Introduction
Individual feed intake (IFI) measurements in aquaculture is challenging to measure because of the nature of the culture system where the fish are reared in group condition. Measuring IFI can enable several advancements in aquaculture development to reduce feed cost and environmental impact, but mainly it is required in selective breeding to select feed conversion ratio (FCR) as a trait. Indeed, the ability to measure IFI is the current limiting factor in achieving appropriate estimation of breeding values for FCR . Current methods are invasive with X-ray imaging and RFID-tag in feeding devices, or neglect social behaviour with individual tanks. Here we propose a new approach in measuring IFI at group condition using computer vision (CV). CV enables the automated interpretation of complex visual data to derive actionable biological insights. The idea is to have a practical method for measuring IFI in groups, to maximize the fish capabilities in feed conversion with selective breeding.
Method
European seabass (200g average) were stocked in 240L aquaria at densities of 10, 15, and 20 individuals to facilitate CV development. Over 5 days, fish were filmed daily for one hour during ad libitum feeding, using a custom distributor that placed pellets near the front glass to optimize video recording of feeding events. Initial and final biometry were recorded, and individual fish were filmed separately beforehand to create a gallery set for re-identification (Re-ID) comparison. To detect feeding events, two YOLOv8 models were developed using 300 randomly extracted frames each, for "PELLET" and "OPEN MOUTH" occurrences. The system identifies a feeding event when a tracked pellet disappears and a proximity search simultaneously detects an "OPEN MOUTH" bounding box. This dual-condition logic ensures accurate consumption tracking within the video data. Individual feeding contributions were quantified using two distinct Re-ID approaches derived from the FISHNESS project dataset (Crestel et al., 2026). The first utilized Vision Transformers (ViTs) trained on over 18090 (divided to 80/20 train and test set) full-body images of European seabass representing 4944 individual fish protographed at 2 to 5 different ages. The second employed a Convolutional Neural Network (CNN) focused specifically on the head region, defined from the snout to the opercula. These models allow the CV system to attribute specific feeding events to individual fish within the experimental groups. In parallel several supplementary models are also built in order to integrate the feeding event with the Re-ID model. We built the "EYE", "FISH" and "HEAD" detection models are built with only 100 frames each with YOLOv8 architecture. The "EYE" detection are used in order to track the fish and assign the Re-ID model identification. The "FISH" model are used to generate crop of the whole fish body and send it to the Re-ID model for identification. The "HEAD" model is similar but crops the head part only. Finally, the whole pipeline works by tracking the eyes of the fish in the aquarium. Where there are head or the whole fish detection by the "HEAD" and "FISH" models within the tracked eye, the bounding boxes are used to crop the frame and send it to the Re-ID model for identification. The IDs are then assigned back to the tracked eye. Then, when there are feeding events detected by the logic explained before, the feeding events are attributed to the closest eye that is detected, tracked and identified. Validation is done manually by carefully evaluating the feed intake output video. Evaluation metrics are based on F1 score as a standard metric. F1 score are calculated as with and . TP means true positive (feed intake happened and detected), FP means false positive (there are no feed intake but it was detected) and FN means false negative (there are feed intake but it is not detected). For the Re-ID, rank-1 (the percentage of correct prediction of the 1st rank) are used for evaluation.
Results and Discussion
All of the detection models were trained successfully with F1 score of >0.8. The head Re-ID model with CNN training resulted in rank-1 score of 100% highlighting a probability of overfitting, that requires refinement with more training dataset or parameters tuning. Meanwhile, the whole fish Re-ID model training with ViTs resulted in rank-1 score of 95.7%. This score is also high but not as perfect as the head CNNs Re-ID model. Both model are tested on images from the 80/20 data split. We continue to build the whole pipeline taking into account the fact that that the modular nature of the pipeline allows for the seamless integration of refined Re-ID architectures as larger, more diverse datasets become available. The whole pipeline was then applied with the logic explained in the Method section. The preliminary results shows that in Aquarium 1 (with 10 fish), 470 minutes of videos were analyzed, 619 feed intake events detected with 85% attributed to an ID. Fish with ID 711 had the highest feed intake with 186 (30% of the total) pellets eaten while fish with ID 622 had the lowest feed intake with only 2 (0.3%) pellets eaten. Other videos are being processed and analyzed. Preliminary validation of the feed intake resulted in F1 score of 0.8 proving that the pipeline is working. The current analyses shows promising results for this pipeline to be used for IFI measurement in group condition of European seabass. Several limitations such as lack of objective validation method for some modules, model overfitting problems and video processing time are areas that need further development. However, this result acts as a proof of concept that measuring IFI with computer vision is possible and with future developments, it can be a reliable method. Several indirect measurements which have been hypothesized to correlate with IFI (e.g., weight loss at fasting, growth rate, lipid content) can also be proven by this method. Overall, this is the first step in creating non-invasive, high-throughput, group conditioned IFI measurement with CV.
Acknowledgment
This work is part of the SEPIAA project (funded by France 2030 and BPI), with experimental support from the Ifremer PEMP station in Palavas-les-Flots, France.
References
Crestel, D., Vergnet, A., Delpuech, E. et al. Quantitative genetics and GWAS reveal population-specific sex-ratio responses in wild European seabass (Dicentrarchus labrax) under various temperature scenarios. Heredity (2026). https://doi.org/10.1038/s41437-026-00841-w.