Research On Real-Time Detection Method Of Steel Plate Defects Based On Reconstruction Network

Using steel in new energy vehicle enterprises,chip equipment manufacturing,and other fields is essential.The quality inspection of steel plates,as a critical part of the manufacturing process,can prevent defective products from entering the market.The detection method based on deep learning can meet the requirements of high-precision steel plate detection.Still,it needs the convolutional neural network to further improve the learning performance of the network with increasing depth.However,when the convolutional neural network’s depth increases,the network’s computing cost will also increase,and the operator reasoning at the hardware end will consume more time,resulting in a slower detection speed.The actual production deployment also increases the cost of hardware.Based on the above challenges,this paper studies the real-time detection method of steel plate surface defects.The main contributions and innovations are as follows:To analyze the influence of rotation Angle on model prediction given rotation deviation of images collected due to equipment jitter or environmental impact in the actual production environment.To make the model still ensure high-precision recognition in the case of external disturbance,it is necessary to do the corresponding data processing to ensure that the model has good robustness to the acquisition of image recognition under interference.Because of the difficulty in identifying minor target defects in steel plate surface images,a convolutional neural network with a multi-branch tree structure was proposed based on the reconstruction network.Each branch realized the decoupling of training time and reasoning time structure through structural reparameterization,and 3×3 convolution was adopted in each branch.The deep optimization of 3×3 convolution in special hardware devices or Cu DNN also optimized the problem of parameter redundancy in neural network reasoning.On this basis,a two-tree attention mechanism was proposed to suppress useless channels and enhance the learning of beneficial channels.Multi-branch information is combined with cross-channel communication to obtain weighted defect features and improve the feature information of the defect region.Subsequently,we use a feature fusion method based on Transformer to solve the problem of convolutional neural networks’ restricted local receptive field.To balance positive and negative samples in detection,the Focal loss function was introduced to reduce information loss.The results of experiments on NEU and DAGM data sets show that our method achieves the balance of accuracy and detection speed and can effectively detect steel plate surface defects in real-time.Aiming at the contradiction between the detection accuracy and detection speed of steel plate surface defects,an automatic pruning method based on parameter reconfiguration was proposed.The first stage of the form is the training of the weight generation network,and then the structure of the pruning network is re-parameterized.In the reasoning process,the branch connection path is re-parameterized into the convolutional layer so that each layer can have any width without being constrained by dimension matching.Finally,the optimization network is searched.In the experiment,the effectiveness of the structure reparameterization method was discussed,and the comparison of different pruning algorithms,the number of channels after pruning,and the number of weights and FLOPs of other models were carried out on NEU and DAGM data sets,respectively.According to the experimental results,the automatic pruning method based on parameter reconfiguration balances the detection speed and accuracy well in detecting steel plate defect surfaces.