Survey of Scratch Detection Technology Based on Machine Vision

Lemiao Yang; Fuqiang Zhou

doi:10.3788/LOP202259.1415009

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 14 >Page 1415009 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 14, 1415009 (2022)

Survey of Scratch Detection Technology Based on Machine Vision

Lemiao Yang and Fuqiang Zhou^*

Author Affiliations

School of Instrumentation and Optoelectronic Engineering, Beihang University, Beijing 100191, China

show less

DOI: 10.3788/LOP202259.1415009 Cite this Article Set citation alerts

Lemiao Yang, Fuqiang Zhou. Survey of Scratch Detection Technology Based on Machine Vision[J]. Laser & Optoelectronics Progress, 2022, 59(14): 1415009 Copy Citation Text

show less

Simulation experimental results of multi-frame iterative deconvolution algorithm[22]. (a) Original image; (b) segmentation result T0; (c) segmentation result T45; (d) segmentation result T90; (e) segmentation result T135; (f) final result T

Fig. 1. Simulation experimental results of multi-frame iterative deconvolution algorithm^[22]. (a) Original image; (b) segmentation result

T_{0}

; (c) segmentation result

T_{45}

; (d) segmentation result

T_{90}

; (e) segmentation result

T_{135}

; (f) final result

T

Download full size

Fig. 2. Operator templates in 45°, 135°, 180°, 225°, 270°, 315°, horizontal, and vertical directions^[23]

Download full size

Fig. 3. Schematic of two-level labeling technique^[26]

Download full size

Fig. 4. Optimized elliptical Gabor filter and its adjustments^[30]. (a) (b) Optimized elliptical Gabor filter; (c) (d) adjusting to ring Gabor filter; (e) (f) adjusting to ring Gabor filter

Download full size

Fig. 5. Example of a typical CNN architecture

Download full size

Taxonomy	Method	Advantage	Disadvantage
Manual design feature	Gray distribution statistics	Reflect the regularity and local features of the image	Sensitive to noise，suitable for low resolution images
	Transform domain methods	Strong anti-interference ability to noise and changing illumination	Lack of local information，susceptible to feature correlation
	High- and low-dimensional space mapping methods	Strong anti-interference ability to noise，strong adaptability	Poor detection for low contrast or fine scratches

Table 1. Advantages and disadvantages of different scratch detection methods based on manual design feature

Taxonomy	Method	Advantage	Disadvantage
Deep learning	Supervised learning	Has high detection accuracy，less susceptible to light and noise	Requires a large number of labeled images as training data sets
Deep learning	Unsupervised learning	Does not require tagging data sets and human intervention	Susceptible to light，noise and initial values of network parameters

Table 2. Advantages and disadvantages of different deep learning scratch detection methods