Deep learning feature compression and decompression method, system and terminal

Abstract

The invention provides a deep learning feature compression and decompression method and system, and a terminal, and the method comprises the steps: carrying out the compact space transformation of an input original feature through a coding end, and obtaining the compact feature expression of the original feature; calculating an importance coefficient of each channel in the features by using compact feature expression, and performing quantization parameter adaptive distribution of each channel; performing non-uniform quantization on different channels based on the distributed quantization parameters to obtain a quantized multi-channel feature map; and performing feature coding on the quantized multi-channel feature map to complete compression. The decoding end carries out decoding to obtain an inverse quantized multi-channel feature map; self-adaptive performance compensation is carried out on different quantization levels of the inversely quantized multi-channel feature map; and original feature reconstruction is performed on the compensated multi-channel feature map to complete feature decompression. According to the method, the coding end adaptively allocates the quantization parameters based on the image content, and the decoding end adaptively performs performance compensation based on the quantization level, so that the performance of non-uniform quantization on feature compression is improved.

Description

technical field [0001] The present invention relates to the technical fields of image processing and computer vision, in particular to a deep learning feature compression and decompression method, system and terminal based on quantization adaptive distribution and compensation. Background technique [0002] Traditional image compression technology is designed for human visual characteristics, and with the superior performance of deep neural networks in various machine vision tasks, such as image classification, target detection, semantic segmentation, etc., a large number of machine-based AI applications for vision. In order to ensure that the performance of machine vision tasks is not damaged by the image compression process, the analysis-then-compression (Analysis-Then-Compression) mode is adopted to meet the needs of machine vision, that is, the lossless image is directly characterized by the neural network at the image acquisition end. Extraction, and then encode and tr...

AI Technical Summary

Problems solved by technology

[0004] At present, there is no description or report of the similar technology of the present invention, and no similar data at home and abroad have been collected yet.

Images