Deep learning model training method for fine portrait picture segmentation

Abstract

The invention discloses a deep learning model training method for fine portrait picture segmentation. The method comprises: collecting and obtaining portrait pictures, and performing binaryzation to form a training data set; detecting a boundary by using a Canny edge algorithm; expanding the boundary by adopting an adaptive expansion operator to form an edge region, and obtaining a portrait edge / background region annotation graph; and inputting the original image into a deep learning model for training processing, calculating image gradient loss and segmentation cross entropy loss according tothe portrait edge / background region annotation graph in the training processing, and jointly performing training optimization on the deep learning model. According to the method, a rough manual annotation result is used, a segmentation result which is more accurate than manual annotation is trained in a self-supervision mode, the manual annotation cost is greatly saved, and the method can be applied to various scenes such as personnel monitoring, portrait analysis and portrait editing.

Description

technical field [0001] The invention relates to a training method for fine portrait segmentation, in particular to an image gradient loss function based on the image gradient consistency between an original image and a segmentation prediction result map and a training method thereof. Background technique [0002] Portrait segmentation algorithms are widely used in various fields such as video surveillance, portrait analysis, and portrait editing. However, the previous algorithm cannot meet the occasions that require high edge precision. Considering that the deep learning-based model has very high requirements for training and labeling, it is very time-consuming to label high-precision portrait edge details such as hair strands and clothes edges. High-precision portrait training samples are often limited by cost and cannot be realized on a large scale. Therefore, it is of great practical significance to explore training methods that do not rely on high-precision manual labe...

AI Technical Summary

Problems solved by technology

However, the previous algorithm cannot meet the occasions that require high edge precision.

Considering that the deep learning-based model has very high requirements for training and labeling, it is time-consuming to label high-precision portrait edge details such as hair strands and clothes edges

High-precision portrait training samples are often limited by cost and cannot be realized on a large scale

Images