應用影像生成模型於少樣本木材表面瑕疵檢測

Abstract

隨著近年工業4.0的風潮漸起，如何將智慧製造應用於產線中成為產業界十分重視的課題。現今在木板、木材貼皮等木製產品加工產線中，產品瑕疵檢測大部分仍由人工進行，產品瑕疵檢測作為產線中重要的一環，如能導入自動化光學檢測將是邁向產線智慧化很好的一個著力點。然而，對於木材表面這類外觀組成複雜且無規律的檢測目標，較難以傳統電腦視覺演算法找出其中瑕疵。伴隨人工智慧深度學習技術的飛速發展，現今已有多種物件偵測模型被提出，此類模型具有高適應性與可在雜亂背景中找出目標的能力，以其作為全新的木材表面瑕疵檢測手段極具潛力。然而，物件偵測模型需使用大量影像訓練才有較好的辨識效果，在實務經驗中，常常遇到產品樣本數過少導致檢測效果不理想的情況，如能克服此問題，將大幅提升將物件偵測模型用於產線的可能性。 本研究將實際應用物件偵測模型檢測木材瑕疵，評估其是否適合作為木材表面瑕疵檢測的新方法，並探討使用不同特徵擷取器對於模型檢測效果的差異。此外本研究將使用影像生成模型生成木材瑕疵影像，測試在原始樣本數較少的情況下加入生成影像一起訓練是否能提升物件偵測模型的瑕疵檢測效果。最後根據結果提出一套針對少樣本情況下的木材瑕疵檢測流程。 經過本研究的實驗測試，物件偵測模型Faster R-CNN可用於木材瑕疵檢測，且在樣本數量較為充足的情況下，不論使用VGG16、Resnet50、InceptionV2作為特徵擷取器，其都有近六成的瑕疵檢出率。在少樣本的情況下，加入影像生成模型Pix2Pix或SPADE生成的瑕疵影像一起訓練後將提升模型瑕疵檢測能力，其中，使用InceptionV2作為特徵擷取器加上SPADE生成的瑕疵影像進行訓練時，其能將瑕疵檢出率從原本的41.58%提升至89.34%，改善效果十分顯著，以此組合作為少樣本情況下木材表面瑕疵的檢測手段將非常有潛力。

With the trend of industry 4.0 in recent years, how to apply Smart manufacturing into product lines has become the subject greatly focused in the industry. Nowadays, in the product line of wooden product processing, such as wood panel and wood veneer, most defect detection is still proceeded artificially. Product defect detection is one of the key part in the product lines. Thus, if we can introduce Automated Optical Inspection into product lines, this will be a great point to head for intellectualizing product lines. However, for the detected target like woods whose exteriors are complicated and nonregulated, it’s hard to find out the defects via traditional computer vision algorithm. With the rapid development of artificial intelligence and the technology of deep learning, many object detection model have been proposed. The kind of models has high adaptability and capabilities of finding targets from the messy background. To applying them as a measure of detecting defects possesses outstanding potential. Yet, object detection models needs a great amount of image training to have better detecting results. Within practical experiences, it’s quite common to face that not enough samples of products results in unideal detecting results. If we could overcome the problems, the possibility of applying the object detection models to the product lines will be elevated significantly. The research will practically apply the object detection model, evaluate whether it suits for the defects detection of wooden exteriors or not, and study the differences of detection effects on the different Feature Extractor. Besides, the research will use the image generating model to generate images of wooden defects, testing if training with image generating together with fewer original samples can elevate the defect detection effects on the object detection models. The research will eventually propose a best detecting set for wooden exterior defects according to the final result. Through the tests of the research, under the situation with enough samples, the training result of Faster R-CNN has almost 60 percent recall no matter we chose VGG16, Resnet50 or InceptionV2 as feature extractors. Under the situation with fewer samples, adding the generated defect images of Pix2Pix or GuaGan to the training will elevate the recall. For instance, training InceptionV2 as feature extractor with the defect images generated by GuaGan together can elevate the recall from 41.58% to 89.34%, which presented a significant effect on improving the results. Thus, using the set as a measure for the detection for wooden exterior defects showed a high potentiality.