Optimizing Feature Transform Structure for Object Classification

Abstract

We propose a method to optimize parameters and the processing order of feature transform. CNN has been proposed as a method for optimizing parameters of feature transform. The processing order of convolution，pooling and activation function is given by a user on CNN. Then weights of convolution are optimized. On the other hand，in the proposed method we propose a method to simultaneously optimize both parameters and processing order of feature transform not limited to convolution. To realize this optimization， we formulate it as a combinatorial optimization problem and solve by meta-heuristics. Applying the proposed method to pedestrian classification and experimenting with the benchmark data set confirmed that the proposed method is effective.