Comparative Analysis of Pre-Trained CNN Backbones and Machine Learning Techniques for High Similarity in Circular Object Image Classification

This research presents a comprehensive comparative analysis of various pre-trained backbone models and machine learning techniques for output layers in convolutional neural networks (CNNs) applied to the classification of circular objects with high similarity. Given the challenges associated with distinguishing visually similar objects, the study evaluates a diverse set of backbone models, including AlexNet, ConvNeXt (Base, Large, Small, Tiny), DenseNet (121, 161, 169, 201), EfficientNet (B0 through B7), EfficientNet-V2 (L, M, S), GoogLeNet, Inception-V3, and MaxViT-T. For the output layers, six machine learning models-Decision Tree, K-Nearest Neighbors (KNN), Logistic Regression, Naive Bayes, Random Forest, and Support Vector Machine (SVM)-are tested to identify their effectiveness in conjunction with the backbone models. The methodology involves fine-tuning the backbone models on a dataset of circular objects and training each machine learning model on the features extracted by these backbones. Performance is evaluated using metrics such as accuracy, precision, recall, and F1-score, with cross-validation ensuring robust results. The analysis reveals that ConvNeXt Large and EfficientNet-B5 are the most effective backbone models for this task, offering superior feature extraction capabilities and consistent performance across all metrics. Among the machine learning models, Logistic Regression and Random Forest demonstrate the highest performance, providing reliable and accurate classification. The optimal combination for this classification task is found to be ConvNeXt Large or EfficientNet-B5 as the backbone model with Logistic Regression or Random Forest as the output layer. This combination offers the best balance of classification accuracy and processing efficiency, making it ideal for automated systems dealing with circular objects with high similarity. The findings offer valuable insights for optimizing image classification systems in industrial applications, enhancing both automation and quality control processes.