Compressed MobileNet V3:A Light Weight Variant for Resource-Constrained Platforms

• Published in: 2021 IEEE 11th Annual Computing and Communication Workshop and Conference (CCWC) pp. 0104 - 0107
• Main Authors: Kavyashree, Prasad S P, El-Sharkawy, Mohamed
• Format: Conference Proceeding
• Language: English
• Published: IEEE 27.01.2021
• Subjects: CIFAR-10; Compressed MobileNet V3; Computational modeling; Conferences; Convolution Neural Networks; Depthwise Pointwise Depthwise blocks; Memory management; Mobile handsets; MobileNet V3; Object tracking; Task analysis
• DOI: 10.1109/CCWC51732.2021.9376113

Convolutional Neural Networks (CNNs) are ubiquitous in computer vision applications. This is attributed to their excellent performance in image classification which forms the foundation for many complex tasks such as object localization, object tracking, etc. Despite their huge success, the intensive computation, memory bandwidth, and energy requirements have made it difficult to deploy them in low power and resource-constrained platforms. To overcome this, many researchers have designed compact models achieving a tradeoff between model size and accuracy. MobileNet V3, the latest variant of MobileNets is one of the CNN models complying with this trend [1]. It has a model size of 15.3 MB with a validation accuracy of 88.93% on the CIFAR-10 dataset[2]. In this paper, we have modified the baseline architecture to further reduce its size to 2.3 MB while achieving an accuracy of 89.13%.