex algorithm based on the model of the retina is proposed to perfect the effect of Retinex algorithms in the removal of illumination from face images. By reference to the functional structure of the human retina, the single-scale Retinex algorithm(SSR) and an adaptive smoothing Retinex algorithm via contextual discontinuities are used respectively to mimic the performance of each functional layer of the retina. Experimental results show that the average recognition rate of the proposed algorithm achieves 93.98%, improved by 20.51% and 5.4% compared with the classical algorithm based on retinal model and widely used adaptive Retinex algorithm separately. Therefore, the proposed algorithm can satisfy the requirement for the demanding recognition accuracy of engineering fields.Second, as iterative computations result in the decline of the efficiency of Retinex algorithms, a Retinex algorithm based on fast bilateral filters is proposed. We use increment-dimensional bilateral filter to separately estimate illumination and reflecting components of the image thus can remove the noise and preserve face detail features effectively while eliminate image illumination. Meanwhile, the increment-dimensional bilateral filter can convert the tedious calculation of the traditional bilateral filter to a linear convolution so that the proposed algorithm can not only ensure the processing effects but also improve the algorithm efficiency dramatically. Experimental results show that the working time of the proposed algorithm is improved by 46% compared with the widely used adaptive Retinex algorithm, close to the faster Retinex algorithm based on mean shift algorithm in current Retinex algorithms. Moreover, the recognition rate of the proposed algorithm is higher than the above two algorithms, so suitable for some practical applications which demand to get results rapidly.