GELAN-T3: A Novel Lightweight Convolutional Block for Efficient Sugarcane Stem Health Detection

Primarily a source of sugar and secondarily biofuel, sugarcane remains a top globally traded crop. As sugarcane buds are used for propagation, it is necessary to be selective and use only high-quality sugarcane stems to get a higher yield and maintain the quality of the crop. Manually inspecting stems visually presents a monotonous task that requires subjective assessment. Our research develops an efficient computer vision object detection model that works on edge devices with limited computational resources to identify high-quality sugarcane stems from stem clusters. In this work, we employ a single-shot detection framework based on YOLOv12n. We also design a new convolutional block, GELAN-T3, which uses depth-wise convolution (DWConv) layers to bring in efficiency while maintaining high detection accuracy. We show that our architectural changes lead to a significant drop in cost and size, including reduced computational and memory overhead, and that the model can be efficiently deployed in the field while not losing accuracy. We train the model on a dataset of 3,839 sugarcane stem images collected from different farmlands in India, with natural variations such as cluttered background, illumination variation, occlusion, scale, and structure of the crop. The dataset is self-compiled and annotated for two classes of stems: fit and unfit. Our experimental results show that our model attains an mAP@50 of 89.2% and an mAP@50:95 of 62.6%, while using only 1.55M parameters, 4.6 GFLOPs, and a model size of 3.2 MB. Our model also has extremely low training time (0.69 h) and can do real-time inference with 69.8 FPS. Comparisons between our approach and leading models alongside Vision Transformers and CNN-Transformer combinations demonstrate our model's superior balance of both efficiency and accuracy. The presented model represents a viable solution for scalable, precision agriculture while opening avenues for further robustness improvements under highly complex field conditions. The complete codebase used in this study can be accessed at: https://github.com/NCU-Pushpendra/GELAN-T3-codebase .