CM1K Hardware User’s Manual
29
4
The optional recognition stage
The recognition stage of the CM1K is enabled through the RECO_EN pin and can be activated through bit 0 of the
Recognition Status Register (CM_RSR). In such case the recognition stage becomes master and the neurons must
not be accessed by an external master while the BUSY line is high (for more information, refer to timings
constraints in a paragraph below).
If V_EN is low, the data received on the V_DATA bus is directly accumulated in a 256-byte FIFO at each pulse of
V_CLK when V_FV is high.
If V_EN is high, the data received on the V_DATA bus is interpreted as a video signal. The V_LV input signal is then
necessary and defines the number of pixels per line of video. V_FV defines the number of lines per video frame. In
the case V_DATA is not accumulated directly to the FIFO but rather integrated spatially per blocks of pixels. The
calculated average value per block is then accumulated to the FIFO. This operation is called
feature extraction
and
used the 6 registers defining the region of interest and its internal blocks to average.
As soon as V_FV falls, the recognition stage broadcasts the content of the FIFO to the neurons using a series of
Write COMP and one Write LCOMP. It then reads the response of the best match using a Read DIST followed by
Read CAT. This data is latched to the RT_DIST and RT_CAT registers and the DIST_VAL and CAT_VAL lines pulse for
the duration of one G_CLK cycle.
Remark 1
: If the digital input signal is not a video signal (V_EN=0), the V_FV signal must stay low for a minimum of
N + 37 cycles of G_CLK with N being the number of V_DATA sampled during V_FV high. Note that the CAT_VAL
pulse occurs one cycle after the N + 37 cycles. V_FV must be changed at the negative edge of V_CLK.