• A large-scale dataset was developed, including 600 orange samples with detailed physicochemical measurements. • Twelve deep learning–based object detection models and four machine learning regression models were benchmarked. • An improved lightweight YOLOv8s model was designed by optimizing hyperparameters and enhancing detection layers. • A hybrid framework was established by integrating YOLOv8s with Gradient Boosting Regressor (GBR) for multi-task quality assessment. • The lightweight hybrid model achieved 96% mAP@50 and can accurately identify ripeness stages and defect types in real time. • The optimized model was successfully deployed as an Android mobile application for on-site orange quality monitoring. Estimating yield and automating harvesting in orchards with complex growth patterns, variable environments, uncertain climates, and limited planting information are significant challenges. Deep learning has been increasingly used to improve agricultural tasks. The main aim of this study is to develop intelligent algorithms and design a user interface tool for the accurate and rapid detection of oranges. This work proposed a hybrid deep learning-based object detection and machine learning regression framework for real-time orange ripeness and quality assessment. A dataset of 600 oranges collected from multiple provinces and conditions was analyzed, incorporating initial image processing and physicochemical measurements (TA, TSS, maturity index, vitamin C, and total sugar). Imaging was performed using an adjusted computer vision system, and handcrafted features were extracted to support machine learning-based prediction of internal quality attributes. After comparing and optimizing machine learning models, gradient boosting regression (GBR) achieved the best performance for predicting orange properties, with R², RMSE, and MAE of 99%, 0.17, and 0.24, respectively. Subsequently, 12 object detection algorithms were assessed on improved image datasets with 16000 images. To reduce computational complexity while maintaining high inference speed and real-time processing, a lightweight YOLOv8s was developed by optimizing hyperparameters through an ablation study and integrating feature pyramid network layers to improve the model's performance at detecting small objects. Improved YOLOv8s successfully predicted 5 main maturity levels and 3 defect classes in oranges, achieving 96% mAP50 and 75% mAP50:95, with response and training times of 8 milliseconds and 4.9 hours, respectively. Then, improved YOLOv8s are coupled with optimized GBR and converted to TFLite for designing intelligent applications. The Android application can predict properties, ripeness level, and quality management with high precision.
Bohlol et al. (Sun,) studied this question.