This thesis explores possible improvements to the existing defect detection sys-tems used by Tata Steel Europe in South Wales. Exploring the possibilities of ac-curacy increases and further uses for in-house surface quality systems presents Tata Steel Europe with methods and considerations on how to improve their current surface defect management tools. This research considers three dif-ferent aspects of the defect management process and tests alternatives to the systems either currently used or not yet in place.The first section considers an ensemble object detection method using Gabor filters, histograms, and a random forest classifier. The algorithm was applied originally as an ensemble method and then as a case-study it was split into two combinations of the methods to determine what level of ensemble complexity was optimal. The images used were a set of various types of steel defect im-ages provided by Tata Steel Europe. The ensemble method achieved acceptable results compared to the existing systems, but the medium-complexity method was optimal regarding overall accuracy, false negative rate, and speed. The sec-ond section used a set of weld hole images which were split into three quality grades by defining characteristics of the feature. These images were used to test three different neural networks, an R-CNN version of GoogLeNet, and Faster R-CNN versions of ResNet-50 and ResNet-101. These networks were tasked with classifying the quality of weld holes in steel. Accurately detecting weld holes is vital in steel production as certain production processes require speed changes for welds. As the hole punches degrade over time, so does the quality of the weld hole. When weld holes are of poor quality, they can be missed or wrongly detected. All three networks detected the weld holes very well, but classifying the quality grade was approximately 60% accurate for both ResNets and 79% accurate for the GoogLeNet. These tests highlighted the importance of data quantity and quality, including lack of bias in data. The third section looks at how colour filters and greyscale methods can affect the images used for detection and classification. An investigation into coloured light sources has not been fully explored at Tata Steel in South Wales before. Having worked with the image data for sections one and two, it highlighted how important the quality of these images is. Steel defect samples were supplied, alongside their lab-confirmed defect label. Scans were taken of clean and oiled steel samples with different coloured filters and a variety of common greyscale methods were used to turn these images from RGB colour to greyscale images. The greyscale values of defect to clean steel were calculated for a specific region of interest on each steel sample. This value was used as a measure of contrast between clean and defect steel. The yellow filter with the Decolorize method produced the highest contrast image, higher than using no filter at all. For oiled sam-ples, using no filter with the Decolorize method produced the best contrast and the orange filter with the Decolorize method produced the best contrast out of the filtered images. The greyscale methods used had significant effect on the contrast of the image.The significance of this thesis is that it informs of the difficulties in developing surface inspection of steel defects due to several factors, such as environment and variation of defect shape, size, colour, and frequency. The work undertaken in this study has highlighted the need for a larger pilot line to further research both the capturing of the perfect image, and finding the most accurate detection and classification methods for each grade of steel.
Continuous Vital Monitoring During Sleep and Light Activity Using Carbon-Black Elastomer Sensors
