大綱
|
摘要
瑕疵檢測對於螺絲產品之表面檢測扮演重要腳色。具有缺陷的螺絲產品大幅琳影響螺絲本身的機械性質及其安全性,因此檢測螺絲產品的品質是二貧重要的生作。本論文研究目標主要利用深度學習方法於即時檢測螺絲產品上的瑕疵特徵。螺絲產品本身因不同的取像位置產生不一致的螺紋成像結果,影像處理方法容易受環境光源影響,使得以自動化光學檢測方法之參數設定及挑選標準樣板影像技術帶來挑戰。為了克服現行自動化光學檢測之瓶頸,學者將非監督式異常檢測方法導入於自動化光學檢測系統之表面瑕疵檢測。然而異常檢測模型無法很好的重建異常影像,以致後續造成瑕疵檢測不佳。為了提升異常檢測於螺絲瑕疵之重建能力,本研究提出基於循環生成網路之異常檢測模型,以檢測螺紋瑕疵影像。為了探討噪聲特徵對於異常檢測模型的影響,提出的方法將與自編碼器及去噪自編碼器做比較。根據結果顯示,提出的基於循環生成網路之異常檢測模型優於其他兩種去噪方法。本論文提出的方法可提升模型重建影像之品質。此外,CAE、U-Net及 CGAN
三種編碼解碼模型亦用於驗證提出方法的泛用性。為了達到即時檢測,本研究方法進一步將提出方法部署於運算神經棒,以加速 AI模型於邊緣装
置之推論速度。本論文提出方法能以高準確率、低過殺率、低漏檢率、高速度於檢測螺絲表面瑕疵。
闚鍵字:螺絲影像、自動化光學檢測、深度學習、異常檢測、邊緣運算
Abstract
Defect detection plays an
important role in assessing the surface quality of screw products.
Defective screw products greatly affect the mechanism of the screw,
thereby affecting its safety. Investigations into the methods of
defect inspection are important for screw products. This dissertation
studies the screw surface defect detection based on deep learning
approaches, with a view to detecting defect thread patterns in real
time. Because of different capturing positions with different
appearance results of the thread features, and inconsistent real-world
variables such as lighting and shadow changes, it is challenging for
traditional image processing techniques of adjusting thresholds and
selecting standard template images to achieve high detection accuracy.
Researchers have recently introduced unsupervised anomaly detection
techniques into automated optical inspection systems for surface
defect detection. However, the anomaly networks are usually unable to
reconstruct abnormal images into satisfactory normal images, resulting
in poor performance for defect detection. To strengthen the
reconstruction ability for detecting defect screw images, this study
proposes a novel strategy of CycleGAN-based image denoising method
training on anomaly networks, which can detect defects in thread
images with only a few representing images for anomaly network
training. In order to explore the effects of noise features on the
anomaly network, the normal-based and noise-based denoising methods
trained on the anomaly detection model are utilized as
benchmarks for comparing with the proposed
CycleGAN-based denoising method so as to assess its feasibility.
According to the indicators of PSNR, SIM and IoU, the results
demonstrate that the CycleGAN-based image denoising method is better
than the other two denoising methods. The proposed method in this
dissertation has an ability to restore more textures, whereas the
other two methods blur the textures and edges of reconstructed images,
leading to ineffective detection of the defect regions. The
experimental results also verify that the proposed CycleGAN-based
denoising method can explore more useful features from the inputs and
restore it, so that the reconstruction images can be dramatically
improved. Moreover, three encoder-decoder networks, namely CAE,
U-Net and cGAN, are used to verify the
generalization of the proposed strategy. Both the CAE and U-Net model
obtain similar results for detecting defective regions in the
restored image.
To achieve real-time
screw defect detection, the proposed method is further deployed on a
neural compute stick (NCS) device to accelerate the inference speed
for the Al model. In terms of detection accuracy of the NCS device,
the predictive results with the accuracy of 96%, 96% and 93% for the
three types of screw dataset, respectively, are obtained. The results
also demonstrate that the inspection time of screw defect is six fps
for the NCS device, which meets the industrial standards for real-time
screw defect detection. The proposed method in the dissertation can
detect screw defect, imeges with higth acouraos, low underkil rae. low
overkil rater, an sigpl foueting speed.
Keyword:
Screw image, Automated Optical Inspection, Deep learning, Anomaly
detection, Edge computing.
|