Academic Journal of Computing & Information Science, 2023, 6(3); doi: 10.25236/AJCIS.2023.060305.
Yankaiqi Li
St. Cloud State University, Saint Cloud, Minnesota, 56301, USA
In the field of motor vehicle recognition, the use of neural network models has become the standard, and the tuning of hyperparameters and loss functions has been shown to be an effective way to improve the performance of these models. However, when using classical convolutional network architectures (e.g., ImageNet) and training them on motor vehicle images with random labels, the overparameterization problem can lead to suboptimal results and an increased risk of recognition failure. P. Ismailova et al. proposed a solution to this problem with the use of weight averaging, which resulted in the development of the simple and effective Stochastic Weight Averaging (SWA) optimizer. In this paper, we apply the SWA method to optimize the original recognition model and demonstrate significant improvements in accuracy through the use of different learning rate schemes with various traditional optimizers. We also identify suitable hyperparameter values to enhance the model's generalization abilities through several experiments, reducing the waste of resources in the motor vehicle recognition task and improving the recognition accuracy of fine-grained images in general, thus increasing the efficiency of related fields.
Random Average Weighting, Fine-Grained Recognition, Generalization capability, Computer Vision
Yankaiqi Li. An Improved Vehicle Fine-Grain Identification Algorithm Based on Stochastic Weight Average. Academic Journal of Computing & Information Science (2023), Vol. 6, Issue 3: 35-41. https://doi.org/10.25236/AJCIS.2023.060305.
[1] Izmailov P., Ashukha A., Gulin A., & Vetrov D. (2018). Averaging weights leads to wider optima and better generalization. In Conference on uncertainty in artificial intelligence (pp. 412-421).
[2] Keskar N. S., Dauphin Y. N., & Socher R. (2017). On large-batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836.
[3] Krause J., Stark M., Deng J., & Fei-Fei L. (2013, December). 3D object representations for fine-grained categorization. In 2013 IEEE international conference on computer vision workshops (pp. 554-561). IEEE.
[4] Information on: https://github.com/AustinReese/UsedVehicleSearch.
[5] Timgaripov. (n.d.). Timgaripov/SWA: Stochastic weight averaging in Pytorch. Retrieved from https://github.com/timgaripov/swa.
[6] Zhang, H., Zhu, Y., Hu, S., He, T., & Sun, J. (2020). SWA object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 5104-5112).
[7] Zhang X., Shi Z., & Chen L. (2020). Expression recognition method based on cascade network optimized by SWA. Electronic Science and Technology, 33(9), 16-20.