FloraFusionNet: A Discrete Hybrid CNN–Transformer Framework with Attention Optimization and Comparative Evaluation for Fine-Grained Flower Classification

Abstract Categorization of flowers is important for numerous application areas, such as biodiversity monitoring, agricultural automation and botanical research. Deep learning has introduced a number of new neural network architectures which have shown good performances to more difficult image classification problems. In this paper, we provide a detailed comparative analysis among deepest deep learning networks for multi-class flower recognition task on the Oxford 102 Category Flower Dataset which is recognized benchmark in terms of high intra-class variability and interclass similarity.We test and compare how different deep architectures, that is traditional CNNs (VGG16, ResNet50 and EfficientNet-B0) versus the recent ViT and Hybrid CNN-Transformer perform. The evaluation is based on various performance metrics including overall classification accuracy, precision, recall, F1-score, model complexity (number of parameters), training efficiency and inference latency. All models are fine-tuned based on generalized transfer learning strategy and we also make it strictly uniform to provide a fair comparison of different approaches. Experimental results show that hybrid models so created, where the model can be seen as stacked CNN based encoder with attention-based transformer encoder outperforms individual component achieving peak accuracy of 97.85% on extended test set. However, CNN-based models have faster convergence and less computing cost which are suitable for resource-limited scenarios. This research presents a unified and comprehensive understanding of the trade off among model performance, efficiency, simplicity, and accessibility, which is worth referring to when deploying deep learning solutions into practical botanical classification systems.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-17
DOI
https://doi.org/10.21203/rs.3.rs-8496596/v1
Primary Topic
Advanced Neural Network Applications
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

FloraFusionNet: A Discrete Hybrid CNN–Transformer Framework with Attention Optimization and Comparative Evaluation for Fine-Grained Flower Classification

Sudeep Varshney, Ankur Choudhary, Abhilasha Varshney
Research Square
Advanced Neural Network Applications
preprint

FloraFusionNet: A Discrete Hybrid CNN–Transformer Framework with Attention Optimization and Comparative Evaluation for Fine-Grained Flower Classification

Sudeep Varshney, Ankur Choudhary, Abhilasha Varshney
preprint en

Abstract

Abstract Categorization of flowers is important for numerous application areas, such as biodiversity monitoring, agricultural automation and botanical research. Deep learning has introduced a number of new neural network architectures which have shown good performances to more difficult image classification problems. In this paper, we provide a detailed comparative analysis among deepest deep learning networks for multi-class flower recognition task on the Oxford 102 Category Flower Dataset which is recognized benchmark in terms of high intra-class variability and interclass similarity.We test and compare how different deep architectures, that is traditional CNNs (VGG16, ResNet50 and EfficientNet-B0) versus the recent ViT and Hybrid CNN-Transformer perform. The evaluation is based on various performance metrics including overall classification accuracy, precision, recall, F1-score, model complexity (number of parameters), training efficiency and inference latency. All models are fine-tuned based on generalized transfer learning strategy and we also make it strictly uniform to provide a fair comparison of different approaches. Experimental results show that hybrid models so created, where the model can be seen as stacked CNN based encoder with attention-based transformer encoder outperforms individual component achieving peak accuracy of 97.85% on extended test set. However, CNN-based models have faster convergence and less computing cost which are suitable for resource-limited scenarios. This research presents a unified and comprehensive understanding of the trade off among model performance, efficiency, simplicity, and accessibility, which is worth referring to when deploying deep learning solutions into practical botanical classification systems.

Research Square
Accenture (Switzerland) (CH), Accenture (United States) (US), Sharda University (IN)
Advanced Neural Network Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon