YOLOv8 vs. SSD for Apple Detection in Agriculture: Performance, Overlap Handling and Volume Estimation

Fruit recognition is of great significance for automated harvesting, monitoring, and counting in agriculture. This paper presents two apple recognition models based on YOLOv8 and Single Shot MultiBox Detector (SSD) algorithms respectively and compares their performance under identical conditions. We annotated an apple image dataset from the Beijing Society of Image and Graphics, trained both models on this dataset, and juxtaposed their loss curves. The accuracy of the two algorithms was calculated and compared, revealing performance variations in scenarios involving aspect ratio changes, target overlap, and small target detection. By evaluating multiple statistical accuracy metrics and recognition speed, we conclude that YOLOv8 outperforms SSD in most cases but lags behind SSD in handling aspect ratio alterations and overlapping targets. This paper briefly analyzes the factors underlying these disparities, providing insights for algorithmic improvements. Additionally, we derived a formula to estimate apple volume using camera parameters and field-of-view data. Our research enhances the efficiency and accuracy of apple detection, demonstrating the potential of deep learning in agricultural automation applications.