A Two-Stage Deep-Learning Model for Detection and Occlusion-Based Classification of Kashmiri Orchard Apples for Robotic Harvesting

Purpose The process of robotic harvesting has revolutionized the agricultural industry, allowing for more efficient and cost-effective fruit picking. Developing algorithms for accurate fruit detection is essential for vision-based robotic harvesting of apples. Although deep-learning techniques are popularly used for apple detection, the development of robust models that can accord information about the fruit’s occlusion condition is important to plan a suitable strategy for end-effector manipulation. Apples on the tree experience occlusions due to leaves, stems (branches), trellis wire, or other fruits during robotic harvesting. Methods A novel two-stage deep-learning-based approach is proposed and successfully demonstrated for detecting on-tree apples and identifying their occlusion condition. In the first stage, the system employs a cutting-edge YOLOv7 model, meticulously trained on a custom Kashmiri apple orchard image dataset. The second stage of the approach utilize the powerful EfficientNet-B0 model; the system is able to classify the apples into four distinct categories based on their occlusion condition, namely, non-occluded, leaf-occluded, stem/wire-occluded, and apple-occluded apples. Results The YOLOv7 model achieved an average precision of 0.902 and an F1-score of 0.905 on a test set for detecting apples. The size of the trained weights and detection speed were observed to be 284 MB and 0.128 s per image. The classification model produced an overall accuracy of 92.22% with F1-scores of 94.64%, 90.91%, 86.87%, and 90.25% for non-occluded, leaf-occluded, stem/wire-occluded, and apple-occluded apple classes, respectively. Conclusion This study proposes a novel two-stage model for the simultaneous detection of on-tree apples and classify them based on occlusion conditions, which could improve the effectiveness of autonomous apple harvesting and avoid potential damage to the end-effector due to the objects causing the occlusion.