Monitoring road surface anomalies is crucial to avoid potential harm to pedestrians, vehicles, and vehicle users. These factors have motivated the development of systems that employ signal processing and machine learning algorithms to detect these irregularities automatically by sensing the vehicle's vibration. The wavelet transform (WT) has been frequently used for signal representation and feature extraction to solve signal classification problems. However, selecting a mother wavelet to perform the WT is challenging due to the need for methods that help to validate the time¿frequency representation. Furthermore, machine learning algorithms require a sizeable sample size to train them, and setting their parameters is considered a time-consuming task. This paper proposes to compute the Continuous Wavelet Transform (CWT) through Generalized Morse Wavelets (GMWs) on vehicle's vertical acceleration data for pavement transverse cracking detection. The parameters gamma and beta of the GMWs and the CWT's voices per octave (VPO) were selected based on two metrics, the Heisenberg area and the Average Reconstruction Mean Squared Error. The scalogram generated through the CWT and GMWs was used to fine-tune pre-trained convolutional neural networks (CNNs) through transfer learning, such as GoogLeNet, SqueezeNet, and ResNet18. The proposed methodology set 1 VPO for the CWT, while the parameters gamma and beta of GMWs were set as 3 and 1.3333, respectively. The 5-fold cross-validation results of fine-tuning the CNNs showed that SqueezeNet provided a higher average validation sensitivity (84.3290 ± 5.5704) than GoogLeNet and ResNet18; however, the average validation specificity of SqueezeNet (92.7061 ± 7.2403) was lower.
