FruitVision: Dual-Attention Embedded AI System for Precise Apple Counting Using Edge Computing

Abstract

In this work, we developed and enhanced an artificial intelligence (AI)-centered hardware framework. This framework integrates the Nvidia Jetson Nano processing unit with a Depth AI camera. Our primary goal was to create an improved version of the YOLOv7 algorithm to quantify apple fruits using edge computing resources. We curated a dataset of 9,000 images of apple fruits to support this effort. Within the enhanced YOLOv7 architecture, we introduced a novel dual attention mechanism called the Global-SE Unified Attention Mechanism (GSEAM). This mechanism was designed to improve the accuracy of object detection by combining spatial and channel-oriented attention mechanisms, significantly enhancing the model.s contextual understanding and object recognition in various settings. The incorporation of GSEAM, along with the Gaussian Error Linear Unit activation function, was a deliberate effort to boost the YOLOv7 architecture.s ability to capture intricate contextual details and hierarchical features. Our system.s performance was rigorously evaluated across six key performance metrics and compared with other pretrained models. We achieved a precision of 99.54%, recall of 98.94%, F1-score of 99.71%, and average precision of 99.13%. This system has proven to be a valuable tool for real-time apple fruit counting, with practical applications for farmers.