Yu. A. Bolotova, A. K. Kermani, V. G. Spitsyn
Authors designed 2-levels network based on hierarchical temporal model for object recognition task. The network was approved on printed and handwritten symbols on monochrome and colored background recognition. Initial pictures were preprocessed by Gabor filter. Filter parameters were obtained experimentally. Two ways of forming the input data for the network were designed. The way of taking into account different orientations leads to better recognition results.
The experiment was held on 2 samples of pictures. The first sample consists of 40 categories of printed pictures on a white background for train and on colored background for test. The size of pictures is 32×32 pixels. Train subsample contains 122 pictures. Test subsamples are presented by 192 initial pictures and 1077 randomly shifted pictures. The second sample represents MNIST database of handwritten digits (10 categories). The train sample’s size is 60000 pictures and the test sample’s size is 10000 pictures size of 29×29 pixels.
In the current work the influence of modeled saccadic movements on the recognition result was also investigated. The recognition results of the network trained on video sequences is better than the network trained on separate pictures. Finally, parallelization of main functions of the system speeds up the processes of training and testing.