Hybrid Multi-View Embedded Vision for Functional Quality Assessment of Cut Roses Using a Raspberry Pi

Abstract

Manual grading of cut roses is affected by operator subjectivity, fatigue, and variability in theinterpretation of commercial quality attributes. This study presents a low-cost hybrid multi-viewmachine-vision system for functional rose quality assessment implemented on a Raspberry Pi 4. Threetask-specific cameras were used to acquire complementary views: a lateral view for flower-headmorphology and petal condition, an upper view for flower-head diameter estimation, and a frontalview for stem color and apparent length analysis. The processing pipeline combines an eight-classconvolutional neural network (CNN) with classical image-processing methods and rule-baseddecision logic. The CNN was developed from more than 600 manually labeled RGB images resized to128 × 128 pixels and trained using Adam and categorical cross-entropy. Func-tional testing yielded95.3% accuracy for dominant-color classification and 91.7% accuracy for flower-opening assessment.Detected flower-head diameters ranged from 4.6 to 6.9 cm, while the integrated prototype operatedcontinuously for more than four hours. These results demon-strate the functional feasibility of multiview embedded inspection under controlled acquisition conditions. Further work is required toestablish independent class-wise performance, calibrated physical measurements, inference latency,and agreement with documented commercial grading criteria.