Research On Data Augmentation Method Of Surface Roughness Image Sample

In the mechanical processing,measuring workpiece surface roughness is a key step of surface quality detection.At present,surface roughness detection model built based on convolutional neural networks(CNNs)and machine vision has addressed many issues of traditional machine vision roughness detection model,like high challenge in building models,high cost and low efficiency.Hence,CNNs roughness recognition model is drawing more and more attention.However,data insufficiency in training CNNs model brings great challenges for further study and application.To address this issue,a data augmentation method based on deep convolutional generative adversarial networks(DCGANs)and texture feature learning algorithm are proposed in this thesis.The research details are as follow:1.Collecting workpiece surface images and constructing data sets.In this thesis,milled standard blocks are set as research objects in image collecting process.Then,image collecting system is prepared for collecting image samples.After that,preprocessing of collected images is conducted.Finally,data sets are constructed.2.Constructing CNNs roughness recognition model and feasibility verification.Roughness recognition model based on residual networks(ResNet)is constructed and feasibility verification is conducted in this thesis.The experimental results indicate that ResNet roughness recognition model has excellent performance in milled surface roughness recognition task,and the maximum recognition accuracy is 82.70%.3.Conducting surface roughness recognition experiments based on DCGANs data augmentation method.DCGANs is applied in milled surface roughness recognition task for the first time in this thesis.This data generation model is trained to generate milled surface images for data augmentation of CNNs model`s training data set,then enhancing CNNs model`s performance.Data augmentation is conducted in two aspects:diversity and numbers of training samples.The experimental results show that sample diversity augmentation can enhance the performance of CNNs model and achieve performance gain of maximum 4.35% with 60% replacement of training data;the sample number augmentation can also do that and achieve performance gain of maximum 10.10% with 220% sample number amplification of training data.4.Proposing the texture feature learning algorithm and constructing texture feature learning generative adversarial networks(TFL-GANs).In this thesis,TFL-GANs is proposed based on surface texture feature and DCGANs.TFL-GANs is used to improve quality of generated images by enhancing the representation in texture feature field,then to enhance performance of CNNs model.The experimental results indicate that the proposed TFL-GANs model can further improve CNNs model`s performance compared to DCGANs.According to these theoretical analysis and experiments,the ResNet roughness recognition model can achieve great performance in milled surface roughness recognition task.Data augmentation based on DCGANs and TFL-GANs proposed in this paper can address data insufficiency and further enhance the performance of CNNs model.