Surface Defect Inspection For Steel Strips Based On Multiscale Deep Feature Fusion

Steel strips acts as a vital and fundamental material for the steel industry.Any surface defects not treated in time will threaten the steel product quality and cause substantial economic and reputation costs to end customers.The on-line steel strips surface defect detection system is greatly troubled by the defect which has the characteristics of periodicity,wide involved area,uneven concave and convex,diverse appearances,etc...Essentially,the online defect detection is the multiscale defect detection problem in high resolution image under harsh environment.Faced with the challenges of low contrast,large intra-class distance,small inter-class distance and large scale variation,a novel method of smoothing complete feature pyramid networks(SCFPN)is proposed in this paper.In detail,multiscale features are fused by feature pyramid,regression loss employs complete intersection over union(CIo U)loss,which has suppressed the vanishing gradient from bounding boxes regression in training process,obtaining faster fitting speed and higher prediction accuracy.Label smoothing is also introduced to enhance the model generalization ability.Beyond that,in order to meet the demand of real-time detection of steel strips surface defect with limited computing resources,a one-stage object detection network YOLOF with split-attention blocks(SA-YOLOF)is in this paper.The backbone of SA-YOLOF uses Res Ne St which is stacked by split-attention blocks.Models with Res Ne St concentrate more on contributing feature channels which makes output features discriminative.From the perspective of optimization,the encoder adopts single-in-single-out form,encoder broaden the scale range of feature by adding the feature with different receptive fields.For purpose of increasing the correlation between classification branch and regression branch,the decoder introduces Io U-aware to promote the localization accuracy of SAYOLOF.In view of lack of public surface image dataset of steel strips,we open a raw defect dataset of steel strips surface called CSU＿STEEL for the first time,which contains 4 kinds of defects including roll mark,deformation,oxide scale and scratch.A number of experiments based on CSU＿STEEL dataset verify the improvement and effectiveness of two proposed methods.The proposed methods yield competitive results,with 79.5% m AP(with 11 fps)and 78.3%m AP(with 28 fps)on CSU＿STEEL dataset,respectively.Drawing conclusions from the above work,the proposed SCFPN has multi-scale and fine-grained defect detection performance and stability,and is suitable for the application of surface defect detection of steel strips at medium and low rolling speed.The proposed SA-YOLOF not only has higher calculation efficiency and lower memory consumption,but also has better multi-scale defect expression ability and defect identification ability,and is more recommended for high-speed rolling production line.