Metals are used in almost every sector of human life, from household appliances and transportation equipment to high skyscrapers. One of the main problems that associated with metals is their corrosion. Corrosion degrades metal's quality. This process decreases the strength of the metal structure and changes the color of its appearance. Early corrosion detection in metals is essential to ensure appropriate treatment against corrosion. Certified experts usually do corrosion classification. Limited resources of experts are one of the problems of corrosion detection. For this reason, it is necessary to have a method that does not depend on experts for corrosion detection and classification which uses artificial intelligence (AI). One of the main parameters to measure image classification performance is accuracy. However, accuracy in Convolutional Neural Networks (CNN) is a challenging problem if there is a limited number of images dataset with slight differences between each class. This research aims to improve the performance of CNN for the classification of corrosion images, including increasing accuracy. ResNet-18, as a pretrained CNN model, is used in this research. Some strategies were done, such as preprocessing and augmentation dataset, transfer learning, fine-tuning and modifying ResNet-18 network architecture, and tuning the hyperparameter of ResNet-18. We use the corrosion dataset, which has multiple classes. It has 600 different images that are divided into five different classes. Each class has 120 images. This research reveals that the proposed method increases accuracy from 72,22% to 93.89%, precision score of 0.94, recall of 0.94, and F1-score of 0.94. © 2025 IEEE.