An Innovative Method Based on Dilated CNN for Enhancing Image Classification Accuracy

In the domain of image classification, traditional convolution models are limited in their ability to accurately capture features across different scales within an image. Likewise, dilated convolution models suffer from the issue of generating uniform responses for features across all scales, leading to suboptimal performance in certain image classification tasks. To address these challenges, this paper presents a novel hybrid convolution model, referred to as i-HDC, which integrates traditional convolution models with dilated convolutions. The architecture of i-HDC is introduced and subsequently evaluated on the GTSRB data set. Comparative assessments are conducted against both traditional convolution models and dilated convolution models, while maintaining consistent experimental conditions. The results demonstrate that i-HDC exhibits enhanced convergence speed and superior generalization capabilities compared to its counterparts. Notably, it achieves a remarkable classification accuracy of 99.86% with a reduced training period, thereby significantly elevating the accuracy of classification tasks.