Regular detection and cleaning of carbon deposits in automotive engines can help maintain normal vehicle operation and reduce the negative impact on the environment. However, traditional detection methods are unable to meet the needs of large-scale automotive inspection. This study employs deep learning techniques for the intelligent detection of internal engine carbon deposits captured by endoscopes. However, many challenges including large intra-class variance, limited inter-class variance, and small sample sizes in certain categories are induced by the dataset of carbon deposits poses. Traditional data augmentation methods tend to obscure key semantic information, while the classification efficiency of conventional models remains low. In this study, an Adaptive Intra-class Part Swapping (A-INPS) method was introduced to expand the dataset by randomly substituting key discriminative regions within the same class using Grad-CAM. Additionally, a Feature Refinement Attention Module (FRAM)-RepVGG model was proposed, leveraging RepVGG to achieve lossless compression and enhance detection efficiency. The FRAM incorporated in the model enhances focusing and reduces redundant feature extraction in terms of space and channel. The experimental testing results show that the algorithm has an accuracy of 91.59% and a superior detection speed of 57.55FPS on GTX1080Ti, which is more suitable for deployment in real industrial environments.
Xiang et al. (Tue,) studied this question.