Lightweight Aircraft Surface Damage Detection Algorithm Integrating With Attention Mechanism

The aircraft will be affected by external factors such as harsh environment conditions during the flight and will cause various damages.These damages pose a great threat to the safe flight of the aircraft and require handling them timely and effectively during maintenance work on the ground.The problem of aircraft surface damage detection can be classified as a object detection problem in the field of computer vision.Object detection algorithms are mainly divided into two categories:traditional image processing and pattern recognition methods,and object detection methods based on deep learning.Compared with traditional algorithms,object detection algorithms based on deep learning require large-scale samples and achieve higher detection accuracy.However,the object detection algorithm based on deep learning usually adopts a complex deep neural network structure with large-scale parameters,and the computational complexity is relatively high.It is of great significance to reduce the amount of model parameters and increase the calculation speed in some application scenarios which require fast or real-time monitoring.This thesis constructed a lightweight aircraft surface damage detection algorithm integrating with attention mechanism,aiming at accelerating the speed of detection while maintaining the detection accuracy or reducing less.The main research contents of this thesis are as follows:This thesis applies lightweight models and lightweight methods to design and implement lightweight deep neural network structures.In order to build a lightweight model,this thesis uses a lightweight basic network structure to reduce the amount of model parameters,and then adds the shallow feature map channels and design a module for feature fusion of the shallow information during the down-sampling layer and deep information during the up-sampling layer to maintain the detection performance of lightweight model.In terms of lightweight methods,this thesis uses knowledge distillation method by migrating the feature expression capability of heavy network to lightweight networks.This thesis applies the offline distillation method based on the output feature response to the CenterNet detection part,and designs the corresponding loss functions.Experimental results show that the lightweight model and knowledge distillation method effectively reduce the amount of parameters and calculations of the model,and increase the detection speed of the model while maintaining a certain detection accuracy.Compared with the heavy neural network,the detection accuracy of the lightweight model has decreased.In response to this problem,this thesis constructs a lightweight attention mechanism module named CAM,which improves the detection accuracy of the model with a bit improvements of parameters and computational complexity.Experimental results show that the lightweight deep neural network model integrating with the CAM structure achieve a detection speed of 65.27FPS on the aircraft surface damage dataset and the average accuracy is 56.59%when the intersection ratio threshold is 0.3.The lightweight aircraft surface damage detection algorithm designed in this thesis contains a small amount of parameters and low computational complexity.It can achieve rapid damage detection while maintaining detection performance or reducing less.