Application of an improved YOLOv9 algorithm for stem node detection in high-resolution global sugarcane images

This study aims to address the problem of poor stem node detection and recognition accuracy in high-resolution global sugarcane images, which is crucial for accurately identifying sugarcane stem node locations during automated sugarcane cutting. Since the sugarcane stem node occupies a very small area of the entire image, only about 0.2%, the feature blurring problem caused by scaling the original image to fit the network input size requirements often results in the loss of important information. In order to overcome this problem, this study conducts an in-depth study on the problem of small target feature loss caused by high-resolution images and proposes an improved algorithm STMFF-YOLOv9 (small target multiple feature fusion YOLOv9) based on YOLOv9-c. The algorithm uses a multi-scale stitching strategy to enhance the feature acquisition of small objects in image data training, designs an FDS (fusion down sampling) module with an improved feature fusion network downsampling process to reduce the loss of small target features, and implements a SE channel attention mechanism to recalibrate the fused channel features, thereby strengthening the focus on useful features. Experimental results showed that STMFF-YOLOv9 significantly improves the detection accuracy of sugarcane stem nodes in global images. Compared to YOLOv9-c, it achieves an 8.3% increase in precision, a 13.1% increase in recall rate, a 10.3% increase in mAP0.5, and a 9% increase in mAP0.5:0.95. Compared to other models, STMFF-YOLOv9 also demonstrates superior detection performance, effectively proving its capability to detect global sugarcane stem nodes.