In an article published in the journal PLOS One, researchers from China proposed a novel architecture for detecting sugarcane stem nodes using the YOLOv5 framework, named G-YOLOv5s-SS. They showed that this algorithm can achieve high accuracy and efficiency in identifying sugarcane stem nodes, which are crucial for the mechanization of sugarcane cultivation.
Study: Lightweight Detection of Sugarcane Stem Nodes: G-YOLOv5s-SS. Image credit: Generated using DALL.E.3
Background
Sugarcane is an important crop that serves as the main raw material for sugar production and a vital resource for the chemical, light, and energy sectors. However, traditional techniques of sugarcane cultivation are time-consuming, labor-intensive, and inefficient, which pose challenges to the development of the sugarcane industry. Therefore, there is a need to adopt mechanized and intelligent practices in sugarcane cultivation.
One of the key steps in sugarcane cultivation is seed cutting, which requires segmenting sugarcane to achieve a high germination rate and minimum seed consumption. Since the sugarcane buds grow on the stem nodes, it is essential to conserve the integrity of these nodes while seed cutting. Hence, the automatic, fast, and accurate identification of sugarcane stem nodes is a game-changing innovation that can improve the production of sugarcane seeds.
Existing methods for sugarcane stem node identification are mainly divided into deep learning and traditional machine vision methods. Traditional machine vision methods rely on features such as grayscale, texture, edges, and color of the sugarcane for identification, but they suffer from low recognition speed, high environmental requirements, and poor robustness and generalization ability. Deep learning methods, represented by convolutional neural networks (CNNs), can extract valuable feature information from images and achieve effective target detection in challenging environments.
However, most of the existing deep learning methods focus on enhancing the accuracy and speed of the algorithm without considering the model’s lightweight characteristics. A highly complex model demands greater memory and computing power, which contradicts the trend of equipment miniaturization and significantly raises the cost of mechanizing sugarcane planting for mobile and embedded devices.
About the Research
In the present study, the authors proposed a lightweight architecture for the detection of sugarcane stem nodes based on the YOLOv5 framework, which is an improved one-stage target-detection algorithm based on YOLOv3. YOLOv5 has a small model size, fast training, and reasoning speed, and is flexible, making it suitable for various applications. The paper aims to reduce the model’s size, complexity, and computational burden while maintaining high detection accuracy and efficiency for identifying sugarcane stem nodes.
The novel architecture G-YOLOv5s-SS is designed by applying the following strategies:
- Some modules were removed from the backbone network of YOLOv5 to reduce the number of down-sampling times and detection layers. This was done to fully utilize the shallow-level feature information that is beneficial for detecting small targets such as sugarcane stem nodes.
- Ghost modules were introduced. They used a simple linear transformation to generate phantom features, replacing redundant features generated by regular convolution. This results in more effective feature maps with fewer parameters and computations.
- The SimAM attention mechanism was incorporated. It can derive 3D attention weights for the feature map without adding additional parameters, enabling the network to learn more discriminative neurons and improve the feature extraction ability.
The researchers used an intelligent transverse sugarcane seed pre-cutting machine and collected a sugarcane stem node dataset consisting of 870 images with approximately 16,000 stem nodes and divided them into sub-datasets of 700 training, 70 verification, and 100 test images. They used these sub-datasets to train and test the proposed algorithm.
Furthermore, the performance and complexity of the proposed model and other algorithms are evaluated using metrics such as average precision (AP), model size, number of parameters, and floating-point operations (FLOPs).
Research Findings
The results show that G-YOLOv5s-SS architecture achieved the highest AP of 97.6%, which is 0.6% higher than that of the YOLOv5 baseline network and higher than other mainstream one-stage target detection algorithms such as YOLOv4, YOLOv6, and YOLOv717. Meanwhile, the model size, parameters, and FLOPs of G-YOLOv5s-SS were 2.6MB, 1,129,340, and 7.2G, respectively, which are equivalent to 18%, 16%, and 45.6% of YOLOv5s, and much lower than other algorithms.
The proposed novel architecture or algorithm has potential applications in sugarcane cultivation, especially in the mechanization of sugarcane seed production. This can be integrated into the sugarcane seed pre-cutting machines to improve their efficiency and accuracy and reduce their cost and power consumption. Moreover, this can be applied to other agricultural tasks that require object detection, such as fruit identification, crop disease and pest identification, and animal behavior detection.
Conclusion
In summary, the proposed G-YOLOv5s-SS algorithm is a lightweight and efficient architecture for the detection of sugarcane stem nodes. This architecture achieved higher accuracy and performance in sugarcane stem identification compared to other existing mainstream one-stage target detection. The authors suggested that future research could consider identifying sugarcane stem nodes in complex environments, such as without fully peeling off sugarcane leaves, to further improve the model’s robustness and generalization ability.
Deep Learning Transforms Solar Cell Design