A Lightweight Binarized Convolutional Neural Network Model for Small Memory and Low-Cost Mobile Devices

In recent years, the high cost of implementing deep neural networks due to their large model size and parameter complexity has made it a challenging problem design lightweight models that reduce application costs. The existing binarized suffer from both memory occupancy big number trainable params they use. We propose convolutional network (CBCNN) address multiclass classification/identification problem. use binary weights activation. show experimentally using only 0.59 M is sufficient reach about 92.94% accuracy on GTSRB dataset, similar performances compared other methods MNIST Fashion-MNIST datasets. Accordingly, most arithmetic operations with bitwise are simplified, thus used accesses reduced by 32 times. Moreover, color information was removed, which also drastically training time. All these together allow our architecture run effectively in real time simple CPUs (rather than GPUs). Through results we obtained, despite simplifications removal, achieves classical CNNs lower costs embedded deployment.