Randomized nonlinear two-dimensional principal component analysis network for object recognition

In order to capture nonlinear structures within data and more representational image features, this paper investigates a multi-stage convolutional neural network with predefined filters. The first two stages are the cascaded blocks consisted of random Fourier mapping, two-dimensional principal component analysis and activation operation. Among that, the approximate method based on Gaussian kernel is used to map the original image to random feature space. Subsequently, convolution filters are learned by two-dimensional principal component analysis. Next, the batch normalization and Gaussian linear error unit activation operation are followed. Afterward, the maximum pooling is utilized to further reduce dimensions of intermediate features. With binary hashing and encoding, the statistical histogram will be obtained and served as the higher-order feature of original image. Experiments have been carried out around the task of object recognition, and quantitative results demonstrate the proposed network has significantly advantageous both in terms of accuracy and computational time compared to the existed algorithms.