Abstract: Long-term accumulation of carbon deposit in car engines can result in decreased engine power, increased fuel consumption, and diminished emission performance, highlighting the critical importance of timely engine detection and cleaning for effectively mitigating the impact of carbon deposit. A novel model based on non-rigid features has been proposed to discriminate the degree of carbon deposit. Firstly, deformable convolution is employed in the model to adjust the offset position and amplitude of the convolution kernel, enhancing the effective receptive field of the network and extracting non-rigid feature information. Subsequently, neurons are weighted based on the correlation of pixels within the kernel region using an adaptive exponential metric pooling kernel to capture more precise feature information. Finally, a feature improvement module based on a self-attention mechanism is incorporated to extract comprehensive contextual information from feature maps. The model’s test accuracy after experimental testing is 87. 1%, and the indexes have increased by 2. 5% on average compared to the original model. Demonstrating the capability of our proposed method to extract effective features for carbon deposit degree discrimination. The approach has the potential to theoretically justify the widespread promotion of the degree model discrimination for carbon deposit.