Car engine sounds recognition based on deformable feature map residual network.

Area of Science:

• Acoustics
• Machine Learning
• Signal Processing

Background:

• Recognizing car engine sounds is challenging due to difficulties in extracting features from time-frequency images.
• Existing methods, including dictionary learning and convolutional neural networks, have limitations in accurately identifying sounds under various operating conditions.

Purpose of the Study:

• To propose a novel method for car engine sound recognition using a deformable feature map residual network.
• To enhance the extraction of salient features from time-frequency representations of engine sounds.
• To improve the accuracy of car engine sound classification.

Main Methods:

• A deformable feature map residual network incorporating offset and convolutional layers was developed.
• Offset layers adaptively shift feature map pixels, focusing on regions of interest.
• Extracted features were fused with Mel frequency cepstral coefficients and refined using a squeeze and excitation block before classification.

Main Results:

• The proposed method achieved an accuracy of 84.28% on a car engine sound dataset.
• The deformable convolution residual network demonstrated superior performance compared to existing state-of-the-art methods.
• Significant improvements were observed in recognizing car engine sounds across diverse operating conditions.

Conclusions:

• The deformable feature map residual network offers an effective approach for car engine sound recognition.
• The method's ability to concentrate on relevant features enhances classification accuracy.
• This technique represents a notable advancement in the field of acoustic event detection for automotive applications.