Open Access

ARTICLE

An Enhanced Image Classification Model Based on Graph Classification and Superpixel-Derived CNN Features for Agricultural Datasets

Thi Phuong Thao Nguyen¹, Tho Thong Nguyen¹, Huu Quynh Nguyen², Tien Duc Nguyen³, Chu Kien Nguyen¹, Nguyen Giap Cu^4,*

1 Faculty of Computer Science and Engineering, Thuyloi University, 175 Tay Son, Dong Da, Hanoi, 100000, Vietnam
2 Faculty of Information Technology and Communication, CMC University, 11 Duy Tan, Dich Vong Hau, Cau Giay, Hanoi, 100000, Vietnam
3 Information Technology Center, Thuyloi University, 175 Tay Son, Dong Da, Hanoi, 100000, Vietnam
4 Faculty of Economic Information System and E-Commerce, Thuongmai University, Ho Tung Mau, Cau Giay, Hanoi, 100000, Vietnam

* Corresponding Author: Nguyen Giap Cu. Email:

(This article belongs to the Special Issue: New Trends in Image Processing)

Computers, Materials & Continua 2025, 85(3), 4899-4920. https://doi.org/10.32604/cmc.2025.067707

Received 10 May 2025; Accepted 17 July 2025; Issue published 23 October 2025

Abstract

Graph-based image classification has emerged as a powerful alternative to traditional convolutional approaches, leveraging the relational structure between image regions to improve accuracy. This paper presents an enhanced graph-based image classification framework that integrates convolutional neural network (CNN) features with graph convolutional network (GCN) learning, leveraging superpixel-based image representations. The proposed framework initiates the process by segmenting input images into significant superpixels, reducing computational complexity while preserving essential spatial structures. A pre-trained CNN backbone extracts both global and local features from these superpixels, capturing critical texture and shape information. These features are structured into a graph, and the framework presents a graph classification model that learns and propagates relationships between nodes, improving global contextual understanding. By combining the strengths of CNN-based feature extraction and graph-based relational learning, the method achieves higher accuracy, faster training speeds, and greater robustness in image classification tasks. Experimental evaluations on four agricultural datasets demonstrate the proposed model’s superior performance, achieving accuracy rates of 96.57%, 99.63%, 95.19%, and 90.00% on Tomato Leaf Disease, Dragon Fruit, Tomato Ripeness, and Dragon Fruit and Leaf datasets, respectively. The model consistently outperforms conventional CNN (89.27%–94.23% accuracy), VIT (89.45%–99.77% accuracy), VGG16 (93.97%–99.52% accuracy), and ResNet50 (86.67%–99.26% accuracy) methods across all datasets, with particularly significant improvements on challenging datasets such as Tomato Ripeness (95.19% vs. 86.67%–94.44%) and Dragon Fruit and Leaf (90.00% vs. 82.22%–83.97%). The compact superpixel representation and efficient feature propagation mechanism further accelerate learning compared to traditional CNN and graph-based approaches.

---

Keywords

---