Comprehensive optimization of Deeplabv3 for mobile devices using lightweight architectures

Authors: Malyshev P.V.
Published in issue: #6(101)/2025
DOI:
Category: Informatics, Computer Engineering and Control | Chapter: Information Technology. Computer techologies. Theory of computers and systems
Keywords: convolutional neural networks, segmentation, digital image processing, deep learning, optimization
Published: 09.12.2025

Deep convolutional neural networks (CNNs) have become a mainstay for image segmentation tasks, making them indispensable for a variety of applications including autonomous systems and mobile devices. However, high computational resource requirements make the use of models such as Deeplabv3 difficult on resource-constrained devices. In this paper, we investigate how Deeplabv3 can be optimized by replacing ResNet-50 with lighter architectures such as ResNet-18 and MobileNetV2. The reduction in resource consumption without significant loss of model accuracy using the lighter architectures is investigated.

References

[1] Usmani I.A., Qadri M.T., Zia R., Alrayes F.S., Saidani O., Dashtipour K. Interactive effect of learning rate and batch size to implement transfer learning for brain tumor classification. Electronics, 2023, vol. 12, art. no. 964. https://doi.org/10.3390/electronics12040964

[2] Zhou P., Xie X., Lin Z., Yan S. Towards Understanding Convergence and Generalization of AdamW. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46, vol. 9, pp. 6486–6493. https://doi.org/10.1109/TPAMI.2024.3382294

[3] Kolarik M., Burget R., Riha K. Comparing normalization methods for limited batch size segmentation neural networks. 43rd International Conference on Telecommunications and Signal Processing (TSP), Milan, Italy, 2020, pp. 677–680. https://doi.org/10.48550/arXiv.2011.11559

[4] Gou J., Yu B., Maybank S.J. et al. Knowledge distillation: a survey. Int. J. Comput. Vis., 2021, vol. 129, pp. 1789–1819. https://doi.org/10.1007/s11263-021-01453-z

[5] Shorten C., Khoshgoftaar T.M. A survey on image data augmentation for deep learning. J. big data, 2019, vol. 6. https://doi.org/10.1186/s40537-019-0197-0

[6] Cordts M., Omran M., Ramos S., Scharw?chter T., Enzweiler M., Benenson R., Franke U., Roth S., Schiele B. The Cityscapes dataset for semantic urban scene understanding. URL: https://markus-enzweiler.de/downloads/publications/cordts15-cvprws.pdf

[7] van de Ven G.M., Tuytelaars T., Tolias A.S. Three types of incremental learning. Nat Mach Intell., 2022, vol. 4, pp. 1185–1197. https://doi.org/10.1038/s42256-022-00568-3

[8] Liang T., Glossner J., Wang L., Shi S., Zhang X. Pruning and quantization for deep neural network acceleration: a survey. Neurocomputing, 2021, vol. 461, pp. 370–403. https://doi.org/10.1016/j.neucom.2021.07.045