| The research presented in this report deals with a particular problem of image segmentation analysis, specifically the separation of objects of interest from background. We present a neural network approach to find an optimal-edge detector able to segment a wide variety of images. Next we present a more robust and unsupervised search technique through a genetic algorithm. Finally we use the strength of a neural network to quickly compute segmentation, combined with the genetic algorithm output for training when needed.; Our first technique uses classical Frei-Chen filters to preprocess input for a neural network trained to detect object edges. This technique is simple in that it contains a receptive field of 5 x 5 image pixels and produces an output of 2 x 2 pixel area. This receptive field scans the input image of any size.; Our second technique uses a classical approach based on state-space techniques for segmentation. These techniques are used in a genetic algorithm to search the state space with better certainty of finding a solution.; In our last presentation, we combine the strengths of both previously stated techniques. The goal is to leverage more out of the learning that has been embedded in the trained neural net, and the exploratory nature of the genetic algorithm. We accomplish this through a mesh of pre-trained neural nets, identifying incorrect responses through a systematic approach, and correcting the network weights localized to the faulting nets with the genetic algorithm output. This technique has been tested on many simulated and real images. |
