Research on Fruits Image Target Detection Based on the Fusion Technology of ResNet-18

Abstract

China, as a major producer of various fruits, faces significant challenges in efficiently identifying fruits during sorting and processing, particularly when dealing with large-scale datasets. Traditional methods, such as YOLOv5, may incur high computational costs in terms of time and space when applied to extensive datasets. To address this issue, this study employed ResNet-18, a deep convolutional neural network (CNN) known for its robust feature extraction capabilities and pre-trained on large-scale image datasets. The approach involved constructing a ResNet-18 model to extract image features from a dataset of harvested fruit images. The model was trained to identify various fruits and subsequently applied to recognize fruits in test datasets. ResNet-18 leverages residual connections to overcome gradient vanishing problems, enabling the effective capture of complex features from fruit images. Experimental results demonstrated that the ResNet-18 model achieved a recognition accuracy of 88.6% for fruit identification. This high accuracy highlights the efficiency of ResNet-18 in handling fruit recognition tasks, significantly reducing computational costs while improving detection accuracy. The findings indicate that ResNet-18 provides a robust solution for fruit recognition in orchard management and robotic picking systems, offering a promising alternative to traditional methods in large-scale agricultural applications.