MinorNet: A Lightweight neural network for battlefield scene classification
Shi Guo, Yong Ni, Ke Xing, Yang Liu, Wei Ni · 2021
Computer vision has a wide range of applications, and the current demand for intelligent battlefields is increasing. However, most research on CNN (Convolutional Neural Network) models did not consider both accuracy and lightweight. To achieve an efficient neural network model for battlefield scene classification, we propose a lightweight and efficient neural network named the MinorNet. The design of MinorN et refers to the advantages and avoids the shortcomings of the previous models. Min bottleneck is an essential part of Minor Net. The Min bottleneck contains Depthwise Separable Convolution, channel shuffle, linear transformation, and residual part. They can further improve model capabilities and efficiency without increasing inference computation. This paper establishes a novel dataset named battlefield scene dataset of 7 different kinds of a real battlefield scenes. As a result, our MinorNet achieves 88.3% with only 2.4M parameters in the battlefield scene dataset, being about 0.5x smaller yet still more accurate 3.7% than the previously most commonly used lightweight place classification model.