EXPERIMENTAL COMPARISON OF CONVOLUTION NEURON NETWORK ARCHITECTURES
DOI:
https://doi.org/10.17770/het2020.24.6742Keywords:
CNN, Convolution Neuron Network, Model, TensorflowAbstract
In this work, authors experimentally compare latencies of convolution neuron network architectures. Authors measured only recognition time. Four architectures were applied in the experiment: AlexNet, AlexNet Separated, MobileNetV1 and MobileNetV2. Models were trained using Fruits360 dataset. The Android mobile application was developed to measure latency on mobile devices. The smallest latency authors obtained using AlexNet Separable model, but the smallest size was provided by MobileNetV2.Downloads
References
Konvolucionāli neironu tīkli vizuālai atpazīšanai [tiešsaiste], [atsauce uz 07.04.2020.]. Pieejams: https://cs231n.github.io/convolutional-networks/
Alexnet arhitektūra [tiešsaiste], [atsauce uz 07.04.2020.]. Pieejams: https://medium.com/@smallfishbigsea/a-walk-through-of-alexnet-6cbd137a5637
A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deepconvolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012
Andrew G., et al. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications, 2017. [tiešsaiste], [atsauce un 07.04.2020.]. Pieejams: https://arxiv.org/abs/1704.04861
M. Sandler, et al. MobileNetV2: Inverted Residuals and Linear Bottlenecks, 2019. [tiešsaiste], [atsauce un 07.04.2020.]. Pieejams: https://arxiv.org/pdf/1801.04381.pdf
H. Mureșan & O. Mihai. Fruit recognition from images using deep learning. Acta Universitatis Sapientiae, Informatica, vol. 10, pp. 26-42, 2018.
Tensorflow Lite [teiššaiste], [atsauce uz 07.04.2020.]. Pieejams: https://www.tensorflow.org/lite
Samsung S8 mobilā telefona specifikācija [tiešsaiste], [atsauce uz 07.04.2020.]. Pieejams: https://www.sammobile.com/samsung/galaxy-s8/specs/
Android Studio laika mērīšanas funkcija [tiešsaiste], [atsauce un 07.04.2020.]. Pieejams: https://developer.android.com/reference/java/lang/System
