YOLO-CMST: Towards accurate pineapple flowering induction using YOLO-based models with the Cross Multi-Style Translator

Abstract

Recently, You Only Look Once (YOLO) models have been widely used in agricultural applications, such as precise flowering induction for pineapples. These models help ensure uniform maturity, optimize harvest schedules, and improve overall quality. However, YOLO models face challenges, particularly the extra effort required to collect diverse datasets and manually label them if we want to enhance model performance. To address these challenges, we propose YOLO-CMST, a solution that integrates the Cross Multi-Style Translator (CMST) and the Intermediate Domain Transformation (IDT) algorithm. This module generates synthetic pineapple images in a variety of styles while ensuring the position of the pineapple core aligns with the original images. As a result, it allows the original labeled files to be reused for synthetic pineapple images, eliminating the need for additional data collection or manual labeling during training. Based on this, we developed a self-propelled pineapple flowering spray machine that autonomously detects pineapple cores using YOLO-CMST and directs the nozzles for precise spraying. Furthermore, to achieve the practical application of the proposed framework, we deployed the model on low-cost, real-time computing systems, such as the Intel Neural Compute Stick (NCS). Experimental results demonstrate that the proposed detection system provides adequate speed for real-world field applications, ensuring both efficiency and reliability in spraying floral inducers. Additionally, the performance of YOLO models can be effectively improved by over 7% in F1-score and 4% in mean Intersection over Union (mIoU) with CMST, while maintaining the same size of the training dataset.