Multi-scale attention-based feature learning for jujube fruit bruise detection and variety classification

Abstract

Automated fruit classification plays a central role in post-harvest quality assessment, particularly in reducing dependence on manual inspection within smart agriculture systems. In this work, a deep learning framework, MSA-TNet, is introduced to address jujube fruit classification across two related tasks, binary bruise detection and multi-class variety recognition. The study uses two publicly available datasets comprising 1,464 original images for bruise detection and 1,716 original images for variety classification, with provider-supplied augmented versions used exclusively to expand the training partitions. The datasets were partitioned at the original-image level before augmentation to avoid leakage from augmented derivatives across subsets. For comparison, six established transfer learning models, including ResNet50, EfficientNet-B0, DenseNet121, MobileNetV2, ConvNeXt-Tiny, and ConvNeXt-Small, were evaluated under consistent experimental settings. Built upon a pretrained ResNet50 backbone, MSA-TNet integrates hierarchical multi-scale feature extraction, channel-wise attention, and feature fusion, introducing a moderate increase in computational complexity. Training is carried out using the AdamW optimizer, and performance is assessed using accuracy, precision, recall, F1-score, and ROC-AUC. MSA-TNet achieves 99.09% accuracy and macro-F1 for bruise detection and 99.61% accuracy with 99.61% macro-F1 for variety classification. These results indicate competitive performance across the two classification tasks. In terms of computational complexity, MSA-TNet requires 28.4 M parameters and 4.8 GFLOPs, representing a moderate increase over the ResNet50 backbone. The findings demonstrate that integrating hierarchical multi-scale representations with channel-wise attention can provide competitive feature learning performance while maintaining a moderate computational complexity.