The alpaca fiber textile industry in Peru faces the challenge of modernizing its production chain to maintain competitiveness. In the early stages, such as fiber opening, quality control remains manual, resulting in slow, subjective, and error-prone processes that reduce the commercial value of the fiber. In this context, visual inspection systems based on computer vision represent a promising alternative to standardize quality and optimize efficiency. This research proposes an automatic system for detecting impurities in alpaca fiber (dirt and straw) through deep learning. A proprietary dataset of 1,524 annotated and validated images was constructed. The methodological pipeline included preprocessing, data augmentation, and the training of YOLOv8 and YOLOv11 models under transfer learning. Experimental results show competitive performance: YOLOv11 achieved 90.4% precision in straw detection and 87.5% in dirt detection, while YOLOv8 obtained higher recall (87.4%) in straw detection. The comparative analysis revealed that YOLOv11 provides a better balance between precision and recall for dirt detection, whereas YOLOv8 demonstrated greater efficiency in convergence. These findings confirm that computer vision can automate critical inspections, enable scalability, and improve the quality of the final product.