Improving mango cold-damage and bruise detection using thermal imaging and flexible spectral sensing

Abstract

Influenced by chemical processes such as respiration, ethylene production, and oxidation reactions, fruits are prone to cold damage and bruising during storage and transportation. To this end, this paper presents an ensemble learning network based on dual-modal data fusion for real-time detection and assessment of mango quality. On one hand, we developed a non-destructive classification system that integrates thermal imaging technology with flexible visible/near-infrared spectral sensing, capturing the physicochemical characteristics of mangoes both on the surface and internally. On the other hand, the CNN-ATT-BiLSTM-based ensemble learning framework effectively enables data fusion from the feature layer to the decision layer. Results indicate that mangoes can be classified into four categories: bruised, cold-damaged, both cold-damaged and bruised, and normal. The CNN-ATT-BiLSTM network achieved an overall accuracy of 95.24%. Experiments on grading pipelines demonstrated the system's strong stability. This research contributes new methodologies to the development of fruit quality assessment technologies.