Evolutionary CNN-based architectures with attention mechanisms for enhanced image classification
Afsaneh Shams, Kyle Becker, Drew Becker, Soheyla Amirian, Khaled Rasheed · 2024
This extended study builds upon prior research, serving as an extension of our previous study presented in “Evolving Efficient CNN-Based Model for Image Classification” [19]. Here, we delve deeper into Convolutional Neural Network (CNN) architectures and their performance on the CIFAR-10 dataset, expanding upon the insights gained from our earlier work. Beginning with an ECNNB [19] model that excelled on simpler datasets, we progress to examine advanced iterations featuring attention mechanisms like CBAM and MobileViTv2. Our empirical analysis demonstrates consistent performance improvement with each enhancement, culminating in the AECNNB with MobileViTv2 as the most efficient model. Notably, CBAM integration alone led to a substantial 7.27% improvement in average accuracy, and the final AECNNB model achieved an 86.89% average accuracy, marking a significant 9.77% improvement over the ECNNB [19]. This underscores the significance of architectural sophistication and the potential of advanced attention mechanisms, particularly MobileViTv2, for optimizing CNNs in complex image classification tasks. Our findings provide valuable insights for future neural network development and applications, building upon the foundation established in our earlier study.