This paper presents a novel approach to defect detection in visual data, specifically focusing on gear wheel images. The system combines unsupervised and supervised learning methods to detect and classify defects accurately. The methodology involves three main phases: anomaly detection, clustering, and classification. 1. **Anomaly Detection**: An autoencoder is trained to reconstruct input images without anomalies, identifying differences between the original and reconstructed images. These differences are then clustered using the DBSCAN algorithm, which groups similar anomalies without predefined cluster numbers. 2. **Clustering**: The DBSCAN algorithm identifies clusters of anomalies, allowing for the definition of regions of interest (RoIs) based on the number of points in each cluster. 3. **Classification**: The pre-trained Xception network classifier is used to classify the RoIs into predefined categories, such as damaged edges, scratches, and surface damage. The system was tested on a dataset of 139 images, with 78 correct samples and 61 tested samples. The U2S-CNN network achieved industry-acceptable results in the first two phases but showed room for improvement in the classification phase. The main challenges included accurately detecting small defects and optimizing lighting conditions for better contrast. The paper concludes that the proposed system effectively combines the strengths of unsupervised and supervised learning, providing a more comprehensive and accurate defect detection solution compared to traditional methods like YOLO or autoencoders alone. Future work will focus on improving defect detection accuracy, particularly for small defects, and optimizing lighting conditions to enhance detection capabilities.This paper presents a novel approach to defect detection in visual data, specifically focusing on gear wheel images. The system combines unsupervised and supervised learning methods to detect and classify defects accurately. The methodology involves three main phases: anomaly detection, clustering, and classification. 1. **Anomaly Detection**: An autoencoder is trained to reconstruct input images without anomalies, identifying differences between the original and reconstructed images. These differences are then clustered using the DBSCAN algorithm, which groups similar anomalies without predefined cluster numbers. 2. **Clustering**: The DBSCAN algorithm identifies clusters of anomalies, allowing for the definition of regions of interest (RoIs) based on the number of points in each cluster. 3. **Classification**: The pre-trained Xception network classifier is used to classify the RoIs into predefined categories, such as damaged edges, scratches, and surface damage. The system was tested on a dataset of 139 images, with 78 correct samples and 61 tested samples. The U2S-CNN network achieved industry-acceptable results in the first two phases but showed room for improvement in the classification phase. The main challenges included accurately detecting small defects and optimizing lighting conditions for better contrast. The paper concludes that the proposed system effectively combines the strengths of unsupervised and supervised learning, providing a more comprehensive and accurate defect detection solution compared to traditional methods like YOLO or autoencoders alone. Future work will focus on improving defect detection accuracy, particularly for small defects, and optimizing lighting conditions to enhance detection capabilities.