Efficient instance segmentation for strawberry in greenhouses using YOLOv8n-MCP on edge devices

Abstract

The labor cost in agriculture is gradually increasing, making it necessary to develop robots for strawberry picking. These robots require accurate strawberry localization, which remains challenging using machine vision. While instance segmentation can improve positioning accuracy, current algorithms are inefficient on edge computing devices during robot navigation and ineffective for recognizing strawberries in elevated cultivation. This paper proposes an improved YOLOv8n model (YOLOv8n-MCP) optimized for edge computing during robot navigation. The network implements three key improvements: 1) MobileNetV3 as the backbone, enhancing strawberry feature extraction under varied lighting while reducing parameters and GFLOPs; 2) a new Cross-scale Feature Fusion Module (CCFM) as the Neck, improving detection of strawberries at varying distances; and 3) Partial Convolution (PConv) to enhance C2f and Head components, further reducing network parameters and GFLOPs while improving FPS. Experimental results show that compared to YOLOv8n, YOLOv8n-MCP reduces parameters by 69 %, GFLOPs by 56 %, and increases FPS by 42 %. Tests on Nvidia Jetson Xavier NX demonstrate that YOLOv8n-MCP achieves 49.5 FPS, significantly outperforming the original YOLOv8n’s 37.6 FPS, effectively meeting the requirements for strawberry instance segmentation during robot navigation with edge devices.

Abstract from the paper’s OpenAlex record; open copy licensed cc-by-nc-nd. Read at the source ↗

The text the model read, with the source record’s inline markup (italics, sub- and superscripts) rendered and its character entities decoded; quote offsets refer to the raw text.