Abstract <p>Thermographic data is highly class-imbalanced and scarce while considering the temporal thermal profiles for automatic defect detection using deep learning. Training a supervised deep learning model requires a significantly equal amount of data. Unsupervised deep learning with one-class classification approaches has recently been introduced in thermography for composite inspection. This article proposes an autoencoder-driven anomaly detection model for automatic defect detection in quadratic frequency modulated thermography. The proposed model utilizes the pretrained stacked denoising convolution autoencoder (SDCAE) to extract deep features and feed them to a local outlier factor (LOF) for defect detection. This work analyzes the performance of the proposed SDCAE-LOF on a quick-responsive mild steel specimen with artificially embedded defects of various sizes at different depths. The performance is compared with the CNN-based deep anomaly detection model and other autorncoder models using multiple metrics to confirm the superior defect detection capability of the proposed method.</p>