Salient Object Detection Of Strip Steel Surface Defects Based On U-Net

The iron and steel industry is one of the pillar industries in our country.When its surface is defective,it will not only affect the appearance of products,but also affect the performance of products.The detection of strip surface defects is still a challenging task due to the complex variation of strip surface defects,high noise interference and low contrast.Existing defect detection methods have the following disadvantages while accurately segmenting defect objects.Firstly,for small significant defect objects,it is difficult for existing models to segment complete defect objects from the compact background.In addition,when the defect background is cluttered or has low contrast,the current methods tend to introduce background noise,and it is difficult to restore the complete defect area from the complex background.To solve these problems,a convolutional multi-scale feature fusion U-Net network(CMU-Net)based on encoder and decoder structure is proposed in this thesis.The main research contents of this thesis are as follows:(1)Aiming at the detection integrity problem of small defect objects,an improved encoder structure based on Conv Ne Xt network was proposed.In the classic U-Net network,the encoder part uses shallow convolution with a fixed size of 3×3 for feature extraction.In this way,defect information of different scales will be lost,and it is difficult for the network to detect the complete defect structure.In this thesis,an improved serial connection multi-scale feature extraction encoder based on Conv Ne Xt structure is proposed to replace the original convolution module.The improved encoder structure has deeper network layers and can effectively obtain deep semantic information.The improved encoder architecture uses Conv Ne Xt blocks as the base module,with serial connections in a ratio of 1:1:3:1,while each Conv Ne Xt Block uses a residual structure.Simulation results show that the improved network can obtain complete defect information,achieve maximum feature utilization,and improve the performance of network segmentation.(2)To solve the problem of introducing background noise into image feature restoration,this thesis proposed that channel shuffling residual module(SSR)and attention mechanism decoder module(SDB)constitute a new decoder.The original decoder of U-Net has poor feature reduction ability and is easy to introduce noise.In this thesis,a new encoder is constructed by alternately connecting SSR and SDB to integrate shallow spatial features and deep semantic features,and gradually restore the detected segmented area.Channel shuffling technology is used in the channel shuffling residual module to make full use of the feature information obtained by the network in the encoder part.The depth separable convolution and residual structures in the structure are easy to network training.The compression excitation network is introduced in the decoder module of the attention mechanism,and the attention weight coefficient learned in the compression excitation network is allocated to the corresponding defect area to reduce the interference of irrelevant background information.Compared with existing defect detection methods,CMU-Net can accurately segment complete defect objects and effectively filter out irrelevant background noise.A large number of experimental results show that the proposed CMU-Net is superior to the existing detection methods and has good segmentation performance.