ICACT20220350 Slide.25
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
Thanks for your attention! Now! Floor is open for Q&A. Welcome any comment or question.
|
|
ICACT20220350 Slide.24
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
References
|
|
ICACT20220350 Slide.23
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
The quality of the cast product is one of the important manufacturing industry. In this paper, we proposed an effective approach to surface defect classification through data from cast products. In image detection and classification research, deep learning model is developing with improve performance. we will research effective methods for detecting surface defects of products, Improved Deep Learning Models, applying various denoise techniques, and quality inspection.
|
|
ICACT20220350 Slide.22
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
The last is the Conclusion.
|
|
ICACT20220350 Slide.21
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
As shown in Fig. 5(a), it can be seen that when the Xception model is evaluated with normal data, the accuracy is 99.16%. (b) shows an accuracy of 95.78% when tested with Gaussian noise applied. (c) shows that 97.62% of the Xception model was measured as an experimental result by wavelet denoising of the noise data. From the experimental results, it can be seen that the image is improved by wavelet denoising when there is a noise filter. As shown in the figure, Epoch was set to 25 to perform the experiment. In the graph, the red line indicates the accuracy of the test, and the green line indicates the loss value of the test progress.
|
|
ICACT20220350 Slide.20
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
first is the precision, which is the ratio of true to what the model classifies as true. And Recall is the ratio to the truth that the model will predict. Recall and precision are trade-offs, and the expression is (2). The accuracy of a model is the most intuitive indicator. The formula is expressed as (3).
|
|
ICACT20220350 Slide.19
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
The data set used in the experiment is a submersible pump impeller among casting products. A submersible pump impeller is a pump component that promotes the flow of a fluid. The Casting Image on the left is a Normal product, and the Image on the right is a Fault product.
|
|
ICACT20220350 Slide.18
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
Next is the Experiment and Result.
|
|
ICACT20220350 Slide.17
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
In this experiment, the wavelet coefficient thresholds of the two methods were selected and proceeded. There are two different methods for wavelet coefficient threshold selection, BayesShrink and VisuShrink. The VisuShrink approach uses a single universal threshold for all wavelet detail coefficients.
The second is the BayesShrink algorithm. It is an adaptive approach to the wavelet soft threshold in which a unique threshold is estimated for each wavelet subband.
|
|
ICACT20220350 Slide.16
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
A Gaussian filter is applied to the image data to assume that the noise emitter is collected for quality inspection in the industrial field.
Gaussian noise usually occurs during image compression and transmission. When the image is compressed, the image is reduced, and when it is restored to its original state, a value generated through an error rather than the existing image is entered due to various factors. In order to proceed with this experiment, a Gaussian noise filter was applied to the dataset.
|
|
ICACT20220350 Slide.15
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
Xception consists of three structures.Fig.3.
Entry flow has 2 Conv layers and 6 Separable Conv layers.
Middle flow consists of 3 Separable Conv layers and repeats 8 times. Exit flow consists of 4 Separable Conv layers. And there is batch normalization at the end of all conv layers and separable conv layers. And Max Pooling and ReLu functions are used.
|
|
ICACT20220350 Slide.14
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
In this paper, We proposed Defect Detection Model for Casting Product using Wavelet Transform Denoise and Xception. The overall flow of the proposed structure and model to refer the implementation process is shown in Figure.2.
|
|
ICACT20220350 Slide.13
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
Next is the Proposed Model.
|
|
ICACT20220350 Slide.12
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
F. Chollet proposed a CNN model with Inception as its underlying structure. It is a structure in which Depthwise Separable Convolution is repeatedly stacked together with Residual. Xception consists of three structures.
|
|
ICACT20220350 Slide.11
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
The application of the wavelet threshold for denoise in the image proceeds as follows 3 formula. Wavelet coefficients are obtained by wavelet decomposition of noisy input data. and Wavelet Threshold The best estimate is obtained through the wavelet coefficients. When the noise filter is removed, the value can be obtained by inverse wavelet transform of the wavelet coefficients.
|
|
ICACT20220350 Slide.10
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
The Fourier transform is a function that transforms point data to facilitate frequency analysis, and as a Fourier function containing only information about a single frequency, it was difficult to apply it to various fields. Wavelet can analyse signals in all aspects in space and frequency, and has excellent performance in continuous control.
|
|
ICACT20220350 Slide.09
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
The quality of precision casting products requires accurate quality inspection because quality errors occur due to micrometer units. In the casting process, quality inspection is carried out through warehousing inspection, dimensional inspection, hardness inspection, and tissue inspection.
At this time, the surface is verified for appearance defects such as stains, foreign substances, and scratches, and the surface inspection is performed with a vision system and a visual inspection. Recently, machine vision technology based on artificial intelligence plays a very important role in detecting defects and defects of parts in manufacturing production lines. In this regard, vision systems are being studied in various fields by combining image processing technology and artificial intelligence.
|
|
ICACT20220350 Slide.08
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
Quality control is to maintain and improve quality by finding the cause of defective products through statistical methods. also Accurate quality inspection is necessary to manage good quality.
The following inspections are carried out in the quality inspection process. The inspection process consists of import inspection, purchase inspection, process inspection, final inspection, and shipment inspection.
|
|
ICACT20220350 Slide.07
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
Second is the Related Work.
|
|
ICACT20220350 Slide.06
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
In order to make a casting product, the quality inspection must be carried out strictly according to the specification standard. However, Depending on the factory environment and Aging equipment, low-quality data may be extracted and inaccurate results.
Therefore, To solve the above-mentioned problem, we propose to improve defect detection by denoise extracted from industrial sites through Wavelet Transform and Xception.
|
|
ICACT20220350 Slide.05
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
Casting is the basic process of manufacturing. It is a method of forming a shape by pouring a liquid material into a mold and hardening it. Casting feature is that it is easier to create a detailed and three-dimensional shape than forging, which is produced by applying pressure to the metal surface. Therefore, it has been widely applied to the production of handicrafts and parts with complex shapes. Recently, the heat resistance of casting has increased and the structure control technology has been developed a lot.
|
|
ICACT20220350 Slide.04
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
Since the advent of the 4th industrial revolution, the smart manufacturing industry is continuously growing due to digital transformation, and smart convergence technology is expanding in line with the era of change and speed.
In the manufacturing industry, the quality of manufacturing is one of the important processes to meet the needs of consumers, and quality improvement through quality control is a very important factor.
|
|
ICACT20220350 Slide.03
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
First is the introduction.
|
|
ICACT20220350 Slide.02
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
The presentation will proceed from the following order of contents.
|
|
ICACT20220350 Slide.01
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
 |
Hi everybody, my name is Juyong park. For ICACT 2022, I have cooperated with two partners from Sungkyunkwan University. Mr. Seokju Oh and Professor Jongpil Jeong.
Our presentation today is about a Defect Detection Model for Casting Product using Wavelet Transform Denoise and Xception.
|
