Enhanced yolov8 with attention mechanism for tomato ripeness detection in smart agriculture / Jalal Uddin MD Akbar

Faculty of Computing

Thesis (Master of Science) -- Universiti Malaysia Pahang - 2026

Includes bibliographical references

Accurate object detection and classification are crucial in precision agriculture for optimizing yield and quality, particularly for determining tomato ripeness stages. However, automated visual assessment in smart agriculture environments faces challenges like variable lighting, occlusions, and subtle ripeness variations. While state-of-the-art detectors like YOLOv8 are fast, their standard architectures can struggle with the fine-grained feature sensitivity needed for accurate ripeness grading and present a trade-off between performance and the computational efficiency required for real-time, edge deployment. This study addresses these limitations by developing and evaluating YOLOv8-ResCBAM, an optimized model tailored for reliable and efficient multi-stage tomato ripeness detection under challenging conditions. The methodology involved curating and preprocessing a comprehensive tomato dataset with extensive augmentation. Key optimizations included strategically reducing layers in the YOLOv8s architecture to create a lightweight backbone and systematically integrating and evaluating several attention mechanisms (CBAM, GAM, ECA, Shuffle Attention, ResCBAM) with this modified architecture. The Residual Convolutional Block Attention Module (ResCBAM) was identified as the optimal mechanism for enhancing focus on subtle ripeness indicators and improving effectiveness in challenging conditions. Performance evaluation confirmed the superiority of YOLOv8- ResCBAM, achieving impressive mAP@0.5 of 91.2%, a significant 6.8% improvement over the baseline YOLOv8s. The model also demonstrated substantial gains in precision (89.2%), recall (86.5%), and F1-score (87.8%). These accuracy improvements are attributed to the synergy between the lightweight design and ResCBAM’s enhanced feature focus, effectively handling occlusion and lighting variations. Critically, this was achieved alongside high computational efficiency (40 FPS, 7.84M parameters, 17.2 MB size), validating its suitability for real-time, resource-constrained deployment. This research contributes a validated, optimized model balancing accuracy and efficiency for tomato ripeness detection, advancing AI-driven crop monitoring for improved operational efficiency and resource management. Future work includes extending the model to other crops, testing in diverse outdoor conditions, and integration with IoT systems. This study provides an effective, optimized deep learning solution, offering insights into tailoring vision models for specific precision agriculture needs and paving the way for more sustainable farming practices.