Mobile terminal image segmentation model training method and device

Abstract

The embodiment of the invention provides an image segmentation model training method and device for a mobile terminal. Comprising the steps of obtaining an image segmentation training sample set, an initial model, an assistant model and a student model; training the initial model by adopting the first fine annotation data to obtain a teacher model which is used for annotating the image segmentation training sample set; adopting the teacher model to label the coarse label data, and generating second fine label data and difficult case data; combining the second fine annotation data, the difficult case data and the first fine annotation data to obtain a total data set; training the teaching assistant model by adopting the total data set; and based on the total data set, adopting the trained assistant model to perform distillation training on the student model to obtain an image segmentation model which is used for performing image segmentation processing on the mobile terminal. A large model is used for guiding a small model to carry out training step by step, so that speed, light weight and precision are taken into consideration.

Description

technical field [0001] The present invention relates to the technical field of image processing, in particular to a method for training an image segmentation model for a mobile terminal, a training device for an image segmentation model for a mobile terminal, an electronic device, and a storage medium. Background technique [0002] Image segmentation is widely used in a variety of scenarios, among which sky segmentation is mostly used in image beautification, defogging and post-production. Traditional sky segmentation uses Gaussian filtering or edge extraction operators to process images, which is time-consuming and computationally intensive and difficult to apply in the real-time field. However, the neural network model trained by deep learning usually has a large number of parameters. Although the effect is good, it cannot be deployed on mobile terminals such as mobile phones. It usually needs to be transmitted to the cloud server for processing through the network. It lac...

AI Technical Summary

Problems solved by technology

Traditional sky segmentation uses Gaussian filtering or edge extraction operators to process images, which is time-consuming and computationally intensive and difficult to apply to real-time fields

However, the neural network model trained by deep learning usually has a large number of parameters. Although the effect is good, it cannot be deployed on mobile terminals such as mobile phones. It usually needs to be transmitted to the cloud server for processing through the network. It lacks robustness and high cost under extreme conditions.

Images