Semi-supervised Leaf Instance Segmentation Based on IOU Score Prediction and Orthogonal Regularization

In the development of smart agriculture, highthroughput and high-precision fully automated plant phenotyping data acquisition and analysis techniques are hot spots for future research. With the application of deep learning techniques in plant phenotyping, the efficiency and performance of plant image recognition have been significantly improved. Nevertheless, instance segmentation of plant leaves, a key task in plant phenotyping research, has not yet been extensively studied or reached the desired performance level. This is mainly due to the large number of leaves in an image as well as the high cost of manual annotation. Therefore, this study proposes to adopt a semisupervised learning framework for instance segmentation of plant leaves. This method requires only $1 \%-10 \%$ of the data to be labeled, and utilizes the unlabeled data as additional training signals, which leads to a model performance approaching that of fully supervised segmentation. This method not only reduces the labeling cost, but also significantly improves the segmentation accuracy, providing a faster and more accurate means for plant growth monitoring and yield assessment. In the semi-supervised leaf instance segmentation task, we introduce a state-of-the-art guided distillation framework. Given the single class of leaf dataset and the limited amount of semi-supervised training samples, we use orthogonal regularization to optimize the model parameters and reduce overfitting. In addition, we added an IOU-based pseudo-label generator to increase the number of high-quality pseudo-labels and further improve the model performance. Our improved method significantly improves the performance of mask-mAP50 on the LSC public dataset, e.g., from 24.70 to 26.66 at 1% labeling ratio, and from 26.5 to 33.28 at 2% labeling ratio, thus validating the effectiveness of the method.