Reproduce “Classifying Shelf Life Quality of Pineapples by Combining Audio and Visual Features” on OpenResearch

About this paper

Yi-Lu Jiang, Wen-Chang Chang, Ching-Lin Wang, Kung-Liang Hsu, Chih-Yi Chiu

This paper presents a non-destructive method for classifying pineapple quality into four categories by integrating audio and visual features using a cross-modal deep learning approach. The proposed system achieves 84% accuracy, demonstrating that combining tapping sounds and visual data improves shelf-life prediction compared to single-modality methods.

Key insights

• Audio features from tapping sounds, particularly from side tapping with unidirectional microphones, provide more informative data for predicting pineapple quality than visual features alone.
• Cross-modal learning effectively leverages complementary information from both audio and visual modalities, outperforming unimodal and simple ensemble approaches.
• An audio-major sampling strategy during training, combined with a moderate sample size (around 3,200 samples), leads to optimal model performance and efficient training.

Summary by alphaXiv.