Computer Vision for Multi-Category African Plum (Safou) Defect Detection

Résumé

The African plum (Dacryodes edulis, Safou) is a nutritionally and economically significant crop across West and Central Africa, yet it suffers from post-harvest losses of 40-50% due largely to the absence of standardised quality assessment systems. This thesis investigates the application of state-of-the-art computer vision architectures to the automated multi-category classification of Safou fruit quality, leveraging an expert-annotated dataset of 4,507 images across six quality grades : Unaffected, Bruised, Cracked, Rotten, Spotted, and Unripe, collected across three agro-ecological regions of Cameroon. A comprehensive benchmark is conducted over twelve architecture families, spanning classical CNNs and modern attention-based models, all trained via ImageNet transfer learning and evaluated using accuracy, weighted and macro F1 scores, and per-class confusion matrix analysis. Results show that most models achieve accuracies in the range of 65%-71%, with ResNet-50, DenseNet121 and EfficientNet-B0 as top performers at 71.01%, 70.27% and 70.71% respectively, while ViT-B/16 and ConvNeXt Tiny lag behind at 60.65% and 62.28%, reflecting their higher data requirements. Consistently across all architectures, Unaffected and Unripe achieve the highest recall, while Bruised remains the most challenging class due to its visual overlap with adjacent defect categories. These findings underscore the intrinsic difficulty of fine-grained Safou classification and motivate future work on dataset expansion, data augmentation, and class imbalance mitigation, constituting the first systematic multi-architecture benchmark for this commercially important African crop.

Citer ce document

Accès au document

Ce document est hébergé sur son dépôt institutionnel d'origine.

Auteur(s)

Statistiques

Consultations : 5

Téléchargements : 0