Image Classification with A-MnasNet and R-MnasNet on NXP Bluebox 2.0

Bluebox 2.0 by NXP Semiconductors, which has goal of enabling autonomy in vehicles for ADAS applications, is used to enhance car capabilities to perform sensor fusion and run AI algorithms. It focuses on sensor data coming from radars, lidars, and cameras. This research focuses on enabling computer vision application, Image Classification, by implementation of Convolutional Neural Networks in Bluebox 2.0. In this paper, two CNN architectures namely A-MnasNet and R-MnasNet have implemented on Bluebox 2.0. These models have been derived by Design Space Exploration of MnasNet, a CNN architecture, proposed by Google Brain team in 2019. These models have been trained and tested on CIFAR-10 dataset. The model size and accuracy of A-MnasNet are 11.6 MB and 96.89% and that of R-MnasNet are 3 MB and 91.13% respectively. They outperform the MnasNet architecture which has an accuracy of 80.8% and a model size of 12.7 MB. These neural networks can also be used to perform other computer vision applications.

[1]  Duo Li,et al.  HBONet: Harmonious Bottleneck on Two Orthogonal Dimensions , 2019, 2019 IEEE/CVF International Conference on Computer Vision (ICCV).

[2]  Mohamed El-Sharkawy,et al.  R-MnasNet: Reduced MnasNet for Computer Vision , 2020, 2020 IEEE International IOT, Electronics and Mechatronics Conference (IEMTRONICS).

[3]  Mohamed El-Sharkawy,et al.  A-MnasNet: Augmented MnasNet for Computer Vision , 2020, 2020 IEEE 63rd International Midwest Symposium on Circuits and Systems (MWSCAS).

[4]  Diganta Misra,et al.  Mish: A Self Regularized Non-Monotonic Neural Activation Function , 2019, ArXiv.

[5]  Quoc V. Le,et al.  AutoAugment: Learning Augmentation Strategies From Data , 2019, 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).