Performance Evaluation of Modern Object Detection Models for Automated Fruit Recognition in Smart Agriculture

Abstract

An accurate and reliable image-based fruit detection system is essential for advancing agricultural tasks such as yield mapping and robotic harvesting. This paper benchmarks five state-of-the-art object detection frameworks on a merged dataset of 10 common fruit classes. The challenge of scarcity of high-quality fruit dataset detection of multiple classes of fruit is addressed. Data merging and augmentation techniques are used to increase data samples and capture variability. Multiple architectures are used to explore computational trade-offs of varying deep learning networks. Model accuracy and inference speed are evaluated for considerations of future deployment in an agricultural environment. This study offers guidance on model selection for fruit detection based on accuracy and latency requirements. Future research should evaluate latencies on resource-constrained devices and should prioritize improving accuracy while reducing overall complexity and inference speed for fruit detection tasks.