Detection and Segmentation of Grape Bunch by Integrating Channel Attention and Large Kernel Attention

Detecting and segmenting fruits in an orchard environment is a vital technique in multiple applications of precision agriculture, such as automated harvesting and yield estimation. This study aims to improve the accuracy and robustness of detectors for detecting and segmenting fruits on trees by integrating detection and segmentation models based on channel and large kernel attention. The proposed Triplet-Large Kernel Attention (TLKA) module inherits the advantages of channel and large kernel attention. It was integrated with YOLOv7 to achieve real-time object detection and segmentation. Several experiments were conducted to verify the effectiveness of the proposed TLKA module. These included comparative attention mechanisms from small to large input image scales, comparative analysis of different attention mechanisms through Grad-CAM visualization, and test experiments with integrated comparisons on mid-term fruits (immature, intermediate, and mature) under three different light conditions (morning, noon, and afternoon). The proposed TLKA module achieved higher accuracy than comparative attention at different input scales. Finally, the proposed model was used to predict the yield of both grapes. The TLKA-YOLOv7 outperformed all other investigated models in terms of grape bunch detection and segmentation and obtained more competitive results in yield prediction.