Depth-Wise Based Convolutional Neural Network for Street Imagery Digit Number Classification
Okeke Stephen, Young-Jick Jang, Tae-Soo Yun, Mangal Sain · 2019
We present a lightweight convolutional neural network model for street view house numbers recognition and classification. Our ultimate goal is to build a high performance, compact and memory efficient deep learning model for classifying digits of real-world house numbers which is suitable for deployment into devices with limited computational powers. Our model applied world class ground breaking convolution neural network architecture for embedded vision proposed as backbone architecture. The backbone architecture uses a streamlined deep learning architecture with depth-wise separable convolutions in performing computer vision tasks. We performed extensive experiments with high degree hyper parameters turning to increase the mode's accuracy, minimize training and validation loss respectively. In order to improve the model accuracy and performance, we deployed random image argumentation technique and batch normalization process during training. We obtained 99.6% training accuracy and demonstrated the effectiveness of our model on randomly selected samples of the street view test dataset and achieved a commendable accuracy of 92.0%.