FLSSnet: Few labeled samples segmentation network for coated fuel particle segmentation

Abstract

Current image semantic segmentation networks based on convolutional neural networks require extensive training samples and pixel-level annotations to achieve satisfactory segmentation performance. This paper proposes a novel FLSSnet segmentation network based on an encoder-decoder structure. A modified MobileNetV3 backbone network, a contour-assisted training branch, and an improved semi-supervised training method with a cosine function to dynamically update the pseudo-label training weight are added to the network, which effectively reduces its dependency on labeled samples and improves its segmentation accuracy. Compared with other image segmentation networks, the proposed network does not need too many labeled samples for training to achieve the high segmentation accuracy required for industrial inspection. The effectiveness of the proposed network is validated on the coated fuel particle segmentation dataset and the MICCAI’s 2015 gland segmentation public dataset. It has great industrial application value in automating the thickness measurement of each layer of coated fuel particles and other fields.