Semantic segmentation method under small sample based on variational prototype reasoning

Abstract

The invention discloses a semantic segmentation method under a small sample based on variational prototype reasoning, and belongs to the field of computer vision semantic segmentation. The invention discloses a semantic segmentation method under a small sample based on variational prototype reasoning. According to the invention, variational prototype reasoning is proposed for the first time; semantic segmentation under a small sample is brought into a probability framework; in the probability framework, prototype representation is not vector representation of a fixed numerical value, but distribution, distribution of a hidden space in variational reasoning is used for representing distribution of the prototype, and the generalization ability of the whole prototype is improved under the condition of a small sample so as to adapt to uncertainty represented by the small sample; moreover, an objective function of variational prototype reasoning suitable for image semantic segmentation under a probability framework is proposed for the first time, a semantic segmentation process under a small sample is assisted, and multiple tests prove that a very good segmentation effect is also obtained under the condition of utilizing a single support set image.

Description

technical field [0001] The invention relates to the field of computer vision semantic segmentation, and more specifically, relates to a semantic segmentation method under small samples based on variational prototype reasoning. Background technique [0002] Deep learning has been widely used in semantic segmentation of computer vision, but due to the lack of labeled supporting data in practical semantic segmentation applications, the performance of the learned deep model is affected. At present, the prototype-based method is more popular. Here, the prototype refers to the representation of a class of objects. In the deep learning framework, it is the output generated by a deep neural network based on the label information of the supporting image and its corresponding object. In other words, a prototype is an associative map between input support images and object categories. [0003] For semantic segmentation under small samples, it is basically based on the prototype method...

AI Technical Summary

Problems solved by technology

However, this method is difficult to achieve good results in the case of small samples. In the case of small samples, using a simple vector to represent the prototype will have bias and lack of generalization ability

Images