Training method of depth image/video compression network

Abstract

The invention discloses a training method of a depth image / video compression network, which can eliminate quantization errors caused by additive mean noise on one hand, realize the consistency of training tests through two-stage fine tuning of a decoder, and obviously improve the rate distortion performance of the whole compression network on the other hand. On the other hand, the quantization granularity of the coding network is flexibly controlled by pre-generating the quantization step size adaptive to the data content from the marginal information, which is a new spatial domain code rate allocation strategy, so that the network can adaptively control the bit allocation on the spatial domain according to the image content during quantization. In addition, the two-stage adaptive quantization strategy can be effective for all coding networks adopting additive mean noise, and meanwhile, the stability of coding network training can be remarkably ensured. In conclusion, video / image compression based on the deep neural network can be more universal, flexible and efficient.

Description

technical field [0001] The invention relates to image / video compression coding, in particular to a training method for a deep image / video compression network. Background technique [0002] Image and video compression coding is an important technology in the electronic information age, which helps to reduce the transmission bandwidth and storage consumption of image and video. Image / video compression essentially realizes the effective representation of information by controlling the relationship between the bit rate (the size of the binary data required to characterize the image and video) and the distortion (the difference between the restored image and the original image). [0003] The existing image / video compression based on variational neural network contains a quantization layer, so it will lead to the problem that the direct training gradient cannot pass through the quantization layer. Specifically, taking the image compression algorithm as an example, the nonlinear t...

AI Technical Summary

Problems solved by technology

However, in the actual encoding and decoding, the entire quantization layer will use the method of direct rounding to obtain discrete hidden layer variables, which causes the problem of inconsistency between training and testing, which in turn makes the encoding performance of the network drop greatly.

At the same time, since the additive mean noise usually ranges from -0.5 to 0.5, the quantization step size is limited, resulting in only integer quantization of hidden layer variables, which in turn limits the flexibility of the compression network.

Images