A Convolutional Neural Network Button Defect Detection Algorithm Based On Parallel Features And Attention Mechanism

With the continuous progress of my country's industrial level,the influence of the button sound field on the apparel industry is becoming more and more important,and the quality of buttons also seriously affects the sales of clothing.Buttons are an important part of most garments,but in their production and processing,they may often be interfered by a series of unknown factors such as processing errors,abnormal machine operation,and damage to production molds,resulting in various kind of flaws.Common defects include missing button holes,deformation of button holes,cracks on the button surface,color difference of buttons,stains on button surface,abnormal button texture,and button shape damage.Now the main method of manual detection is used for button detection.A big problem with this type of method is that the detection efficiency is not high and the detection accuracy is difficult to guarantee,which greatly restricts the development of enterprises.Under the current process of national science and technology development,it has become possible to replace humans with machines for related inspection work.There are many button manufacturers in the market that urgently need to replace manual inspections with fast-developing deep learning and other technical means,in order to achieve the goal of enterprises to improve production efficiency.To help companies improve production efficiency,a defect detection model of resin button and metal button based on deep learning is proposed.Using algorithms to automatically identify button images to detect button defects instead of manual detection is of great significance to enterprises.In this thesis,a set of button defect detection algorithm is established based on parallel feature extraction and attention mechanism.The main research of this thesis includes three parts: the button defect detection algorithm based on parallel feature extraction network;the button defect detection algorithm of parallel feature extraction network with channel attention mechanism;the accurate classification algorithm of multi-defect buttons.The main innovations of the paper are as follows:(1)A button defect detection algorithm based on Parallel Feature Extraction Network(PFEN)is proposed.The algorithm uses a deeper and wider convolutional neural network CNN-B and a shallower and narrower convolutional neural network CNN-A to extract the deep features of button images,and finally perform feature fusion interaction to obtain the final features maps.Before training,the deep and wide network CNN-B is pre-trained on the Image Net dataset in advance.In the actual training process,the CNN-B feature extraction network has only the last three convolutional layers at the end of the network trained by the standard button image data sets,and the rest of the network layers are frozen,while the CNN-A feature extraction network is completely trained by the standard button images data set.Finally,experiments prove that the proposed parallel feature extraction network can perform accurate defect detection tasks on metal buttons and resin buttons after training on a small standard button image data set.(2)The channel-based attention mechanism is added to the CNN-A network of the parallel feature extraction network to establish a parallel feature extraction network(Parallel Feature Extraction Network with Channel Attention,PFEA)combined with the channel attention mechanism.By further correcting the activation value of the feature map extracted by the network,the network can more accurately identify effective feature information.Finally,through comparative experiments,it is proved that after adding the channel-based channel attention mechanism,all indicators of the model have been improved.(3)Button production often produces multiple types of defects,and a single button sometimes has multiple types of defects at the same time.Enterprises need to obtain button defects in time to facilitate their targeted adjustment of related production processes.Through the establishment of a new multi-defect button image data set and the establishment of a multi-defect button detection algorithm(Multi-defect Button Detection Algorithm,MBDA),it can accurately classify multi-defect buttons and obtain the exact type of button defects to meet the needs of enterprises to improve production processes.