Radoslav Skoviera, Ivan Bajla, and Julia Kucerova
machine learning, pattern recognition, hierarchical temporal memory, image saliency
The essential goal of this paper consists in extending the functionality of the bio-inspired intelligent HTM (Hierar- chical Temporal Memory) network towards two capabili- ties: (i) object recognition in color images, and (ii) classifi- cation of objects located in "clutter color images. The for- mer extension is based on development of a novel scheme for application of three parallel HTM networks which sepa- rately process color, texture, and shape information in color images. For the latter HTM extension we proposed a novel system in which HTM is combined with a modified model of computational visual attention. We adopted the results of [1] and [2], and added new elements [3] for the calcula- tion of image saliency maps. The proposed algorithm en- ables to locate individual objects in clutter images automat- ically. For computer experiments a special image database has been created to simulate ideal single object images and cluttered images with multiple objects. The recognition performance of the HTM alone and in combination with a salient-region detection method has been evaluated. The evaluation of the attention subsystem shows promising re- sults in the sense that the system satisfactorily locates sev- eral objects in clutter color images with non-homogeneous background. Our pilot study confirmed that the proposed attention system can improve the HTM’s capabilities for object classification in cluttered images. However, as ex- pected, the system cannot match the HTM’s recognition accuracy achieved on single object images.
Important Links:
Go Back
