Dynamic graph-enhanced deep network for robust infected fruit detection through adaptive CNN-GNN feature fusion

Automated detection of infected fruit remains challenging due to complex lesion distributions, variable illumination, and partial occlusions, which limit the effectiveness of conventional convolutional models that rely primarily on local receptive fields. This study aims to improve infected fruit detection by incorporating spatial relational reasoning while maintaining computational efficiency for practical deployment. A hybrid deep learning framework is proposed that integrates convolutional feature extraction with graph-based relational modeling. A transfer-learned convolutional neural network generates spatial feature representations, which are transformed into a graph structure encoding feature similarity and contextual proximity. A graph neural network with attention-based message passing captures long-range spatial dependencies across fruit surfaces. An adaptive fusion mechanism assigns sample-dependent weights to convolutional and graph-based predictions based on feature complexity, enabling dynamic decision integration. Experimental evaluation on a public fruit quality dataset demonstrates that the proposed framework achieves superior classification performance compared with representative transfer learning and handcrafted feature-based approaches. The model attains a validation accuracy of 96.3% and consistently improves precision, recall, and F1-score across multiple fruit categories. By unifying local feature extraction with relational spatial reasoning through adaptive fusion, the proposed method provides a robust solution for infected fruit detection under diverse visual conditions. The framework offers practical potential for automated quality assessment in horticultural and agricultural supply chain applications.