Research On The Detection Method Of Aluminum Surface Defects Based On Convolution Neural Network

In the production process of industrial aluminum products,pitting corrosion,crack,intergranular corrosion and galvanic corrosion defects of different degrees often occur due to reasons such as chlorine-containing water solution,severe temperature change of quenching,grain boundary impurities and contact metal potential difference.From the performance of various materials,surface defects have a significant impact on the performance of aluminum materials,not only causing a sharp decline in material mechanical properties,reducing the service life of finished products,but also causing structural damage and endangering safety.Therefore,it is particularly important to achieve efficient and accurate detection of aluminum surface defects.However,the task of detecting the surface defects of industrial aluminum products is challenging,which makes it difficult to achieve high-quality detection with existing detection technologies.In recent years,the application of convolutional neural network has become more and more extensive,which provides a new idea for defect detection.Convolution neural network has unique advantages for the detection of complex and changeable material surface defects due to its flexible and stable characteristics.Therefore,this paper uses convolution neural network to carry out the research on the detection method of aluminum surface defects.Considering the large proportion of small targets in aluminum surface defect samples,sparse samples,and complex and variable characteristics,this paper divides the detection task into three stages,with the main research contents as follows:(1)An adaptive detection method of aluminum surface defects based on transfer learning is studied.Firstly,based on the construction of an aluminum surface defect detection dataset,the dataset is filtered and classified through defect data representation,and the sparse defect categories in the sample are automatically amplified to varying degrees.Secondly,in view of the overfitting problem caused by the sparse sample of aluminum surface defects and the large proportion of small targets,based on the idea of transfer learning,the deep residual learning strategy is used to obtain the defect feature map from noisy images.Through the fusion of attention mechanism and adaptive deformable convolution,the corresponding weight matrix is added to the defect feature map to achieve fine feature extraction.Then,a feature pyramid structure is constructed to achieve multi-scale feature information fusion.Finally,the obtained context feature information is used for category prediction and border regression.The experimental results show that the research of adaptive aluminum surface defect detection method based on migration learning can significantly improve the accuracy of defect target classification and location without increasing the parameters of feature extraction network.(2)An aluminum surface defect segmentation model based on convolution and edge is constructed.In view of the difficulty of defect segmentation caused by noisy background information,the idea of first locating and then segmentation is adopted.The defect feature is extracted by convolution,and the feature is classified and regressed to obtain the defect target area.Then the final segmentation result is obtained by frequency domain filtering,threshold segmentation and edge processing.The model sends the information aggregated by the feature extraction module to the regression classifier for defect location and classification,and the output end uses the edge algorithm to achieve accurate defect segmentation.The experimental results show that the aluminum surface defect segmentation method based on convolution and edge can effectively complete the aluminum surface defect online segmentation task.(3)Research on model optimization and defect Numerical calculation methods.In order to solve the problem of model parameter redundancy and low detection efficiency,a quantitative pruning and Tensor RT operator fusion strategy are used to optimize the model parameters.Under the premise of ensuring the accuracy,first quantize the model to INT8 to improve the throughput and conduct channel pruning training;Secondly,it optimizes the allocation of GPU memory and bandwidth by combining the nodes in the kernel and the kernel automatic adjustment strategy;Finally,the dynamic tensor video memory and data flow parallel strategy are used to minimize the memory occupation and effectively reuse the tensor memory.In addition,for the evaluation of aluminum defect data,contour fitting was first used to express the segmentation results,and the area and aspect ratio of the defect target were calculated using subpixel methods;Secondly,a series of material performance experiments were conducted to determine the tensile strength,elongation,surface wetting tension,and antioxidant activity of defective aluminum materials,which to some extent reflects the relationship between defects and material properties.