Wafer-defect maps can provide important information about manufacturing defects. The information can help to identify bottlenecks in the semiconductor manufacturing process. The main goal is to recognize random versus patterned defects. A patterned defect shows that a step in the process is not performed correctly. If same defect occurs multiple times, then the yield can rapidly decrease. This article proposes a method for yield improvement and defect recognition by using a feed-forward neural network. The neural network classifies wafer-defect maps into classes. Each class represents certain defect on the map. The neural network was trained, tested and validated using a wafer-defect maps dataset containing real defects inspired from manufacturing process.