dc.description.abstract | The textile company applied AOI to detect a defect on fabric automatically. In their case, AOI failed to overcome this problem because there are differences defect categories. AOI has defect with categories among others; fold, black thread, broken thread, hooked thread, needle mark, mosquito, stain, thread-off, white spot and uneven thickness. Whereas, Textile company has different defects categories, its categories among others; Thinning, missing weft, stop mark, weft mark, loose weft, warp break, mechanical section and fold back. These differences caused more than 90% of captured images are not defect (overkill). A Professional Inspector in Textile Company reclassified those overkill images into real defect and non-defect images. Our goal in this work is, to reduce a Professional Inspector working loads by proposed two approaches. The first approach is, building a tiny and light CNN architecture as classifier model. While second approach is, combining Autoencoder to reconstruct the dataset as an input to CNN model that built in first approach. The result shows that the models are able to reduce a Professional Inspector working loads up to 90% with maximum FNR 5% and FPR less than 5%. | en_US |