Research On Detection Algorithm Of Fine Apparent Damage Of Metal Workpieces Based On Deep Learning

Metal workpieces are widely used in the field of electronic instruments.Metal workpieces with substandard apparent quality affect the performance of the entire product directly.Traditional product apparent damage detection is mainly based on manual visual inspection,which cannot be adapted to large-scale industrial production.This topic takes the three precision machining workpieces in the notebook as the research object.In view of the apparent damage of precision machining workpieces,which has the characteristics of irregular shape,similar color to the background,and small damage.And combined with the efficient,fast,and non-contact advantages of deep learning,we build a micro-appearance damage detection system for metal workpieces based on deep learning.This paper first compares and analyzes the performance of a variety of imaging equipment,selects hardware equipment to build an image acquisition system.According to the actual imaging effects of the workpieces under different illumination angles,working distances,and lighting methods,we formulate targeted lighting scheme for three metal workpieces.We collect a large number of apparent images of metal workpieces through the imaging system,and we manually label the samples by labelme software to build a sample set of apparent damage of the workpiece,and perform data enhancement on the original positive sample images with apparent damage to improve the robustness of the detection model.Aiming at the problem that some invalid background areas in the original image of the workpiece will cause detection interference,a set of workpiece area localization algorithm is designed.Through image preprocessing,noise interference is removed and image details are preserved.The local enhancement algorithm of the workpiece area is used to improve the local contrast of the workpiece body.The workpiece edge detection algorithm is used to obtain the edge of the workpiece body.Combined with binarization and morphological processing,scattered noise points are removed and the workpiece body is smoothed.The edge and contour extraction algorithm obtain the contour of the workpiece edge,and finally the contour is filtered to obtain an image containing only the workpiece area.Aiming at the problem that the original segmentation network has poor segmentation effect on the surface of the workpiece,such as shallow scratches and abrasions,we propose a multi-level fusion method of global and local features to improve the detail perception ability of the network.The coordinate attention mechanism is introduced into the decision network to focus on the local area of interest to improve the classification accuracy.For the problem of imbalance between damage samples and background samples,we introduce weight factors to optimize loss function of the two-stage network,which can improve the detection effect of subtle apparent damage.In this paper,we realize the quantitative analysis of micro apparent damage,and complete the software visualization interface design of the detection system for micro apparent damage.