A Fruit Detection Model in Complex Environments Based on Attention Mechanism and Feature Fusion

In complex orchard environments, harvesting robots face significant challenges in identifying fruits accurately due to fruits of various sizes, varying degrees of sunlight intensity, and occlusion. To resolve this issue, we propose a fruit detection model in this paper, which is called AMMFF-YOLO and based on the YOLOv8s architecture. Firstly, a small target detection layer is introduced to enhance the model's ability to identify small-sized fruits. Secondly, a light-adaptive module is proposed, which increases spatial dimension information without reducing the channel dimension, enriching the model's perception ability under different sunlight intensities. This module is named C2f-CS, which stands for C2f module based on channel and spatial (CS) attention mechanisms. Furthermore, a context information enhancement module (CIEM) is proposed, which employs dilated convolutions and skip connections to fuse feature information from multiple scales, enlarges the receptive field, and improves the model's ability to recognize target fruits and backgrounds under occlusion. Lastly, the Inner-MPDIoU Loss is used as a bounding box regression loss function to improve the mod-el's localization ability and detection accuracy. In the detection experiments of strawberries, apples, and citrus in complex orchard environments, the detection accuracy of the proposed model is higher than the detection accuracy of other state-of-art models 0.4 % ~ 8.8 %. The experimental results validate the effectiveness of AMMFF-YOLO in fruit detection under complex orchard environments.