In this paper, we describe an approach to inspection of solder joints on printed circuit boards by using a circular illumination technique and a neural network classifier. The illumination technique, consisting of three tiered circular colour lamps and one colour camera, gives good visual cues to infer 3D shape of the solder joint surface. A general aspect of this inspection is that the shape of the solder joint tends to greatly vary according to soldering conditions. Due to this, a neural network classifier based on a supervised version of Kohonen learning vector quantization (LVQ) is proposed to automatically and efficiently make classification criteria of the solder joint shapes according to their quality. The practical feasibility of the proposed approach is demonstrated by building a prototype inspection machine and testing its performance.