This thesis is concerned with the role of symmetry in low-level image segmentation. Early detection of local image properties that could indicate the presence of an object would be useful in segmentation, and it is proposed here that approximate bilateral symmetry, which is common to many natural and man made objects, is a candidate local property. To be useful in low-level image segmentation the representation of symmetry must be relatively robust to noise interference, and the symmetry must be detectable without prior knowledge of the location and orientation of the pattern axis. The experiments reported here investigated whether bilateral symmetry can be detected with and without knowledge of the axis of symmetry, in several different types of pattern. The pattern properties found to aid symmetry detection in random dot patterns were the presence of compound features, formed from locally dense clusters of dots, and contrast uniformity across the axis. In the second group of experiments, stimuli were designed to enhance the features found to be important for global symmetry detection. The pattern elements were enlarged, and grey level was varied between matched pairs, thereby making each pair distinctive. Symmetry detection was found to be robust to variation in the size of matched elements, but was disrupted by contrast variation within pairs. It was concluded that the global pattern structure is contained in the parallelism between extended, cross axis regions of uniform contrast. In the third group of experiments, detection performance was found to improve when the parallel structure was strengthened by the presence of matched strings, rather than pairs of elements. It is argued that elongation, parallelism, and approximate alignment between pattern constituents are visual properties that are both presegmentally detectable, and sufficient for the representation of global symmetric structure. A simple computational property of these patterns is described.