Image processing method and system, electronic equipment and storage medium

Abstract

The invention discloses an image processing method and system, electronic equipment and a storage medium. The image processing method comprises the steps of inputting RGB values of all pixel points inan image into a deep neural network model subjected to integer quantization, wherein the output and the weight of each layer of network in the deep neural network model are quantified into integer; inversely quantizing the integer probability array output by the deep neural network model into a floating point type probability array, wherein the probability array comprises the probability that each pixel point in the image is a foreground portrait. According to the invention, RGB values of all pixel points in an image are inputted to a deep neural network model subjected to integer quantization, the output and the weight of each layer of network in the deep neural network model are quantified into integer, the whole-course integer operation of the deep neural network model in the process of calculating the foreground portrait probability is ensured, and compared with the floating-point operation in the prior art, the bandwidth is saved, the time delay is reduced, and the power consumption is reduced.

Description

technical field [0001] The present invention relates to the field of image processing, in particular to an image processing method and system, electronic equipment and a storage medium. Background technique [0002] The portrait background replacement technology is essentially a portrait matting task plus a portrait background fusion task. Specifically, foreground portraits are accurately estimated from natural images or videos containing portraits, and they are seamlessly blended with preset background images. [0003] Portrait matting belongs to the task of image matting, which is essentially a pixel-level regression task, which can be expressed by the following formula: [0004] I i = α i f i +(1-α i )B i ,α i ∈[0,1] [0005] Among them, the value I of the i-th pixel i by foreground point F i and background point B i Weighted according to the above formula, α i Foreground transparency describes the probability that the pixel belongs to the foreground portrait,...

AI Technical Summary

Problems solved by technology

[0008] The above-mentioned map-cutting schemes provided in the prior art either require manual interaction and cannot be completed automatically, or are all floating-point network models, which require a large amount of calculation and serious memory consumption, and can only be run on cloud servers or desktops

For mobile phones, edge devices, etc., these solutions are limited by memory, power consumption, speed, etc. and cannot be used

In addition, different from the classification problem, the matting of portrait background replacement needs to accurately regress an alpha, which is very sensitive to the loss of quantization accuracy. If the existing floating-point model is directly converted to the fixed-point model, there will be a serious loss of quantization accuracy, resulting in Obvious visual artifacts, such as severe quantization aliasing or quantization moiré at the edges of the alpha map

Images