A Neural Computation Model Based on nCRF and Reverse Control Mechanism - Developing Concepts in Applied Intelligence - page 49

4 Results

We confirmed the change range of the size of RF of retina GC according to the
physiological data. So, in the process of experiment, the change range was limited and
cannot enlarge or reduce arbitrarily. We got the schematic diagram (as shown in
Fig. 3) of RF dynamic adjustment process after tracking. In Fig. 3, red means that RF
has reached the final size, and needn't to change; blue means that RF need to enlarge
further; yellow means that RF need to reduce further; the real line denotes the actual
size after this iteration; the broken line denotes the RF size that be replaced gradually
in the process of iteration, which shows the change trail of RF. Observing the final
size of RF (red RF) in Fig. 3, we found that the smallest RFs are always on the border
of the image. In the background wall, the information is uniform within each color
block, so the RFs representing the information within color block reach to the
maximum. The RFs representing the information of the lamb, pillows, cup, flowerpot,
the bottom of the sofa don't change when they just reach the edges. Therefore, from
Fig. 4 we can see that RFs best represent the image information by dynamically
adjusting their size in the limited change range.

Fig. 3. The schematic diagram dynamic
adjustment process.

Fig. 4. The difference in image
understanding between human vision and
machine vision.

5 Discussion

The property of nCRF is not changeless. The nCRF can adjust its filter characteristics
according to the spatial frequency components of images. And it is flexible toward
changing contrasts and brightness of stimulations. Along with the changing of image
spatial properties, nCRF sometimes turn into high spatial frequency filter and
sometimes into low one. The nCRFs of most neurons are facilitatory under the
condition of low contrast or low brightness, which shows that nCRF enhance the RF