Research On Crack Image Expansion Method Based On Improved GAN Networ

Nowadays,with the development of China’s transportation industry,tunnels account for an increasing proportion of daily transportation facilities,and the efficient detection and prevention of tunnel cracks have attracted a focus of social attention gradually.Traditional image augmentation methods for generating samples,such as geometric rotation,cropping,and stitching,are effective in traditional tasks such as crack image binary classification detection,but cannot meet the requirements of high sample requirements such as fine grained multi classification of crack images.With the development of deep learning theory quickly,the deep learning method based on the concept of transfer learning is gradually applied to various image expansion,and the sample quality which is generated by it is high enough to meet the needs of tasks with high sample requirements.Previous studies have pointed out that traditional image augmentation is not suitable for tasks with high sample requirements.However,using different deep learning sample augmentation methods in different scenarios can result in different sample generation effects.When using the deep learning method of encoding and decoding process for image expansion in a single dataset,the generated image sample features are single,and the extracted image features are not rich enough to meet the requirements of model training robustness,but it is difficult to meet the requirements of model generalization ability.When using the deep learning method of transfer learning concept to expand images in a variety of data sets,the generated sample features are rich,and the training of complex models takes a long time.Given the issues of lack of feature extraction in a single dataset and generation efficiency in multiple datasets in the two different sample expansion scenarios mentioned above,this thesis will focus on the crack image expansion problem based on an improved GAN network.The specific content is as follows:(1)In order to ensure that sample expansion can be carried out on a single dataset and generate a certain quality of samples,experimental comparisons are conducted based on an improved VQGAN network model.First,the original VQGAN model uses the default hyperparameter in the original model to expand the samples,uses the commonly used evaluation indicators FID and IS of the confrontation generation network to evaluate the quality of the generated samples,and puts the generated samples into a simple binary network built with the Alex Net model as the backbone for classification.Then,a Feature Fusion module is introduced into the encoding and decoding structure of the original model to enrich feature extraction.Finally,an improved convolution module was used to replace ordinary convolution and Pytorch dynamic learning rate adjustment strategies for model lightweight.The generated samples were also evaluated for indicators and classification accuracy.The experiment showed that the introduced feature fusion module can enrich the extraction of deep features in the model,and the improved convolution module and dynamic learning rate adjustment strategy accelerated the training of the model.On a single crack sample dataset it can generate available samples during sample expansion efficiently.(2)In order to ensure that crack samples with more features and certain quality can be generated in a state with similar crack feature datasets,an improved Cycle GAN network model is used for experimental comparison.First,the original Cycle GAN model uses the default hyperparameter in the original model for sample expansion.After using indicators and classification accuracy evaluation,the global feature fusion module and channel attention mechanism are introduced into the codec structure of the original model to enhance the extraction of features.The improved packet convolution and depth separable convolution are used instead of ordinary convolution to form a double lightweight module,Using indicators and classification accuracy for evaluation,experimental comparisons show that the introduction of the global feature fusion module and attention mechanism enhances the model’s extraction of overall image information,while the two-layer lightweight module reduces the number of parameters in the network layer,accelerating model training.When expanding samples on multiple crack sample datasets with similar crack characteristics,available samples can be generated efficiently.