Low-light image improvement apparatus and method

1. A low-light image improvement apparatus, comprising:

an image component decomposition network module that analyzes a light image, a low-light image, and a mid-light image to decompose reflectance and illumination, respectively, wherein the mid-light image is generated using the light image and the low-light image; and

a component improvement network module configured to include a mid-teacher network model that extracts a first feature map with improved reflectance and illumination of the mid-light image using the reflectance and illumination of the light image and a student network module that distills the extracted first feature map and then extracts a second feature map for the reflectance and illumination of the low-light image based on the distilled first feature map and acquires an image with improved light by reflecting a structural component of a multi-band near-infrared image in the second feature map.

2. The low-light image improvement apparatus of claim 1 , wherein the student network model comprises:

a first encoder configured to distill a first feature map and extract a second feature map for reflectance and illumination of the low-light image based on the distilled first feature map;

a second encoder configured to receive the multi-band near-infrared image and extract a third feature map;

a cross-attention transformer unit configured to fuse the third feature map in the second feature map to reflect a structural component of the multi-band near-infrared image;

a first decoder and a second decoder configured to reconstruct the reflectance and illumination using the fused feature map, respectively; and

a synthesizing unit configured to generate the image with improved light by multiplying the reflectance and the illumination reconstructed by the first decoder and the second decoder.

3. The low-light image improvement apparatus of claim 2 , wherein the cross-attention transformer unit fuses a cross-attention map of a self-attention map for the second feature map and the third feature map.

4. The low-light image improvement apparatus of claim 2 , wherein the image component decomposition network module has a separated convolutional neural network model, and the separated neural network models share weights with each other.

5. A low-light image improvement method, comprising:

applying a light image, a low-light image, and a mid-light image to an image component decomposition network to decompose reflectance and illumination, respectively, wherein the mid-light image is generated using the light image and the low-light image;

training the mid-teacher network model to improve the reflectance and illumination of the mid-light image using the reflectance and illumination of the light image, and then knowledge distilling an output of an encoder of the mid-teacher network model into a student network model; and

inputting the reflectance and illumination of the low-light image to the student network model, extracting a feature map for the low-light image based on the distilled knowledge, and reflecting a structural feature of a multi-band near-infrared image in the feature map for the low-light image to generate an image with improved light.

6. The low-light image improvement method of claim 5 , wherein the generating of the image with improved light comprises:

extracting a feature map for the multi-band near-infrared image;

generating a fused feature map by fusing the feature map of the multi-band near-infrared image with the low-light image feature map based on a cross-attention algorithm to reflect a structural component of the multi-band near-infrared image;

reconstructing reflectance with improved light and illumination with improved light using the fused feature map; and

generating an image with improved light by multiplying the reflectance with improved light and the illumination with improved light.