Optimizing a Spatial Ring Filter for Edge Extraction Using Convolutional Neural Network

The effectiveness of using convolutional neural networks to optimize the parameters of a spatial-frequency ring filter that provides contrasting edge detection is investigated. To create a data set, arbitrary images in the form of test objects and their Fourier transform are used. It was found that, value regardless of the internal and external radius, the intensity maximum is detected in the test figure corners of a square and a triangle. However, these values affect the uniformity of energy distribution along the contour of the figures. The energy distribution along the contour of the test circle figure occurs in the same way, virtually size regardless of the internal and external annular diaphragm radius. As for the contour width, it increases in direct proportion to the inner radius size. A convolutional neural network with 8 layers was trained. The images were classified into two groups according to the required contrast in order to determine the optimal parameters of the bandpass filter for identifying edges in an arbitrary test image. The criterion for dividing the training set into two classes is the specified contrast threshold value. After 10 epochs of training the convolutional neural network, an accuracy rate of 0.836 was obtained for the “hook” test image.