Yapay Zeka Hızlı Bakış & Çıkarımlar
AI InsightThe results demonstrate that the proposed ResNet-18 + LDA framework, strengthened by the feature-grouping strategy, provides high accuracy, strong class discrimination, and computational efficiency for multi-class image classification
📌 3 Maddede Bu Makale
Abstract
In this study, a hybrid classification approach combining ResNet-18 and Linear Discriminant Analysis (LDA) is proposed for eight-class animal image recognition. ResNet-18 is used as a pretrained feature extractor, producing 1,000-dimensional deep features that capture high-level visual patterns from each image. Unlike conventional methods that use these features directly, the study introduces a feature-grouping strategy in which the 1,000 features are reorganized into structured sequences of 30–60 elements to create more balanced and discriminative input vectors. These grouped features are then classified using LDA, which reduces dimensionality and enhances class separability. Experiments on an eight-class animal dataset (cat, cow, deer, dog, goat, hen, rabbit, sheep) achieve an overall accuracy of 95.6%. Analysis of the confusion matrix shows minimal misclassification, mainly between visually similar categories, while t-SNE visualizations confirm clear separations among classes. The results demonstrate that the proposed ResNet-18 + LDA framework, strengthened by the feature-grouping strategy, provides high accuracy, strong class discrimination, and computational efficiency for multi-class image classification