Facial expression recognition (FER) has been acknowledged as a significant modality that could bring facial expression into human-machine interaction and make the interaction more efficient. However, the ability of FER to operate in a fully automated and robust manner is still challenging. Illumination effects, for example, make the facial expression images always contaminated with different levels of ambient noise (such as brightness variation) in a cluttered background. Thus, this paper aims to investigate the illumination effects (brightness variations) on facial expression recognition using empirical mode decomposition reconstruction techniques. In this framework, firstly, the noisy facial expression images were simulated with the illumination effects using different brightness levels of 30%, 40%, 50%, 60%, and 70%. Then, the EMD will decompose the noisy facial expression images into a small set of intrinsic mode functions (IMF), namely IMF1, IMF2, IMF3, and residue. Based on property held by EMD, the signals are decomposed into several IMF components, each with a different time scale. Because the last several IMFs represent the majority of illumination effects, various reconstruction techniques for IMFs have been investigated at various brightness levels. Feature reduction techniques Principal component analysis (PCA) and linear discriminant analysis (LDA) have been employed to reduce the high-dimensional space of IMF features into low-dimensional IMF features. The reduced IMF reconstructions were then used as input to the k-nearest neighbour classifier to recognise the seven facial expressions. A series of experiments have been conducted on the JAFEE database using various reconstruction IMFs together with PCA plus LDA. Based on the results obtained, the reconstruction of IMF1 + IMF2 + IMF3 shows the highest accuracy in high illumination conditions, which is 99.06%.