On-Line Detection Of Welding Surface Defects Of Chair Parts Based On Improved SSD Algorithm

Welding is a traditional means of connecting parts,which is indispensable in Anji area’s chair manufacturing.The detection of welding defects is the most important work in chair manufacturing.In order to ensure the firmness and reliability of the chair and good appearance,it is necessary to carry out quality inspection on the welded parts.In the traditional detection,manual inspection is mostly used.It has the disadvantages of high labor cost,slow speed and unreliable detection results,which brings great difficulties to the quality inspection.With the rapid development of visual detection technology,the detection based on image processing and the target detection based on deep learning are available now.This paper mainly proposes a defect detection method based on improved SSD algorithm,which can meet the detection accuracy and real-time detection speed of welding defects,so as to realize a more intelligent automatic production line processing method,which has a certain research significance.The main research work and results are as follows:(1)By exploring the background and significance of welding defect detection,and by studying the three development stages of defect detection,it is found that deep learning is of great significance in defect detection.Through the research of several typical target detection algorithms,the research direction and technical points of improvement of this paper are determined.(2)The improvement of SSD algorithm in backbone feature network and the means of feature fusion are proposed by using the backbone feature extraction network of Dense Net.The parameters of the model are reduced and the reuse of features is improved,so that the model has faster detection ability and stronger feature extraction means.The output of the backbone feature extraction network is divided into four different scales.The idea of BIFPN weighted feature fusion is introduced for multiscale feature fusion to improve the detection ability of welding defects with complex background or small shape.(3)It is proposed to make the welding defect data set in the real processing scene,and expand the data through two different data enhancement methods with external and internal characteristics,so as to solve the problem of over fitting the model caused by the small number,difficulty in preparing and uneven distribution of defect detection data sets.(4)The improvement of loss function,optimization algorithm and learning rate setting strategy based on accurate positioning are proposed.By replacing the smooth L1 loss function in the original SSD algorithm as CIOU loss,the ability to accurately locate the welding defect features with dense segments is improved;Adam optimization algorithm and Warmup + Cosine learning rate strategy are used to improve the convergence effect during training and the ability of defect identification.(5)The embedded device with efficient edge computing capability is introduced,and the improved algorithm is quantified and accelerated by using Tensor RT technology,so that the computing speed of the algorithm on the edge device reaches 32 fps,which can meet the online detection speed without losing the detection accuracy.According to the needs of defect detection,the development of welding defect detection system from camera SDK to embedded device identification and server training is completed.The work leads to the innovative development of efficient welding defect detection of chair parts.