주조품 분류 정확도 향상을 위한 적대적 생성 네트워크(GAN) 이미지 데이터 확장

Abstract

Casting is a method of making metal through molds by changing solid metal into liquid state, which has the advantageof being able to yield a large number of products with complex shapes. Owing to the advantage, it is widely used tomanufacture jewelry, artwork, surgical implants, and impellers in automobiles and ships. However, low quality productscan be produced due to pinholes, sand blows, shrinkage cavities, and cracks that are well-known issues in casting. Especially using a defective impeller, a rotating element of a centrifugal pump that accelerates fluid outside from thecenter and transfers the power of fluid kinetic energy, causes a significant damage to its pump and/or workers nearby dueto its high pressure. Therefore, foundries endeavor to catch any defectives before sending them out to purchasers. However, foundries are usually small or medium-sized enterprises. It is difficult for them to hire additional experiencedworkers to catch more defectives or install photographing and imaging-storing devices to keep track of a large amountof product images for analyses. The foundries usually have a few inspectors to catch defective products and, due to ashortage of manpower and human inaccuracy, defective products are often classified as non-defective products. Thisstudy shows that a combination of classic augmentation and self-attention generative adversarial network improves theaccuracy of classifying non-defective and defective impellers by augmenting a limited amount of image data that can beeven manually photographed. Combining classic augmentation and self-attention generative adversarial networkoutperforms the sole use of classic augmentation in generating quality images for convolutional neural network.