Video Description Model Based on Feature Reinforcement and Knowledge Supplement

The video captured in low illumination environment often has low contrast, much noise, unclear details and other problems, which seriously affects the subsequent target detection, segmentation and other computer vision tasks. Most of the existing low-light video enhancement methods are built based on convolutional neural networks. Because convolutional can not make full use of the long-term dependence between pixels, the generated video often has the problem of detail loss and color distortion in some areas. To solve the above problems, a twin low-illumination video enhancement network model integrating local and global is proposed. The local feature extraction module based on deformable convolution is used to capture the local features of video frames, and a lightweight self-attention module is designed to capture the global features of video frames. Finally, the feature fusion module fuses the extracted local features and global features to guide the model to generate more realistic and detailed enhanced videos. The experimental results show that it has certain engineering value.