Key points are not available for this paper at this time.
The aim of this study is to employ deep learning models for microstructural image classification. The microstructures provide valuable insights into material's history and properties, but manual classification is time consuming, labor-intensive and requires expertise, as these images are particularly challenging and complicated. The research experiment involves evaluating three deep Convolutional Neural Networks (CNNs): AlexNet, GoogLeNet, and SqueezeNet. These models were trained and tested on five colored microscopic image datasets differs in classes number, image sizes and quantities. The datasets were obtained from experimental heat treatments conducted under different conditions on AISI 4140 low alloy steel specimens. The comparison focused on accuracy, elapsed time for training, and the impact of class numbers and characteristics on classification performance. Results demonstrated high accuracies ranging from 86.76% to 98.33%, with SqueezeNet showing superior performance for this task. Faster training intervals were recorded for the dataset with lower classes and quantity of images.
Ibrahim et al. (Sat,) studied this question.