Complex Road Recognition CNN Network Based on Multi-Label Learning
Haoxiang Gan, Han Zhang, Wanzhong Zhao, Yuhan Liu · 2024
The road information preceding the vehicle serves as a fundamental prerequisite for intelligent driving systems. In comparison to other data requirements of autonomous vehicles, road information exhibits a more two-dimensional character. Among the array of sensors currently equipped on autonomous vehicles, it is only the visual sensors that possess the capability to detect and interpret the road surface. The majority of road recognition algorithms rely heavily on multi-classification methods, which, however, tend to introduce significant data imbalance issues. The process of further refining pavement types exacerbates this problem. Based on the Road Surface Classification Dataset (RSCD), this study uses the diffusion model to generate category images with less data, and uses the traditional image augmentation method to augment the original image dataset. Furthermore, a convolutional neural network-based multi-label classification model, incorporating the Zero-bounded Log-sum-exp& Pairwise Rank-based loss functions, is trained to enhance the model's predictive performance and accuracy. This enhanced model ultimately achieved an average accuracy of 86.8%, demonstrating significant improvements in performance.