Computing and Informatics

The detection of carbon deposit degree is of great significance to the maintenance of automobile engine. Due to issues with poor feature aggregation, inter-class similarity, and intra-class variance in carbon deposit data with a small number of samples, model-based discriminative approaches cannot be widely implemented in the market. In order to overcome this technical barrier, the article examines the impact of DCNNs (Deep Convolutional Neural Networks) level on the recognition effect of the degree of carbon deposit, introduces a dropout structure and data enhancement strategy to lower the risk of overfitting brought on by the small dataset, and suggests a recognition method based on the kernel of dual-dimensional multiscale-multifrequency information features to enhance the differentiation characteristic. After experimental testing, the accuracy of this method is 86.9 %, the F1-score is 87.2 %, and the inference speed is 190 FPS, which can meet the practical requirements and provide basic support for the large-scale promotion of the model discrimination.