A neural network for detecting deges in an image

摘要

This research paper describes a neural network architecture based on the instar and outstar networks of Adaptive Resonance theory to enhance and detect edges in an image. The network can discern edge pixels in natural scenes and catagorize them into one of 32 outputs representing differences of 11.25 degrees. The network is built on a series of instar and outstar pairs with a competitive hard limiting network between. The instar network is organized as 32 neurons with 25 inputs each which requires a 5-by-5 subwindow of the input image. A unique weighting and summation techniqeu are used for the instar neurons to decrease the calculation time for the network. The weighting and summation functions produce an asymmetric output so a radial basis network is used to scale and reshape. The iamge intensity levels are contrast enhanced by the network as it detects edges and the feedback mechanism merges pre- and post-neurally processed images. The network was developed with synthetic edge images to determine the best parameters for the network. Both synthetic and photographic iamges were presented to the network to test the final configuration. Piece-wise linear edge elements are correctly classified as to one of 32 angles.