UnderstAnding Bag of Tricks of Deep Learning-Based Semantic Segmentation in Pavement Crack Detection

Abstract

The rapid development of deep learning has significantly enhanced the performance of models in the detection of pavement cracks, thereby facilitating the deployment of deep learning-based approaches into real-world applications. Nevertheless, it is worth noting that deep learning-based crack detection models represent complex amalgamations of deep learning networks and model training strategies, with the latter frequently being overlooked. Therefore, in this paper, we focus on various techniques in data augmentation, and model deployment stages that are commonly employed in deep learning-based semantic segmentation models. Through extensive experiments, the effectiveness of these techniques in crack detection is evaluated, aiming to provide guidance for subsequent crack detection experiments and project implementations. Consequently, the experiments demonstrate data augmentation methods such as color jittering and CutMix can effectively improve model performance by altering the distribution of the training dataset. Additionally, in case of crack datasets with limited samples and severe class imbalance, loss function selection and pre-training weights can be crucial in model deployment.