Search

In recent years, facial expression recognition has gained a growing attention in the field of pattern recognition research. Now many previous researches use deeper model or model integration to improve the recognition rate. For these models are more difficult to train and traditional methods have optimizable recognition accuracy, in this paper, we present a novel single network which combines the classic Resnet with the sobel features which can emphasize edge information effectively. We also adopt 10-crop testing to further improve the performance. The model was tested on the CK+ database and FER2013 database and we observed 95.96% and 71.89% classification accuracy respectively. The results showed that the method achieved better or comparable recognition rate with other state-of art methods.

The traditional bias circuit in an LCD driver, by means of resistors dividing a voltage, consumes power continuously as it works. Based on the properties of liquid crystals with a large parasitic capacitance and resulting in a large time constant during charge and discharge, a novel bias circuit using dynamic resistors is proposed in this paper. The new bias circuit has been applied in a 32/spl times/4 LCD driver. The power dissipation in the biasing circuitry part could be reduced by about 36/spl times/ to more than 100/spl times/.