Convolutional neural networks for pattern classifying based on parameterized predefined sequence of image filters

Convolutional neural networks (CNNs) are used to solve pattern classification problems. As this algorithm is based on establishing a relationship between an image-shaped input and its related output through the CNN structure, the training stage is a significant process in their working basis. This study develops a new-fangled and explainable algorithm to train CNNs. The input filters in the convolution layers are parameterized to keep the filter structure, implementing traditional and explainable image processing filters within the network topology. A back-propagation scheme updates the parameters in the filters and the fully connected section of the CNN. Several traditional image filters (Sobel, averaging, Gaussian, and directional, among others) are used in CNN with a learning strategy that keeps their kernel structures. The method implies that the training of these networks is applied to a single parameter instead of all coefficients in the filters, reducing the uncertainty about how each filter performs the image analysis in CNN. This approach was compared with traditional CNNs considering the analysis of the computational cost (measured in terms of time and floops required for training) and their accuracy results. Three image databases were used to evaluate the proposed algorithm. Using a cross-validation methodology, the new training algorithm based on the filter parameterization achieved higher accuracy (93.7