Imaging and Convolution Neural Network for Surface Finish Quality Assurance on Assembly Line

Surface finish of a product is among one of the important parameters of manufacturing process consideration. A product no matter the configuration without a good surface finish will not wholly be acceptable by the customers unless the roughness is part of engineering consideration. On the assemble line, inspection is usually carried out to check if the product has met its surface finish requirements. Such inspection includes sampling techniques such as single and double sampling etc. These techniques are cumbersome, costly, and sometimes unable to completely uncover all the defects in a batch. For a continuous production line, therefore, it is desirable to deploy an alternative and superior method that can reduce the number of defects, easy to apply, less expensive and allow flow with manufacturing process. This was accomplished by classifying flaws on metal surfaces and detecting roughness on them using Convolution Neural Networks (CNN). 1800 images of metal surface flaws were gathered to create this dataset. With three hundred images of metal surfaces for each of the six defects: pitted surfaces, patches, inclusions, crazing, rolled in scales, and scratches-the flaws were categorized. There were roughly 1620 photos which was trained (90%) and about 180 pictures that was used in model testing (10%). With the model developed images of any other metal surface can be taken on the production line and compare with the established model.