An Improved Binarized Weight Classification Network

摘要

For the visually impaired people, the use of electronic stimulation devices instead of eyes to capture external data is a feasible scheme to assist activities of the blind. In the field of image classification, deep learning has demonstrated a powerful ability, which is reliable to assist the blind people. Deep learning neural networks, using floating-point computing, has high requirements on the storage space and computing power of devices. BinaryNet constrains the weights and activations to -1 and 1, greatly reducing the need for storage space on the neural network and accelerating the computation. However, compared to other networks, BinaryNet performs poorly on the benchmark set. This paper proposes an improved binarized network IBNet. IBNet applies multi-branch convolution, and each branch is combined with convolution kernels in different sizes. IBNet shows state-of-the-art results on different benchmark sets. The test error on the mnist data set is less than 0.8%, and the test error on cifar10 only 0.02%. In the indoor environment database designed by the author, the test error is only 0.2%. The IBNet parameter is less than 75,000, while that of BinaryNet has reached more than 10 million. Moreover, the speed of image processing by IBnet has reached 1,300 frames. Therefore, IBnet owns a huge advantage in deploying hardware.