Strip Defect Classification And Location System Based On Feature Fusion And Target Detection

As one of the most important production materials for industrial production,steel is related to the development of all walks of life.As one of the most used materials in steel materials,strip steel is critical to product quality in many industries.Due to storage,transportation,manufacturing process and other factors,the surface of the strip product may have some defects.Strip defect detection can identify defects such as white spots,black spots,corrosion,roll marks,holes and other defects in the production process in time,and this will improve the quality of strip steel.Therefore,it has been a hot issue in the industry.The strip defect classification and location system based on feature fusion and target detection adopts deep learning technology,introduces the cross-layer feature fusion model into the traditional feature pyramid structure,and uses the target detection algorithm based on feature fusion technology to classify and locate strip defect.In the normal feature pyramid structure,because the features of each layer are separated from each other,the high-level features and the low-level features cannot be combined well.Therefore,the system uses a cross-layer feature fusion model,which is used to perform feature maps of different scales in the feature pyramid.The fusion operation obtains a feature map of a certain size that contains global features,and then merges it into the corresponding layer to obtain a better feature expression.Since the result of the fusion operation has many different scales,in order to better describe the models of different scales,the system proposes four different cross-layer feature fusion models of M2,M3,M4,and M5.The system implements the classification and defect location functions for strip steel defects.It can use boundary boxes and masks to locate defect features from the image,and gives the defect category and confidence.The experimental results show that the cross-layer feature fusion model has better accuracy for the classification and location of strip defects.The overall detection rate of M4 has reached 93.9%,and the detection rate of M3 and M5 fusion feature models has also increased by 1.9%,0.7%.The M3 has the lowest overall false detection rate among all detected targets,accounting for 3.34%,followed by M5 at 4.77% and M4 at 5.78%.The performance of the M4 feature fusion model in terms of network overhead is also optimal.In the experimental environment,the detection overhead is reduced by 10.92 ms,which is 2.6%.The m AP value of the M3 ? M5 feature fusion model is higher than the original model,and M4 is the highest,which is 0.462395.