Illumination Normalisation for Face Recognition Using Generative Adversarial Network

Recently, research on object classification and recognition using machine vision techniques has been increasingly applied across various industrial domains. Among these, face recognition is widely used in numerous applications such as unlocking mobile phones, crime identification and home security systems. However, the facial appearance is significantly affected by variations in illumination; therefore, illumination variation is one of the factors that degrade the performance of face recognition systems. In this paper, based on Retinex theory for face recognition insensitive to illumination variations, we propose a deep learning method that combines adversarial training with normalising illumination components from face images and generates only reflective components from low‐light illumination images. We decompose the original face image into an illumination component and an illumination‐invariant component with a Retinex constraint and then use a generative adversarial network for face image reconstruction. We also design a loss function to ensure visual quality and preserve discriminative information for neural network optimisation. The proposed method removes the influence of illumination from face images locally as well as globally and preserves the detail information of the face well. To evaluate the performance of the proposed method, we tested the Yale B, extended Yale B and CMU. PIE face databases were built in an illumination changing environment and the average recognition rates of 99.418%, 93.658% and 99.55% were obtained by the proposed method, respectively. We also showed improved recognition performance compared with traditional methods. This confirms that the proposed method is less sensitive to illumination variations and is suitable for face recognition.