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

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

Problem Statement
While deep learning has significantly advanced flower recognition accuracy, previous research has predominantly focused on classification metrics, often neglecting the joint requirements of real-time mobile deployment. Smartphones possess limited processing power, memory, storage, and energy resources, making the direct application of heavy convolutional neural networks (CNNs) challenging. Furthermore, existing studies often apply uniform transfer-learning configurations across different architectures or select models based solely on accuracy, failing to account for the specific trade-offs between inference speed, model size, training cost, and architectural complexity required for practical, offline mobile use.

Methodology
The study proposes a mobile-centric framework that evaluates four distinct flower recognition pipelines using the Oxford 102 Flowers dataset (8,185 unique images across 102 classes). The methodology diverges from standard approaches by employing dataset-specific and architecture-specific adaptations rather than a one-size-fits-all transfer learning strategy.

1. Dataset and Preprocessing: The dataset was split into training (60%), validation (10%), and testing (30%) sets using stratified sampling to maintain class distribution. Images were resized to 224×224 pixels and normalized. Data augmentation (rotation, shifting, shearing, zooming, flipping, brightness adjustment) was applied exclusively to the training set to prevent overfitting.
2. Model Pipelines: Four candidate pipelines were developed and evaluated under a common experimental protocol:
   • Hybrid MobileNetV2–XGBoost: A frozen MobileNetV2 backbone extracts 1,280-dimensional deep features, which are fed into a separately optimized XGBoost classifier (150 trees, depth 4).
   • VGG16: A deep CNN architecture with a custom classification head (Global Average Pooling, 128-unit dense layer, ReLU) and a two-stage training strategy (initial head training followed by full backbone fine-tuning).
   • NASNet-Mobile: A neural architecture search-derived model with a regularized head (Global Average Pooling, 256-unit dense layer, 0.5 dropout). Fine-tuning was abandoned as it did not yield consistent improvements; the frozen backbone configuration was retained.
   • MobileNetV2: A lightweight architecture using inverted residual blocks. It utilized a compact head (Global Average Pooling, 0.2 dropout) and a two-stage training strategy involving initial head training and subsequent end-to-end fine-tuning.
3. Evaluation Metrics: Models were assessed using weighted precision, recall, and F1-score (to address class imbalance), alongside computational metrics: inference time, training time, model size, and architectural complexity (parameter count and layer count).
4. Mobile Deployment: The best-performing model was converted to TensorFlow Lite (TFLite), optimized, and integrated into an offline Android application. Deployment performance was benchmarked on physical Android devices (Redmi 10C and Redmi Note 13) using 300 repeated measurements per device.

Key Contributions

• Dataset-Specific Adaptation: The study demonstrates that applying identical transfer-learning configurations to different architectures is suboptimal. Instead, it develops specific classification heads, regularization strategies, and fine-tuning schedules tailored to the learning behavior of VGG16, NASNet-Mobile, and MobileNetV2 on the Oxford 102 Flowers dataset.
• Hybrid Pipeline Development: The construction of a MobileNetV2–XGBoost pipeline to investigate the efficacy of combining deep visual features from a lightweight CNN with a non-neural gradient-boosted classifier.
• Comprehensive Comparative Evaluation: A consistent comparison of the four pipelines based on a holistic set of metrics including predictive performance, training/inference time, model size, and deployment feasibility, rather than accuracy alone.
• End-to-End Offline Deployment: A complete workflow from model selection to physical device benchmarking, including TFLite conversion, Android integration, and latency measurement on real hardware.

Results

• Classification Performance: The adapted MobileNetV2 achieved the highest performance with a weighted F1-score of 0.96, precision of 0.96, and recall of 0.96. VGG16 followed with an F1-score of 0.83, while NASNet-Mobile and the hybrid MobileNetV2–XGBoost achieved F1-scores of 0.80 and 0.79, respectively.
• Computational Efficiency: MobileNetV2 offered the best balance, achieving high accuracy with a model size of 27.74 MB and an inference time of 0.00518 seconds. VGG16 was the most resource-intensive (169.40 MB, 0.05570s inference, 360.70 min training). The hybrid pipeline's reported classifier inference time was negligible but excluded the feature extraction cost.
• Mobile Deployment: The selected MobileNetV2 model was converted to TFLite, reducing its size from 27.74 MB to 2.53 MB. The resulting offline Android application achieved mean preprocessing-plus-inference latencies of 55.57 ms (Redmi Note 13) and 68.11 ms (Redmi 10C) across 300 measurements per device, demonstrating responsive real-time performance.

Significance and Claims
The paper claims that its primary contribution lies not in proposing a new backbone architecture, but in the systematic, model-specific adaptation of established pretrained architectures to a target dataset and their deployment-oriented comparative evaluation. The study concludes that accurate and responsive flower recognition can be achieved on mobile devices without cloud-based processing, provided that model selection considers the trade-offs between accuracy, inference speed, and model size. The results validate that dataset-specific adaptation and deployment-oriented selection are critical for enabling effective real-time mobile applications.