Variable bit rate generative compression method based on adversarial learning

A variable bit rate generative compression method based on adversarial learning implemented in a computer, comprising the following steps:

• step (1):
• constructing training and testing data sets through an image acquisition device;
• step (2):
• constructing a generative compression network based on an auto-encoder structure, wherein the generative compression network completely adopts a convolutional layer and a do-convolutional layer to implement a conversion of inputting an original image-compressing an encoding result-reconstructing the original image;
• step (3):
• on training data collected by the image acquisition device, training the generative compression network constructed in the step (2), wherein a training target of an encoder and a decoder is a rate-distortion optimization function, and a target of a multi-scale discriminator is to distinguish a real image and a reconstructed image, and the multi-scale discriminator and the decoder train alternately;
• step (4):
• according to a target bit rate, calculating a correlation threshold of a feature map to be employed;
• step (5):
• based on a feature map channel redundancy index, calculating a mask, wherein after a training is completed, an original network structure is configured for an image compression and outputting a result of a single compression rate;
• and
• step (6):
• performing a lossless compression and decoding on the mask and the feature map, wherein adaptive arithmetic encoding is employed to further compress a binary mask and a feature map corresponding to the binary mask, respectively.