Research On Detection Algorithms Of Aluminum Surface Defects Based On Convolution Neural Network

As a common industrial material,aluminum has large market demand and high-quality requirements.However,due to various reasons in the production process,defects on the surface of aluminum materials are inevitable,which will reduce the external aesthetics of the products,affect the performance of the products and shorten the service life of the products.Therefore,the detection of aluminum surface defects is very important.There are so many surface defect detection methods.In the past,manual visual inspection was often used.That was flexible,but it was highly subjective and slow.In the past,manual visual inspection was often used.That was flexible,but it was highly subjective and slow.With the rapid development of computer vision technology,the detection method based on deep learning has been widely used in many fields,and the detection effect is outstanding.Therefore,based on the classical deep learning detection algorithm Faster RCNN,this paper studies the aluminum surface defects as the detection object.The main research contents of this paper are as follows:First of all,seven common aluminum surface defects are selected as the research object of this paper,and the defect data format is processed and converted into the PASCAL VOC format commonly used in the field of target detection.In order to solve the problem of too little defect data,a variety of data amplification methods are used to increase the number of data sets to four times of the original.The set of aluminum surface defects is analyzed from many aspects such as defect shape,defect area and defect labeling frame,and the detection difficulties such as irregular defect shape,different scale,and large proportion of small target defects are obtained,which is convenient for the subsequent targeted improvement of the Faster RCNN network.Then,according to the characteristics of aluminum surface defect data set,a detection algorithm of aluminum surface defects based on improved Faster RCNN is proposed.By analyzing various convolution neural networks,ResNet50 is selected as the feature extraction network in this paper,and deformation convolution is embedded in the last two stages of ResNet50 to adapt to the irregular shape of defects.In order to solve the problem of different scales of surface defects,K-means clustering algorithm is used to customize the size and ratio of anchors and optimize the original anchor parameters;The CIOU function is used as the position regression loss function and combined with the cross-entropy classification loss function to form a new loss function,which improves the target positioning accuracy and accelerates the network convergence.Experiments show that the improved network has superior performance,accurate positioning and correct classification when various defects are detected.Finally,aiming at the problem of missed detection and repeated detection,a detection algorithm of aluminum surface defects based on improved Cascade RCNN is proposed.Based on the feature extraction of ResNet50,PAFPN is constructed by using the idea of multi-scale feature fusion.The low-level location information and high-level semantic information are fused multiple times from bottom-up and top-down directions to reduce the loss of feature information and solve the problem of target missing detection.The Soft-NMS algorithm is used to optimize the anchor box selection strategy of NMS algorithm to solve the problem of repeated detection of multiple anchors.The detection results and experimental data show that the improved network performance is better.The detection accuracy is significantly improved,and the situation of missed detection and repeated detection no longer occurs.