Weak supervised semantic segmentation method based on attention mechanism

Abstract

The invention relates to a weak supervised semantic segmentation method based on an attention mechanism, and the method employs a mode of combining image level supervision and detection frame level supervision, and overcomes the defect that too much noise is introduced into a detection frame through the target attention of image level supervision. Meanwhile, the target integrity supervised by thedetection frame is utilized to improve most of the deletion in the class activation map generated by image level supervision. Compared with the existing weakly supervised semantic segmentation technology, the method has a better segmentation effect.

Description

technical field [0001] The invention relates to the field of machine vision, in particular to a weakly supervised semantic segmentation method based on an attention mechanism. Background technique [0002] In recent years, convolutional neural networks have gradually taken a dominant position in computer vision. The advantage is that they can obtain high-quality features from a large amount of data, but at the same time, a large amount of training data is required to support this good result. This shortcoming is particularly prominent in image segmentation problems. Most semantic segmentation methods rely on large-scale and dense annotation data to train deep neural network models. However, it takes a lot of manpower and time to make pixel-level annotations. According to statistics, it takes an average of 4 to 5 minutes for pixel-level labeling of an image, while image-level labeling only takes 2 seconds, and boundingbox labeling only takes about 20 seconds. In contrast, le...

AI Technical Summary

Problems solved by technology

[0004] The current method based on image-level annotation has poor anti-interference, and complex environments may lead to over-segmentation; based on the detection frame method, when the background and the foreground target color are similar, the background noise in the detection frame is easily mis-segmented into the foreground or the foreground It is mistaken for the background to cause omissions; the target object is marked based on graffiti. When the color of the target object changes too much, it is easy to miss the segmentation target

Images