Low-Light In-Room Image Enhancement Network Based on Retinex Theory

Taking pictures indoors always faces low light and high number of shadows, and because of the complexity of the light source, it indirectly leads to more noise in indoor images than in images captured from outdoor scenes. Existing techniques do not work satisfactorily in indoor photographed images, in order to overcome these difficulties we propose a new network: RTI-Net. The network is based on Retinex theory, which decomposes a low-light indoor image into a light image and a reflection image. We add the inception network to the original Retinex network so that the network can both reduce the noise of the image and improve the shortcomings of the original image with insufficient illumination. We also add Color Constancy Loss to the loss function to reduce the image color distortion, so that the final enhanced image color is closer to the real world. The final experimental results show that our proposed network is ahead of other methods in both subjective and objective aspects.