Abstract:With the continuous expansion of the aquaculture industry and the advancement of technology, the production model of the industry has gradually shifted toward greater modernization, mechanization, and automation. This transformation has become the primary trend in industry development, signaling the movement from traditional farming methods to intelligent and automated approaches. Fish fry, as a crucial link in the aquaculture supply chain, play an important role in the entire industry. Accurate fish fry counting is essential for managing the industry effectively, conducting scientific feeding practices, controlling stocking density, and ensuring fair pricing and transparent transactions in the sale of fry. In traditional aquaculture practices, fish fry counting mainly relies on manual methods, which are not only time consuming and labor intensive, but also prone to significant errors. Many farms still use the "pushing method" and "bowl method" for fry counting. The pushing method involves estimating the number of fry in a pile manually, whereas the bowl method estimates this number based on a sample. Both methods are subject to human error and often lead to inaccurate counts. Moreover, these manual methods are not only inefficient but can also harm the fry. During counting, the fry are often handled repeatedly, which can negatively impact their growth and survival, causing stress and affecting fry quality. With the advancement of technology and the development of computer systems, automated devices have been gradually introduced into the aquaculture industry. The advent of automatic fish fry counters has effectively addressed the inefficiencies of manual counting while ensuring accurate and transparent counting. These automated devices use sensors, image recognition, and machine learning technologies to automatically detect and track fry, efficiently completing counting with real-time data collection, high accuracy, and reliability. These tools provide aquaculture operators with a more scientific and convenient way of management, allowing for more precise feeding practices and reducing overfeeding or underfeeding, thus improving farm efficiency. However, despite the widespread use of automated counting technology in aquaculture, several challenges remain. In particular, when dealing with large volumes of fry, existing counting technologies face limitations in efficiency, accuracy, and handling fry overlap. As the number of fry increases, counting accuracy tends to decrease, especially when dealing with smaller fry, where detection systems can make errors. Additionally, the complex environment of aquaculture farms, such as light conditions, bubbles, and debris in the water, can interfere with counting accuracy, making the process complicated for automatic counting systems. Therefore, enhancing the accuracy of automated counting, especially in large volumes of fry or in complex environments, is still a technical issue that needs to be addressed. To tackle these issues, researchers have made significant improvements and innovations in automatic counting technology. The accuracy and efficiency of automatic counting systems have been significantly enhanced by incorporating advanced image recognition algorithms, deep learning techniques, and multisensor fusion technologies. These improved algorithms are effective at separating and tracking targets, achieving precise counting. Furthermore, with the development of simulation technology, virtual testing and simulations have played a crucial role in optimizing and designing automatic counting devices. Simulation allows device performance to be predicted under different working conditions, reducing the need for testing with live fry and minimizing potential losses. It also improves design efficiency and ensures that the stability, safety, and durability of the devices are thoroughly validated before practical application, providing reliable technical support for their implementation. Nevertheless, when handling large volumes of fry counting, challenges related to efficiency and dealing with overlapping fry remain. For instance, when designing automatic fry counting devices, design parameters are often difficult to calculate accurately because of the limitations of the fry and the aquaculture environment. Therefore, machine operating parameters must be adjusted to test counting effectiveness. However, this method wastes both human and material resources, and repeated testing can cause stress and harm to the fry. Simulation allows for testing the operational performance of prototype designs, reducing the need for physical testing and saving costs. This study introduces an improved YOLOV8-ByteTrack algorithm to achieve high-precision detection and tracking of fry. This algorithm focuses on real-time performance and efficiency by combining the lightweight YOLOV8 model for object detection with the precise multi-target tracking capabilities of ByteTrack. YOLOV8, being a lightweight model, reduces computational load while maintaining high detection accuracy, offering fast and stable performance in resource-limited environments. Once the fry are detected, ByteTrack uses an efficient data association strategy to track multiple targets and maintain identity consistency, significantly reducing issues such as ID switching and target loss caused by rapid movement, overlap, or environmental changes. Unlike traditional algorithms that rely solely on high-confidence detections, ByteTrack incorporates low-confidence results by utilizing motion consistency and appearance features, thus improving counting accuracy and continuity. To verify the performance and stability of the proposed fish fry automatic counting device in practical applications, a series of accuracy testing experiments was conducted on fish fry of different sizes. The experiments tested 3–5, 6–8, and 9–12 cm-sized grouper fry, with 200, 250, and 300 fry per group, respectively. The test results showed that the counting accuracy of the device was 99.1% and 99.6% for the 6–8 and 9–12 cm fry, respectively, with a slightly lower accuracy of 98.5% for the 3–5 cm fry. The algorithm achieved an average frame rate of 155 FPS, with a single-frame processing time of approximately 6.5 ms. Moreover, the processing speed at different frame rates demonstrated high real-time stability, with a minimum processing speed of 6.3 ms (158 FPS) and a maximum of 6.6 ms (152 FPS). The lower accuracy for the 3–5 cm fry can be attributed to their smaller size, which makes them more susceptible to background complexity and rapid movement, leading to a slight decrease in detection precision. As fry size increases, their features become more distinct, and their movements become more stable, resulting in higher detection accuracy. These results validate the excellent real-time and efficient performance of the algorithm even with limited hardware resources, meeting the practical needs of speed and accuracy in fish fry counting scenarios. The fish fry automatic counting device proposed in this study offers an innovative solution to improve fry counting precision and efficiency, providing valuable insights for theoretical research and practical applications in aquaculture.
