Abstract
Fruit freshness monitoring represents one of the key research foci in the quality control of fruits and vegetables. Traditional manual inspection methods are characterized by subjectivity and inefficiency, which renders them unsuitable for large-scale and real-time detection demands. Automated detection methods based on deep learning have increasingly attracted attention. In this study, a fruit freshness classification method based on the ResNet-101 network and a Non-local Attention mechanism is proposed. By embedding a Non-local Attention module into ResNet-101, subtle surface feature variations of the fruit are captured, thereby enhancing the model's capacity to identify rotten areas and detect variations in color under complex backgrounds. Experimental results show that the improved model achieves a precision of 94.7%, a recall of 94.24%, and an F1-score of 94.24%, outperforming conventional ResNet-101, ResNet-50, and VGG-16 models. In particular, under complex environmental conditions, the model demonstrates significantly improved robustness in image processing. The combination of the Non-local Attention mechanism with the ResNet-101 model can substantially enhance the accuracy and stability of fruit freshness detection, which is applicable to real-time monitoring tasks in intelligent agriculture and smart logistics.