Lighting condition is a key factor in image quality. This is even more important for human facial images, which contain a wealth of identity information and detailed features. In order to adjust the illumination of human images, various relighting techniques have emerged. However, in the process of human head relighting, the relighted images always face problems such as overexposure, underexposure, and loss of details in the relighted area. Therefore, it is necessary to design an effective quality assessment method for relighted human heads (RHHs) to promote the improvement and optimization of the relighting technology. In response to this urgent need, this paper proposes a light-aware quality assessment method based on multi-task learning. Specifically, the method adopts a randomized patch scheme and uses Swin Transformer for feature extraction. Subsequently, a multi-task learning module is designed to divide the task of RHH quality assessment into two subtasks: light perception and quality regression. These two subtasks share the learned knowledge and promote each other to enhance the quality perception capability of the proposed method. Experimental results show that the method is significantly better than existing quality assessment methods and closer to human visual perception.