Progressive optimization of EfficientNetV2 for classification based on images of corroded objects

Abstract Considering that corrosion is a widespread problem in tropical countries, this study proposes a progressive optimization of EfficientNetV2 for images of corroded objects (corrosion dataset) that can effectively target small and medium-sized corrosion datasets for detection. Compared to other models, the proposed model first adopts EfficientNetV2 as the basic architecture, focusing on the use of only MBConv blocks and MBConv with Fused-MBConv blocks in the hidden layers, as well as the effect of the number of these layers on the model’s classification results. To further improve the performance, this paper attempts to replace the convolutional modules in the input layer with LazyConv, utilizing FReLU and Dy-ReLU as an activation functions in both the input and output layers. The simulation results show that for the medium-sized corrosion dataset in this paper which uses only MBConv blocks for EfficientNetV2 can achieve higher accuracy but lower computational efficiency. Setting a smaller number of layers and replacing the convolutional block in the input layer with LazyConv can significantly reduce the total size of the model and make it more flexible, where the total size of the obtained M2 model being only 58.98 MB, and capable of automatically determining the number of input channels. Using FReLU in the input and output layers can achieve greater stability, with standard deviations of F1-score and Accuracy under five cycles of only 0.0099 and 0.0126, respectively. In addition, the optimized M2 model also offers advantages in terms of both light weight and stability compared to other classic deep learning models. These findings from this study may serve as a foundation for future innovations in the design of corrosion classification models.