Research On Few-Shot Learning Method For Surface Defect Detection Of Steel Plate

As an important branch of computer vision,defect detection has a wide range of applications in industrial production.Although the traditional image processing technology has become the mainstream detection method,it still has certain limitations,such as the inability to quickly detect defects on the steel plate surface and the detection accuracy is low.The method based on the convolutional neural network can use data to train a model that can detect the surface of the steel plate.The model takes less time to detect the surface of the steel plate and has a higher detection accuracy.Low detection accuracy and poor model fitting.To this end,three models for steel plate surface defects are proposed,YOLO-C algorithm,ID-RCNN algorithm and Cascade Detection Network algorithm.They can not only solve the problems encountered above,but also verify their effectiveness.The content of the thesis mainly includes:Aiming at the small sample training problem encountered by the deep learning method in the detection of steel plate surface defects,a method of integrating the attention mechanism and the YOLO model is proposed.On the basis of the YOLO architecture,the attention mechanism module is added to enhance the model’s ability to train data,and the channel and spatial dimensions are enhanced with attention,making full use of data features and improving the generalization of the model.In addition,the generated feature map is calculated and analyzed into a feature map with perceptual direction and sensitive position,which can not only use the input of the feature map,but also enhance the feature representation of the object.Finally,a model with high defect detection ability is obtained under small sample data training.Aiming at the problem of low detection accuracy of the deep learning method on the surface of the steel plate,the ID-RCNN algorithm is proposed.According to the characteristics of the high accuracy of the two-stage network,an attention mechanism is designed.By redesigning the MLP layer,the designed attention The force mechanism is integrated into the Basicblock layer and the Bottleneck layer to improve the model detection accuracy and generalization ability.In order to prevent the possible overfitting problem of small sample data sets in model training,the Cascade Detection Network algorithm is proposed,which uses the designed attention mechanism to enhance the front-end data input layer,and makes full use of the characteristics of the Cascade RCNN cascade detector,take an incremental threshold for each stage,so that each stage has enough data processing capacity to prevent the occurrence of overfitting problems.On this basis,DAGM and NEU-CLS steel plate surface defect datasets are used to verify and evaluate the above model.Among them,the accuracy of YOLO-C algorithm is4% higher than other models,the accuracy of ID-RCNN algorithm is 1.1% higher than other models,and the accuracy of Cascade Detection Network algorithm is 2% higher than other models.Experimental results show that the designed algorithm can effectively detect steel plate surface defects.