Method for reconstructing HDR video based on semi-supervised learning

The invention discloses a method for reconstructing an HDR video based on semi-supervised learning. The method comprises the following steps: 1) preparing a small amount of labeled sample data; 2) respectively inputting the two samples into an optical flow estimation network to respectively obtain aligned LDR frames; 3) respectively inputting the aligned LDR frames into a merge network, obtaining a feature sequence through a hybrid cavity convolution module, and giving more weights to the key region through an attention module to correct the feature sequence; (4) the corrected feature sequences are connected in series and input into a coder-decoder, the coder-decoder obtains an output result by calculating the weight occupied by each pixel of all input frames, and a final HDR frame is obtained through reconstruction according to the output result of the average weighted value of the input frames; and 5) regarding the processes respectively carried out by the two samples as two branches, namely a supervised br