Under the extreme LLL （low light level） conditions (environment illumination less than 2×10^-3lux), the LLL image has the characteristics of low signaltonoise ratio and low contrast, so that the target is difficult to be identified, thus seriously affecting the observation effect. In order to improve the LLL image quality, a convolutional autoencoder deep neural network for image enhancement was designed. In view of the fact that the traditional mean square error loss function cannot meet the human visual perception characteristics, several loss functions including perceptual loss were studied and a novel, differentiable loss function was proposed in combination with the existing full reference image quality evaluation index. Experimental results show that the proposed loss function can improve the detail information of the image while improving the signaltonoise ratio and contrast ratio of the lowlight level image when the network structure does not change. 