Feedforward Neural Network Based on Convex Optimization Theory and Its Application on Urban Information Extraction from Remote Sensing Images

According to Young inequality of convex function´s conjugate properties, a new error function is constructed for feedforward neural network. The function is convex to both connection weight value and hidden layer´s output, so it has no local minimum points. Training speed of feedforward neural network based on the new error function is fast, and network´s training success rate is high. Its training strategy is: first, fix connection weight values before and after hidden layer, optimize hidden layer´s output; then, fix hidden layer´s output, optimize connection weight values before and after hidden layer; just like so, until error demand is satisfied. The new network is applied to extract urban information from remote sensing images. Compared with traditional image information extraction method (maximum likelihood method), its extraction accuracy rate is much higher.