Improved YOLO-V3 metal processing surface defect detection method

Abstract

The invention discloses an improved YOLO-V3 metal processing surface defect detection method. The method comprises the steps of acquiring a metal surface processing image; extracting feature maps of different scales from the feature extraction backbone network, wherein the feature extraction backbone network has no pooling layer, the downsampling factor is smaller than the downsampling factor of Darknet-53, and the convolution layer number is smaller than the convolution layer number of Darknet-53; inputting the feature map into a multi-scale fusion module for local feature fusion to obtain afused feature map; and inputting the fusion feature map into a multi-classifier module for metal processing surface defect positioning and classification. According to the invention, a pooling layer is omitted and a downsampling factor is reduced, and the problem of low-level feature loss generally caused in the down-sampling process of pooling operation is avoided. According to the method, the structural characteristic of high resolution of the deep features is guaranteed; the high resolution characteristic of the low-level features and the high semantic information characteristic of the deepfeatures are combined through local feature fusion; and finally accurate detection of small target defects on the metal machining surface is achieved.

Description

technical field [0001] The invention relates to the field of defect detection, in particular to an improved YOLO-V3 metal processing surface defect detection method. Background technique [0002] The detection of surface defects in metal processing is an important process in the manufacture of metal parts. Common defects on metal surfaces include scratches, crushes, and bruises. In precision mechanical equipment, the above defects may lead to potential problems such as deterioration of working conditions, affect the transmission accuracy of equipment, generate noise, and even cause equipment damage, resulting in huge losses. Affected by tooling iron residues, improper workpiece clamping, physical collisions, etc., surface defects in metal processing often occur, and batch problems are prone to occur. At present, manufacturing enterprises mainly use manual sampling inspection for the detection of metal processing surface defects, which has problems such as relying on manual...

AI Technical Summary

Problems solved by technology

[0004] The above method can detect conventional defects on the appearance of industrial devices through the traditional neural network, but when detecting processing defects on the metal surface, since the defects such as crushing and bruising on the metal surface are usually small in size, the defects in the input image The pixel area is also very small. If the traditional neural network is used, because its feature extraction backbone network usually has a large receptive field, although it is convenient for image classification, it will compromise the spatial resolution, resulting in the inability to accurately identify small objects, that is, the convolutional layer changes. Deep depth leads to a larger receptive field. For example, after the original feature extraction backbone network Darknet-53 of YOLO-V3 downsamples the image by 32 times, the final output feature is equivalent to a point, which will lead to the detection of this type of defective target. The effect is poor, which is not conducive to the detection of small target defects, and will bring problems of poor detection position accuracy and low recall rate

Images