Unsupervised Low Illumination Enhancement Model Based on YCbCr Color Space

Since most of the current Low Light Image Enhancement(LLIE) models are trained under RGB channel, it has problems such as training process difficulties and color deviation of enhancement results. We propose an unsupervised low illumination enhancement model based on YCbCr color space. This model maps the input low illuminance original image based on RGB space to YCbCr space, and then carries out feature learning on the brightness channel Y in YCbCr. First, extract low-level features through standard convolution dimension elevation, and then divide the network into light-enhancement module and detail-recovery module. In the light-enhancement module, conduct deep-global feature extraction and shallow-local feature extraction respectively , to obtain the feature map after light enhancement. Then the feature map is connected and fused with low-level features in the detail-recover module to restore the texture and details lost in the light enhancement process. After obtaining the feature map after detail-recovery module, the new Y connect the feature map with Cb and Cr channels after dimensionality reduction to obtain the final enhanced image. In the process of model training, a new comprehensive loss function is used to realize the constraints on each component of the model. At the same time, the unsupervised learning process is realized by using the generative adversarial network, and the channel attention mechanism is introduced to achieve a better denoising effect. Through experimental comparison, the average NIQE and LOE values of the new model on the test dataset reach 3.88 and 146.2, which has a competitive enhancement effect.