Welding techniques are frequently used in modern industrial facilities. However, the quality of weld is far from perfect. The weld produced may have different types of defects, hence the use of non-destructive testing (NDT) methods is ubiquitous to examine weld quality without causing any alteration in the weld structure. One of the most popular NDT methods used is radiography. Traditionally, the inspection of radiographic images is carried out by visual inspection, which leads to more time and labor-intensive process, as well as being prone to errors. The introduction of computer vision and machine learning techniques allow the interpretation of radiograms to be automated that makes the inspection of radiograms more reliable, reproducible and faster. In this doctoral thesis, new techniques for segmentation, detection and classification of welding defects are proposed to enable automated weld quality assessment. The new technology developed will contribute to both real time and offline automated analysis of weld quality, enabling better interpretation and a rapid correction of the defects. Therefore, a fundamental step to evaluate the quality of the weld is to identify perfectly the geometry of the welding defects.There are different stages for weld defects analysis. Among them, the image segmentation stage which is one of the most significant problems in the pattern recognition process due to the structural complexity of the radiographic images and the often insufficient contrast to extract the region of interest without any a priori knowledge of its shape and location. Given the influence of segmentation in the detection of defects, a section in this thesis is devoted for enhancing the performance of segmentation in the case of weld defect radigraphic images specifically for defects of porosity and lack of penetration, which are the subject of this investigation. This section is followed by the defects detection step which allows determination and interpretation of defect's shape. The final section focuses on determining effective models basing on deep learning technology making it possible to characterize defects so that they become easily identifiable class elements.