This paper proposes a novel defect detection system that combines unsupervised and supervised learning methods to identify and classify defects in visual data, particularly in gear wheel images. The system, called U2S-CNN, consists of three phases: anomaly detection using an autoencoder, clustering of anomalies using DBSCAN, and classification of anomalies using a pre-trained Xception network. The autoencoder is used to reconstruct images and identify anomalies by comparing the original and reconstructed images. Clustering algorithms are then used to group these anomalies into regions of interest (RoIs), which are subsequently classified to determine the type of defect. The system was tested on a dataset of 78 images, resulting in the detection of 177 regions and 205 damaged areas, with 108 regions correctly classified and 69 incorrectly labeled. The system provides a more accurate and efficient alternative to traditional methods like YOLO, autoencoders, and transformers, as it not only detects anomalies but also classifies them into specific defect categories. The study highlights the effectiveness of combining unsupervised and supervised learning for defect detection, particularly in industrial applications where defects may vary in shape, position, and color. The results demonstrate that the U2S-CNN system can accurately identify and classify defects, offering a promising solution for industrial inspection systems.This paper proposes a novel defect detection system that combines unsupervised and supervised learning methods to identify and classify defects in visual data, particularly in gear wheel images. The system, called U2S-CNN, consists of three phases: anomaly detection using an autoencoder, clustering of anomalies using DBSCAN, and classification of anomalies using a pre-trained Xception network. The autoencoder is used to reconstruct images and identify anomalies by comparing the original and reconstructed images. Clustering algorithms are then used to group these anomalies into regions of interest (RoIs), which are subsequently classified to determine the type of defect. The system was tested on a dataset of 78 images, resulting in the detection of 177 regions and 205 damaged areas, with 108 regions correctly classified and 69 incorrectly labeled. The system provides a more accurate and efficient alternative to traditional methods like YOLO, autoencoders, and transformers, as it not only detects anomalies but also classifies them into specific defect categories. The study highlights the effectiveness of combining unsupervised and supervised learning for defect detection, particularly in industrial applications where defects may vary in shape, position, and color. The results demonstrate that the U2S-CNN system can accurately identify and classify defects, offering a promising solution for industrial inspection systems.