Application of neural networks analysis in image feature extraction

Analysing neural network edge detection (NNED) is being presented in a new method in order to gain a close insight into their internal functionality. A training sets consisting of a limited number of prototype edge patterns is proposed in order to analyze the problem of edge detection. The generated set of vectors simply corresponds to standard situations that are clearly understood as edges or non-edges and also permit a controlled distribution of the edges enclosed. The behavior of neural network edge detector's hidden units, as templates, were analyzed into three gradient component: low pass, gradient, and second-order gradients. Although the above only gives some analysis results for the units in the hidden units, it should be clear that a characterization of the neural network as a whole could also be derived from these results. Visual comparisons are conducted here, by image convolution of the weights templates with a real test image to demonstrate the feature extracted by each hidden node of NNED. Different image features extracted by image convolution could provide a simple verification of the investigated gradient analysis method.