All welding processes can give rise to defects, which weakens the joint and can eventually lead to the failure of the welded structure. In order to inspect welds for detects, without affecting the usability of the product, non-destructive testing (NDT) is needed. NDT includes a wide range of different techniques, based on different physical principles, each with its advantages and disadvantages. The testing is often performed manually by a skilled operator and in many cases only as spot-checks. Today the trend in industry is to move towards thinner material, in order to save weight for cost and for environmental reasons. The need for inspection of a larger portion of welds therefore increases and there is an increasing demand for fully automated inspection, including both the mechanised testing and the automatic analysis of the result. Compared to manual inspection, an automated solution has advantages when it comes to speed, cost and reliability. A comparison of several NDT methods was therefore first performed in order to determine which methods have most potential for automated weld inspection. Automated analysis of NDT data poses several difficulties compared to manual data evaluation. It is often possible for an operator to detect defects even in noisy data, through experience and knowledge about the part being tested. Automatic analysis algorithms on the other hand suffer greatly from both random noise as well as indications that originate from geometrical variations. The solution to this problem is not always obvious. Some NDT techniques might not be suitable for automated inspection and will have to be replaced by other, better adapted methods. One such method that has been developed during this work is thermography for the detection of surface cracks. This technique offers several advantages, in terms of automation, compared to existing methods. Some techniques on the other hand cannot be easily replaced. Here the focus is instead to prepare the data for automated analysis, using various pre-processing algorithms, in order to reduce noise and remove indications from sources other than defects. One such method is ultrasonic testing, which has a good ability for detecting internal defects but suffers from noisy signals with low spatial resolution. Work was here done in order to separate indications from corners from other indications. This can also help to improve positioning of the data and thereby classification of defects. The problem of low resolution was handled by using a deconvolution algorithm in order to reduce the effect of the spread of the beam.The next step in an automated analysis system is to go beyond just detection and start characterising defects. Using knowledge of the physical principles behind the NDT method in question and how the properties of a defect affect the measurement, it is sometimes possible to develop methods for determining properties such as the size and shape of a defect. This kind of characterisation of a defect is often difficult to do in the raw data, and is therefore an area where automated analysis can go beyond what is possible for an operator during manual inspection. This was shown for flash thermography, where an analysis method was developed that could determine the size, shape and depth of a defect. Similarly for laser ultrasound, a method was developed for determining the size of a defect.