Research On Surface Defect Recognition Technology Of Aluminum Profile Based On Deep Convolutional Neural Network

Aluminum profiles have excellent corrosion resistance and other properties,and are important raw materials for various industrial products and daily necessities.Due to various unavoidable factors in the production process,there will be some defects on the surface of the aluminum profile,which will affect the quality of the product.At present,aluminum profile enterprises mainly adopt the method of manual identification to detect surface defects,which is slow in detection speed and weak in stability.However,machine vision recognition based on traditional image processing can only detect simple defects.When faced with defects in aluminum profiles with various shapes,the algorithm has poor robustness,and a large number of threshold parameters need to be manually selected,and the degree of intelligence is relatively low.In order to improve the automation level of defect detection,the thesis constructs an aluminum profile surface defect recognition algorithm based on deep convolutional neural network.The main work is as follows:（1）The industrial inspection requirements for aluminum profile defect identification are clarified.After analyzing the surface defects of aluminum profiles,it is found that the defects have the characteristics of large size differences,various shapes,and small size of some defects,which are difficult to detect.Based on this,the overall scheme design is given,and the two-stage target detection model Faster RCNN is selected as the basic network of the thesis.（2）In view of the large size difference of aluminum profile defects,the thesis introduces the feature pyramid module to the model for multi-scale feature fusion,which improves the model’s multi-scale target detection ability.For the flaw of extreme aspect ratio,the thesis improves the size of the anchor to optimize the generation of the preselected box.To enhance the localization accuracy,the thesis also uses the CDIo U LOSS bounding box regression loss function for improvement,and eliminates the error in the quantization process through Pr Ro I Pooling.（3）In view of the problem that standard convolution cannot perform sufficient feature learning for defects of various shapes,the thesis introduces deformable convolution to the network,and the receptive field of the network can be adaptively transformed according to the feature shape,which improves the detection of irregular defects.detection capability.Aiming at the problem of low detection accuracy of small-sized defects such as paint bubbles and dirty spots,the thesis adds a feedback path to the feature pyramid structure so that the network can repeatedly learn the features,and fuse the features learned twice to enhance the features.The detailed information of the map,thereby improving the classification and localization of small-sized defects.To sum up,the thesis constructs a set of aluminum profile surface defect recognition algorithm based on deep convolutional neural network,which improves the detection accuracy of defects.Among them,the detection results of irregular-shaped defects such as scratches and jets,and small-sized defects such as paint bubbles and dirty spots have improved significantly,providing reference value and reference for other types of defect detection.After experiments,the model in the thesis has achieved a m AP of 87.2% for surface defect detection of aluminum profiles,which is 3.8% higher than the existing algorithm;the recognition accuracy rate has reached 97.2%,and the average detection time is 0.169 s,which meets the needs of industrial inspection.,has better practical value.