Deep CNN-Based Grape Image Classification with Enhanced Feature Representation

Abstract

Classifying grape varieties based on images of their leaves is a delicate and challenging task in image classification due to the significant similarity in shape, texture, and color among different varieties. This research proposes a hybrid framework combining machine learning and deep learning to achieve high classification performance in real-world environments. The task is formulated as a problem of multiclass classification, including eight grape classes, using a balanced dataset of 8,000 images collected from real-world agricultural environments. Preprocessing steps were implemented to improve the quality of features, including removing the background to separate the leaf from unrelated elements and edge detection to highlight the details of the structure, such as contours and veins. Deep features are then extracted by multiple pre-trained CNN models, like InceptionV3, ResNet50, VGG16, and VGG19, allowing the model to capture balancing visual representations. Principal Component Analysis is used to remove redundancy and decrease feature dimensionality, then resulting features are fused into a unified representation. Some machine learning classifiers were evaluated, including K-Nearest Neighbors, SVM, Logistic Regression, Random Forest, Decision Tree, and Naïve Bayes. The SVM model demonstrates the finest performance (71.6% accuracy) with balanced precision, F1-score, and recall. Although this accuracy is less than the accuracy described on controlled datasets, it reflects real-world data complexity. Generally, the proposed approach explained that combining deep feature extraction, dimensionality reduction, and feature fusion can efficiently enhance the performance of classification in real-world agricultural scenarios.