A Lightweight Convolutional Neural Network for Real-time Detection of Aircraft Engine Blade Damage

Abstract

To address the large number of parameters and the computational complexity of deep learning models in the field of borescope detection, we propose a lightweight blade damage detection model LSSD using a knowledge distillation algorithm. First, the inverse residual structure is used to lightweight the backbone network of the classic SSD model. Then, the K-means clustering algorithm is used to optimize the scale and number of anchor boxes to reduce the parameters and computational complexity of the proposed model. Second, to ensure that the lightweight model has a certain level of detection accuracy, a feature fusion module CA-FPN combined with coordinate attention and a small damage detection enhancement module W-Inception are embedded. Finally, the knowledge distillation algorithm is used to further improve the detection accuracy of the model. The number of parameters of the LSSD model is 4.99M, the MACs is 3.541G, and the detection speed reaches 32FPS. Compared with the SSD model, the LSSD model reduces the number of parameters by 79.3% and the computational complexity by 88.42%, resulting in a 2-fold increase in the detection speed.

Keywords

• aero-engine blade
• damage detection
• borescope detection
• lightweight convolution neural network
• knowledge distillation