Answers are grounded in this paper and cite their sources.
Apple surface defects represent a critical challenge for automated post-harvest inspection, as they directly affect market value and consumer acceptance. This study aims to enhance defect detection performance by analyzing convolutional neural network architectures applied to multispectral images from the CAPA dataset. The methodology involved benchmarking baseline and attention-enhanced CNN models, including DenseNet121, MobileNetV1 with SE, and MobileNetV2 with SE, under both binary (healthy versus defective) and multiclass classification schemes. The experiments included spectral band ablation to assess the contribution of individual wavelengths and a cascaded design integrating binary and multiclass stages. Results show that MobileNetV2 combined with squeeze and excitation modules and fine-tuning achieved the best performance, reaching 96.9% accuracy and an F1-score of 0.956 in multiclass classification. The green spectral band proved to be the most discriminative, while red-edge and infrared bands provided complementary information. In conclusion, the proposed two-stage CNN framework improves accuracy within the evaluated CAPA dataset conditions.
No takes yet. Share an insight, caveat, or question.
Vintimilla et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: