<![CDATA[Identification of koi fish (Cyprinus carpio) varieties in aquaculture and ornamental fish industries is commonly performed manually through visual observation, making the process subjective, inconsistent, and inefficient, particularly at large production scales. This study aims to develop an automated image-based detection and classification system for koi varieties using the YOLOv3-Tiny algorithm integrated with OpenCV, capable of operating in real-time conditions. The dataset consists of 3,154 images of six koi varieties—Asagi, Bekko, Hikarimono, Kohaku, Sanke, and Showa—which were expanded to 6,360 images through data augmentation techniques. Image labeling and annotation were conducted using Roboflow, while model training was implemented with the Darknet framework in a Google Colab environment supported by GPU acceleration. System performance was evaluated using mean Average Precision (mAP), loss function analysis, and both static image and real-time video testing. Experimental results demonstrate that the YOLOv3-Tiny model is capable of accurately detecting and classifying koi varieties with stable inference speed suitable for real-time applications. The proposed system enhances objectivity, consistency, and efficiency in koi variety identification and shows strong potential for practical implementation in technology-driven ornamental fish farming and trading industries]]>
Copyrights © 2026
