EDT-YOLOv8n-Based Lightweight Detection of Kiwifruit in Complex Environments

Abstract

Automated kiwi harvesting hinges on the seamless deployment of a detection model and the accurate detection of kiwifruits. However, practical challenges, such as the limited computational resources on harvesting robots and occlusions among fruits, hinder the effectiveness of automated picking. To address these issues, this paper introduces EDT-YOLOv8n, a lightweight and efficient network architecture based on YOLOv8n. The proposed model integrates the Effective Mobile Inverted Bottleneck Convolution (EMBC) module to replace the C2f modules, mitigating the channel information loss and bolstering generalization. Additionally, the DySample upsampler, an ultra-lightweight and effective dynamic upsampler, improves feature extraction and resource efficiency when compared to traditional nearest-neighbor upsampling. Furthermore, a novel Task Align Dynamic Detection Head (TADDH) is implemented, incorporating group normalization for a more efficient convolutional structure and optimizing the alignment between the classification and localization tasks. The experimental results reveal that the proposed EDT-YOLOv8n model achieves higher precision (86.1%), mAP0.5 (91.5%), and mAP0.5-0.95 (65.9%), while reducing the number of parameters, the number of floating-point operations, and the model size by 15.5%, 12.4%, and 15.0%, respectively. These improvements demonstrate the model’s effectiveness and efficiency in supporting kiwifruit localization and automated harvesting tasks.