A DETR-like detector-based semi-supervised object detection method for Brassica Chinensis growth monitoring

• Published in: Computers and electronics in agriculture 2024-04, Vol.219, p.108788, Article 108788
• Main Authors: Li, Haoyan, Shi, Fanhuai
• Format: Article
• Language: English
• Subjects: Brassica Chinensis; Crop growth monitoring; Pest control; Phenotypic measurement; SSOD
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2024.108788

•Semi-Supervised-Object-Detection method greatly reduces training cost of Brassica Chinensis growth monitoring.•End-to-end Semi-Supervised-Object-Detection method proposed aligns well with unique characteristics of Brassica Chinensis growth monitoring.•Low threshold filtering and decoupled optimization address class-imbalance and multi-task conflict in Brassica Chinensis datasets. Object detection technology plays a crucial role in crop growth monitoring within smart agriculture. However, data labeling is a costly process necessary for constructing a large-scale dataset, which is essential to prevent overfitting in deep learning-based object detection models. Semi-Supervised Object Detection (SSOD) presents a cost-effective solution to reduce labeling and model training expenses; nevertheless, existing SSOD algorithms fall short in addressing the specific challenges posed by detection tasks in Brassica Chinensis growth monitoring. Specifically, the two-stage object detector cannot be well-suited for scenes characterized by severe occlusion and complex backgrounds. The Non-Maximum Suppression (NMS) may filter out numerous true positives in scenarios with severe occlusion. Moreover, the label assignment enlarges the negative effects of the noise introduced by teacher model’s prediction, resulting in potential divergence. To tackle these challenges, we propose an end-to-end SSOD method based on Detection Transformer (DETR), which streamlines the post-processing without NMS and adopts a more advanced bipartite matching assignment strategy. These modifications tailor the semi-supervised training method to better align with the unique characteristics of detection tasks in Brassica Chinensis growth monitoring. Furthermore, two key techniques: low threshold filtering and decoupled optimization, are introduced to address class-imbalance and multi-task optimization conflict in the tasks, respectively. In the end, we conduct experiments using two self-constructed Brassica Chinensis image datasets to validate the effectiveness of the proposed method, which demonstrates state-of-the-art (SOTA) performance in both tasks. For plant detection, the proposed method achieves an mAP of 74.1 using only 5 % of the total data volume (18 images). In the wormhole detection task, the method achieves an AP50 of 73.7 using 5 % of the total data volume (73 images). These impressive results meet the requirements for practical applications in Brassica Chinensis growth monitoring.