YOLOv8-STEM: ENHANCED OVERHEAD APPLE STEM DETECTION UNDER OCCLUSIONS

Accurate detection of apple stems is crucial for robotic cutting. This study proposed an improved YOLOv8-stem method for apple stem detection in overhead imagery under occlusion conditions. First, several im-provements were made to the YOLOv8 neural network: the conventional convolutional process within the in-termediate neck layer was substituted with the AK Convolution mechanism, a small object detection head was added, and ResBlock+CBAM attention mechanism was incorporated. Second, stem occlusion was determined by analyzing the positional relationship between the detected bounding boxes of stems and apples. The exper-imental results showed that compared to the original YOLOv8, this method improved apple stem detection ac-curacy by 6.0% (from 79.9% to 85.9%) and increased harvesting completeness from 84.2% to 93.2%.