Wavelet based method for edge detection of defect on coated board during production proccess

This thesis considers edge detection of defects on a coated board based on the wavelet transform. Properties of the wavelet transform, above all the possibility to represent singularity in the signal with a few coefficients, gives opportunity to realize the efficient edge detector. This thesis gives a detailed description of existing edge detection methods based on differentiation and Gaussian filtering with in-depth review of the wavelet transform techniques. It is analyzed how authors compare suggested methods with other edge detectors. If was found that only few authors use objective evaluation and those comparisons are based on a synthetic image and cannot be applied to real images. Shortcomings of classical edge detector when it is used with coated board images are shown. The most important properties of the wavelet transform are presented with theoretical background of singularity detection. Characteristics of the signal have influence on the edge detector performances and the model of an edge on coated board is developed. Using this model it is shown that better result are obtain when three, instead of originally suggested two scale of wavelet transform are multiplied. Additionally, it is shown that proposed algorithm can be applied on arbitrary image set and only initially scale for multiplication must be determined. Subsequently, the edge detector evaluation methods are shown and analyzed. It is concluded that objective methodology based on the ground truth images must be used. For the evaluation set of 50 defect images on coated board and 50 corresponding ground truth images are created. The comparison is also performed with on available sets of 50 object and 10 aerial images. Testing of proposed algorithm is done by comparing it with the classical and frequently used edge detector: Sobel, Canny and Marr-Hildreth; with two edge detectors based on the wavelet transform and one newly and commonly used edge detector – SUSAN (Smallest Univalue Segment Assimilating Nucleus) detector. The evaluation shows that proposed method for edge detection of defects on coated board outperforms all other mentioned edge detectors. Also, the proposed algorithm is even slightly better when it is applied to other two sets of images and when noise is added on the object and aerial images differences in performances are increased in favor of the proposed method. This thesis presents a method for edge detection which outperforms classical and the mostly used edge detecto