A Weakly Supervised Semantic Segmentation Model of Maize Seedlings and Weed Images Based on Scrawl Labels

Area of Science:

• Agricultural Science
• Computer Vision
• Deep Learning

Background:

• Fully supervised deep learning for maize and weed segmentation requires extensive pixel-level annotations, which are costly due to complex plant morphology.
• Existing methods face challenges in efficient and accurate semantic segmentation of agricultural images.

Purpose of the Study:

• To develop an efficient weakly supervised semantic segmentation network (SL-Net) that reduces the reliance on pixel-level annotations.
• To improve the accuracy and reduce the cost of semantic segmentation for maize and weed images.

Main Methods:

• Proposed SL-Net incorporating a pseudo label generation module, an encoder, and a decoder, built upon the DeepLab-V3+ model.
• Utilized scrawl labels converted into pseudo labels to train the network, employing a migration learning strategy.
• Improved feature extraction using an enhanced backbone network.

Main Results:

• The pseudo label generation module achieved 83.32% intersection over union (IoU) and 93.55% cosine similarity with ground truth.
• SL-Net reached a mean IoU of 87.30% and average precision of 94.06% in segmenting maize and weeds.
• Achieved superior performance compared to DeepLab-V3+ and PSPNet under both weakly and fully supervised conditions.

Conclusions:

• SL-Net effectively reduces the annotation burden for semantic segmentation in agriculture.
• The proposed method demonstrates high accuracy and efficiency in distinguishing maize plants from weeds using scrawl labels.
• SL-Net offers a promising solution for automated agricultural image analysis.