This technology is a hardware implementation of hand segmentation and recognition system with deep neural network model. By using separable convolution layer and attention module to both increase the execution speed and evaluate performance. The implemented system can achieve the frame rate of 52.6fps and 65.6 GOPS.

The gesture recognition of this design is tested in the OUHANDS data set. When the model size is 1.07MB, the recognition rate can reach 89.25%. Implemented on the ZCU106 development board, the performance reaches 52.6FPS and 65.6 GOPS, and the performance after quantifying on-chip memory is better than the existing depth separable convolutional hardware accelerator, which can reach 7.01 GOPS per Mb of on-chip memory.

This technology uses a deep neural network to implement the hand gesture recognition method, which have a very good recognition rate in a complicated background using only a single CMOS camera. With the proposed neural network hardware accelerator, it is very suitable for use in the smart home appliances which can recognize gestures very quickly and provide a more convenient operating environment for the user.