Optimizing Loss Functions for You Only Look Once Models: Improving Object Detection in Agricultural Datasets

Japan faces a significant labor shortage due to an aging population, particularly in the agricultural sector. The rising average age of farmers and the declining participation of younger individuals threaten the sustainability of farming practices. These trends reduce the availability of agricultural labor and pose a risk to lowering Japan's food self-sufficiency rate. The reliance on food imports raises concerns regarding price fluctuations and sanitation standards. Moreover, the challenging working conditions in agriculture and a lack of technological innovation have hindered productivity and increased the burden on the existing workforce. To address these challenges, "smart agriculture" presents a promising solution. By leveraging advanced technologies such as sensors, drones, the Internet of Things (IoT), and automation, smart agriculture aims to optimize farm operations. Real-time data collection and AI-driven analysis play a crucial role in monitoring crop growth, assessing soil conditions, and improving overall efficiency. This study proposes enhancements to the YOLO (You Only Look Once) object detection model to develop an automated tomato harvesting system. This system uses a camera to detect tomatoes and assess their ripeness for harvest. Our objective is to streamline the harvesting process through AI technology. Our improved YOLO model integrates two novel loss functions to enhance detection accuracy. The first, "VSR", refines the model's ability to classify tomatoes and determine their harvest readiness. The second, "SBCE", enhances the detection of small tomatoes by training the model to recognize a range of object sizes within the dataset. These improvements have significantly increased the system's detection performance. Our experimental results demonstrate that the mean Average Precision (mAP) of YOLOv7-tiny improved from 61.81% to 70.21%. Additionally, the F1 score increased from 0.61 to 0.71 and the mean Intersection over Union (mIoU) rose from 65.03% to 66.44% on the tomato dataset. These findings underscore the potential of our proposed system to enhance efficiency in agricultural practices.