Research On Recognition Method Of Aluminum Profile Surface Defects Based On GAN

In recent years,the consumption and application market of aluminum profiles in China have changed dramatically.Because of its low density,strong corrosion resistance and high plasticity,aluminum has become the second largest metal after steel.However,there are many defects in the produced aluminum profiles,especially the surface defects will seriously affect the quality,safety,usability and aesthetics of the products,so it is very important to identify the problematic aluminum profiles in actual production.In addition,the production environment and process may cause the surface of aluminum profiles to be covered by oil,mud and dust,which makes it more difficult to identify defects on the surface of materials.Necessary measures must be taken to repair these defects.Traditional manual recognition is difficult to cope with the complexity of production environment.Machine vision recognition based on image processing is only applicable to surface defects with fixed shape.Based on the principle of Generative Adversarial Networks(GAN),this thesis applies the data feature extraction and sample generation ability of the deep learning model to overcome the limitations of the traditional recognition technology,and removes the random obstructions and identifies the surface defects of the aluminum profiles.Firstly,this thesis introduces the research background and current situation of surface defect recognition technology for metal materials mainly based on aluminium profiles.Secondly,the basic principle of GAN model is introduced in detail.The advantages and limitations of choosing Deep Convolution Generative Adversarial Networks(DCGAN)with Convolutional Neural Network(CNN)as the model architecture are pointed out.Then a dual-task GAN model,DTGAN,is proposed,which can successively remove random occlusions and identify surface defects of aluminum profiles.The generator of the original model is responsible for autonomous learning to generate pseudo-samples from input noise data for subsequent restoration images.The discriminator is used to discriminate authenticity to assist in completing restoration images,and the repaired samples are used to expand the diversity of the original sample set.In order to obtain high-quality repair effect,a similar perceptual loss function is designed,which helps the repaired area to have more vivid practical significance.Based on this model,an classifier is added,which is combined with the generator to form a new network for the surface classification task of the aluminum profile.In this thesis,the global search and fast convergence of differential evolution algorithm are used to optimize the parameters of DTGAN network architecture.The weight and bias are regarded as individuals in the evolution,and the global optimal solution is obtained through selection,crossover,differential variation and other operations.Then,the optimal solution is set as the initial value of DTGAN,so as to improve the accuracy and convergence of identification of surface defects of aluminum profiles.Finally,the dual task characteristics of the DTGAN model and the guiding effect of similar perception loss on image restoration are verified,and a reasonable proportion of 0.01 is found for the loss.In addition,it shows that the DTGAN model optimized by differential evolution algorithm can improve recognition accuracy by 2.2%on the basis of rapid convergence.