A Comparative Study on the Performance of the Classical and Wavelet Based Edge Detection for Image Denoising on Defective Weld Thermographs

Defect in welded structures is a matter of serious concern. The current practice involves interpretation by inspectors and experts, of radiographs after the weld is completed which is time consuming. With greater emphasis on automation during manufacturing process, automated NDT has gained prominence. One of the Non Destructive Evaluation (NDE) techniques that is being increasingly applied for online monitoring is thermal imaging. Thermal imaging is an advanced NDE technique based on the detection of infrared radiation. IR imaging during welding is based on the fact that a good weld would produce isothermal patterns while a weld with defects or arc misalignment, etc would produce isothermal patterns which are not symmetrical. In automated/ adaptive welding system, IR images are to be captured in real time and based on the isothermal patterns the quality of welding is to be judged. Once thermal features are detected and attributed to defects, it is necessary to quantitatively characterize them for acceptance / rejection. Quantitative characterization requires the application of advanced image processing tools for feature extraction and dimensional measurements. Edge detection is an important image processing operation which helps in detecting the Region of Interest. Traditional approaches that use classical edge detectors work fine with high quality pictures, but often are not good enough for noisy pictures because they cannot distinguish edges of different significance as they primarily focus on the coupling between image pixels on a single scale. Recent developments in Multi Scale Analysis such as Wavelet Transform help to overcome this difficulty. Analysis of the thermal image depicting lack of penetration reveals that compared to the edge detection performance of Classical Sobel & Prewitt filters wavelet based edge detectors provide better results by removing the noise. This paper compares the performance of classical edge detector with the proposed wavelet based edge detection technique that involves multiresolution analysis which significantly improves the results.