A Mobile-Centric Comparative Evaluation of Deep Learning Models for Real-Time Flower Recognition

Abstract Flower recognition has important applications in education, agriculture, and intelligent mobile systems. However, previous studies have primarily focused on classification accuracy, with less attention to the joint requirements of real-time mobile deployment. This study presents a mobile-centric flower recognition framework whose technical contribution lies in the dataset-specific adaptation and deployment-oriented evaluation of established learning architectures. Rather than applying an identical transfer-learning configuration to all convolutional models, architecture-specific classification heads and training strategies were developed for VGG16, NASNet-Mobile, and MobileNetV2 according to their learning behaviour on the Oxford 102 Flowers dataset. In addition, a hybrid MobileNetV2–XGBoost pipeline was constructed by combining deep features extracted from a pretrained MobileNetV2 backbone with an independently optimised XGBoost classifier. The four resulting pipelines were evaluated on a common data split using classification performance, inference time, training time, model size, architectural complexity, and deployment feasibility. On the 2,539-image test set, the adapted MobileNetV2 achieved approximately 96% accuracy and a weighted F1-score of 0.96, providing the most suitable overall trade-off for mobile use. It was subsequently converted to TensorFlow Lite, reducing its size from 27.74 MB to 2.53 MB, and integrated into an offline Android application. Repeated measurements on physical Android devices produced mean preprocessing-plus-inference latencies ranging from 55.57 to 68.11 ms. These results demonstrate that dataset-specific model adaptation and deployment-oriented selection can enable accurate and responsive flower recognition without cloud-based processing.