Research On Model Lightweight For Aluminum Defect Detection Based On Deep Learning

With the continuous development of information technology and the continuous advancement of Industry 4.0,our country’s intelligent manufacturing has new challenges and opportunities.Aluminum is an important non-ferrous metal in production and life,how to ensure the production efficiency and quality of aluminum,which means how to improve the detection of aluminum defects.At the same time of accuracy,it is an important research topic to speed up the detection speed as much as possible and reduce the power consumption cost and machine cost.This thesis designs a lightweight model based on neural network,which can detect10 different defects of aluminum materials,and improves the detection speed of the original PP-YOLO-E model by more than 5 times and reduces the hardware cost to the original 1/4,the hardware power consumption is reduced to 1/67 of the original model.And the specific research content is mainly divided into the following parts:1)This thesis proposes a PSE data enhancement method,which takes positive samples as defects,and uses Mix Up,Mosaic and other data enhancement methods for fusion and data enhancement,which can make full use of positive samples and solve the problem of insufficient samples and uneven distribution of data sets.2)This thesis optimizes based on the single-stage target detection lightweight network PP-YOLO-E in computer vision,and cuts the model through sensitivity analysis and network node importance based on L1 norm,thereby reducing the computational complexity of the model.The amount of parameters reduces the amount of calculation and parameters by about 30% of the original model,and increases the inference speed of the model to 1.5 times the original.Also this thesis uses the method based on feature map to perform knowledge distillation on the cropped PP-YOLO-E model,so that the accuracy loss caused by the cropping of the model can be compensated.3)This thesis proposes offline quantization based on quantization factors,and selects different quantization factors for different weights.For the quantization of activation functions,the loss of different quantization factors is calculated through a small amount of label data,and the best quantization factor is obtained.Converting the model from FP32 to FP16 increases the detection speed of the model by about 2.5 times.4)In the process of model deployment,this paper uses the inference framework Tensor RT to further increase the detection speed by 4.8 times,and reduces the hardware cost to 1/4 of the original,and the hardware power consumption to 1/67 of the original.