WWW.WAS.ORG • WORLD AQUACULTURE • SEPTEMBER 2025 39 (CONTINUED ON PAGE 40) The next problem lies with the fish and the model. The fish can swim in a fuzzy manner, which makes it impossible to track or predict the pattern of swimming. Also, the model cannot keep a fixed tag on the fish as there will be sudden movements and fish cross-over, which makes this measurement process further difficult. However, keeping track of a fish that is closer to the marker for a specific number of frames might be easy as we need the movement for minimal frames in a video to determine the approximate swimming speed of a fish. The final constraint lies in the dimension of a tank; the tank is a three-dimensional space, while the model assumes that the videos are in 2 dimensions. The measurements of the same fish that swim closer to the marker and further from the marker may vary because of the depth/ height factor of the tank. This is a difficult task to overcome. There are two ways we can eliminate this, one being taking measurements only when the object swims closer to the marker or using two markers on both sides, which have the same dimension and have a known distance from each other. Since the fish moves in 3D space, we use a pinhole camera model to measure approximate depth. Then, the fish length can be calculated from the formula, Length (3D) in cm= ( Lxy 2 + Z2)1/2, where Lxy is the length obtained from measurement on the 2D plane using an ArUco marker, and Z is the estimated depth of the tank (Lepetit et al. 2009). Models that use video inputs taken from mobile phones may not be as accurate as a model that uses high-end cameras and a high processing system, although I believe this will be a stepping stone for AI integration without the need for high-priced equipment, making AI available to everyone. The Double-edged Dagger We cannot rely entirely on AI, as an accuracy of 100% is impossible to achieve in theoretical terms. Yet, with future advancements and the development of more detailed models, it might be possible shortly. In this case, the Zid-AI system shows significant progress in non-invasive zebrafish monitoring, achieving reliable fish detection. However, visual occlusion and lens distortion pose challenges. Future work should focus on refining length measurement accuracy and expanding the dataset for more robust results. Testing in diverse environments and integrating real-time monitoring will further enhance its utility in zebrafish research. The accuracy of the model and speed of analysis contradict each other; we can either sacrifice accuracy by a small margin for rapid detection or vice versa. Yet, implementing the latest models, which have lower parameters, can give an improved accuracy even under rapid detection. Camera malfunctions and model overfit are also areas where we need to be cautious. Training the model with a lot of images or data with very high learning cycles (epochs) can lead to model over-fitting, which may affect performance. We cannot use these models in some field that requires much more precision, such as molecular TABLE 1. Statistical analysis of the correlation between different groups measured by AI and actual measurements. Group p_value Correlation Small 0.89452 0.80593 Medium 0.12136 0.95685 Large 0.090876 0.95034 TABLE 2. The measurements taken by the Zid-AI on zebrafish tanks. * Indicates the average measurements of triplicates of the same group. Whereas S, M, and L denote small, maturing, and matured fishes. AI A* B* C* D* E* S 2.17 2.13 2.1 2.1 2.4 M 3 2.8 3.1 2.7 2.8 L 4.33 4.63 4.233333 4.36666667 4.36666667 TABLE 3. The measurement taken by digital photography on zebrafish tanks. * Indicates the average measurements of triplicates of the same group. Whereas S, M, and L denote small, maturing, and matured fishes. Or A* B* C* D* E* S 2.17 2.1 2.233333 1.966667 2.4 M 3 2.82 3.033333 2.6 2.7 L 4.13 4.63 4.1 4.366667 4.266667
RkJQdWJsaXNoZXIy MjExNDY=