Autonomous segmentation and neural network texture classification of IR image sequences

Presents a technique for the texture classification of segmented regions in sequences of infrared images taken from low flying aircraft. Sequences of IR images have been segmented using a technique which integrates edge detection and region segmentation into a single, inherently parallel process. A temporal component is used to ensure that the segmentation is consistent in time. The texture of the segmented regions is characterized using discrete Hermite functions and property co-occurrence matrices. This results in a low order feature vector in which the zeroth coefficient describes the Gaussian noise of the region while the higher orders describe the texture. Neural networks are used to perform an initial classification of the segmented regions, this is then refined using spatio-temporal relaxation labelling to ensure consistency of interpretation, both spatially within an image and temporal within the sequence