A Deep Learning-based Method For The Detection Of Painting Defects On Small Samples Of Ships

Ship painting quality inspection is an important part of ensuring the safe navigation of ships,but in the modern ship industry mainly relies on manual defect detection,which is not only inefficient,but also hazardous to the health of the staff by painting volatile substances.At present,the target detection algorithm has made certain achievements,but the application of deep learning model in the field of ship painting defect detection is not mature due to the sparse amount of ship painting defect data,complex and diverse defects,and high similarity of defects,etc.In order to improve the detection efficiency of ship painting defects,a small sample ship painting defect detection system based on deep learning is built to achieve the refinement and intelligence of ship painting defect detection.refinement and intelligence of ship painting defect detection.The research content of this thesis is as follows:(1)For the problem of small samples of ship painting defect images,this thesis forms the IDCGAN(Improved Deep Convolution Generative Adversarial Networks)model based on the DCGAN(Deep Convolution Generative Adversarial Networks)model of deep convolutional generative adversarial network with optimization and adjustment.Firstly,the Generator-Residual Module Unit(G-RMU)and the Leap-layer Channel Excitation(LCE)are built in the generator of the DCGAN model to enhance the gradient flow of the deep convolutional layer in the generator and improve the quality of the generated images.Secondly,the discriminator incorporates the idea of self-supervised learning,incorporates Spectral Normalization SN(SN)and builds the Discriminator-Residual Module Unit(DRMU),which enables the discriminator to map more rich and detailed features with limited sample data.This enables the discriminator to map more detailed features with limited sample data,and makes the training process more stable.The experimental results show that the IDCGAN model has more significant generative performance than DCGAN.(2)To address the problem of low efficiency of ship painting defect detection,an improved attention module-based ship painting defect detection method is proposed.Firstly,in order to reduce the number of parameters of the detection model,the depth-separable convolution is replaced by the conventional convolution of the path aggregation network in YOLOv4 to form IYOLOv4;secondly,the CBAM attention module is improved to form ICBAM,and multi-frequency channels are constructed in the CAM channel attention module to improve the global average pooling,and the one-dimensional convolution is used instead of the fully connected layer to aggregate the information between adjacent channels to reduce the Finally,ICBAM is incorporated into the IYOLOv4 detection model to form ICBAMIYOLOv4,and the experimental results show that the ICBAM-IYOLOv4 model has stronger detection performance.(3)Design and implementation of a ship painting defect detection system.Firstly,the design criteria and functional requirements of the system are analysed;secondly,the overall framework of the ship painting defect detection system is built;then the development environment and database tables of the system are analysed and designed;finally,the development tools and programming language are used to design and develop the system and implement the functions of each module.