Research On Image Denoising Method For Strip Steel Surface Defect Detection

With the rapid development of intelligent technology,machine vision technology has been significantly improved in the field of accurate and efficient industrial inspection,which is widely used in many aspects of mechanical engineering.As an important raw material for mechanical manufacturing,the surface quality of strip steel is directly related to the performance and life of the final product.However,due to various reasons such as production process,environmental factors and aging of equipment,various defects may appear on the surface of strip steel,such as cracks,corrosion,inclusions and so on.In order to detect and deal with these defects in time and improve product quality,machine vision technology is widely used in strip steel surface defect detection.Image processing is an important part of machine vision detection,and denoising is one of the key steps of image pre-processing.For the strip steel surface defect image,the purpose of denoising is mainly to eliminate the noise interference in the image,highlight the defect features,and provide a clear image for subsequent edge detection,identification and analysis.In recent years,wavelet transform,variational and partial differential equations,and deep learning methods have become the mainstream technologies for machine vision and image processing.In this paper,based on these three,we choose to study the problem of denoising images of surface defects on strip steel based on wavelet and partial differential equations.The main innovative work is as follows:Firstly,in order to overcome the wavelet denoising method is prone to produce ringing effect at the edge or jump of the image,and retain the accurate image detail information of surface defects of strip steel,a state-aware image denoising algorithm is proposed.Firstly,in order to avoid image distortion,symmetric wavelets are used to decompose the image and obtain the wavelet coefficients of each frequency band;secondly,through the theoretical analysis of wavelet thresholding,the concept of wavelet coefficients’ state-aware quantity is introduced,and the high-frequency wavelet coefficients are normalized to obtain the corresponding wavelet coefficients’ state-aware quantity in order to improve the comparability of wavelet coefficients as well as the reliability of the image feature extraction;and then the state-aware quantity is subjected to fourth-order anisotropy and the state-aware quantity is normalized.Then,the fourth-order anisotropic diffusion of the state-aware quantity is used to obtain the state-aware power of the corresponding wavelet coefficients;finally,the state-aware power is applied to the corresponding wavelet coefficients to obtain high-frequency wavelet coefficients that accurately reflect the image information,and the denoised image of the surface defects of the strip steel is obtained by using the wavelet reconstruction algorithm.The experimental results show that the proposed algorithm can obtain accurate image details while suppressing noise.Secondly,in order to improve the edge protection ability of the higher-order variational regularization model on the surface defect image of strip steel,an adaptive higher-order variational denoising method based on edge protection is proposed.Taking advantage of the wavelet coefficients in characterizing the edge and detail information of the image,the wavelet coefficient modes are used instead of the gradient modes to construct the detection function with the function of edge protection.Taking the constructed edge detection function as a regular term,a new higher-order regularization model is established,and the existence of the uniqueness of the solution of the model is proved theoretically.In this,the state-aware based method proposed in this paper is utilized to ensure the accuracy of the obtained wavelet coefficients.The experimental results show that the new model not only improves the denoising ability of the higher-order diffusion model,but also obtains clear and complete edges of the image of surface defects on the strip,demonstrating its advantages in denoising and edge protection.Thirdly,in order to suppress the noise generated by the image of surface defects on strip steel in the process of magnification,an image magnification algorithm with coupled thresholding in wavelet packet transform domain is proposed.Through the analysis of wavelet thresholding theory,the partial differential equation model of wavelet thresholding is established,and the flux function is introduced in order to characterize the rate and direction of information propagation in the image.On this basis,the coupled threshold relation equation guided by the diffusion function is obtained,in which the diffusion function is selected as bidirectional diffusion to ensure that the smooth denoising and edge preservation are achieved at the same time.In order to ensure the similarity between the original image and the enlarged image,the original image is soft threshold as low-frequency coefficients,which are reconstructed with the high-frequency wavelet packet coefficients of the coupledthresholding process to obtain the final enlarged image.The experimental results show that the proposed image enlargement algorithm effectively suppresses the noise of the strip steel surface defect image itself as well as the noise in the enlargement process,and the edges of the enlarged image are clear and continuous,and better results are obtained in terms of visual effect and objective evaluation indexes.Fourth,in order to enhance the stability and interpretability of the deep learning denoising model,a residual learning image denoising method based on diffusion function is proposed.The process of converting a wavelet threshold partial differential equation model into a single residual learning block in a residual learning network(ResNet)is investigated,the framework between partial differential equations and residual learning is established,and the correspondence equations between diffusion function,threshold function,and activation function are obtained.Based on the deep residual convolutional neural network model(DnCNN),the diffusion function is used as the activation function to construct the deep learning denoising network model in order to enhance the expressive ability of the network and make the network obtain more image features.The experiment selects three common diffusion functions as the activation function of the network,which is trained and tested by Severstal steel defect dataset and NEU strip surface defect dataset from Northeastern University.The experimental results show that the diffusion function can achieve the characteristics of the activation function,and can effectively remove the image noise of strip steel surface defects while also obtaining clear and continuous defect image edges.