A Real-time Detection Algorithm Based on Nanodet for Pavement Cracks by Incorporating Attention Mechanism

Abstract

In order to realize real-time detection of road crack defects on edge devices with low computational power, we propose a Real-time detection algorithm of Nanodet pavement cracks incorporating attention mechanism for a wide range of domestic road crack detection practical needs. Based on the Nanodet model, the algorithm in this paper introduces the attention mechanism into the feature extraction structure and assigns different weights to the features to improve the attention of the network crack target. It adopts the SimOTA dynamic label matching strategy to optimize the training process, so that the model can make full use of the data set. Besides, it uses the Hard-Swish activation function instead of the original ReLU function to speed up the network training convergence speed and improve the robustness of the algorithm, so as to achieve fast and accurate recognition of the key parts of pavement cracks. Through experimental validation on the pavement crack data set, compared with several classical target detection algorithms, there are different degrees of improvement in recognition accuracy and detection speed. In this model, the number of parameters is only 0.96M, the occupied memory is 7.6 MB, the floating point computation is 2.21G, and the average accuracy in the test set reaches 95.51%, the FPS is 123 fps. The model, which is 4.32% better than the original model, improves the detection accuracy of the model while maintaining the lightweight characteristics of the original model, and has good prospects for application deployment.