In general, the detection algorithm demonstrated good efficiency in recognizing and counting pirapitinga fingrlings, as observed through the [email protected] values and recall above 95%, which validates the correct predictability of the tested convolutional network. The resulting values were superior to the work of Li and Jiang (2023), that found mAP values of 95.4% and recall of 88% for fish detection in various cultivation systems. However, the tested network faced difficulties in detecting smaller fish.
Although this study did not focus on evaluating the algorithm's performance in complex environmental situations, such as underwater environments, other limitations were noted. Obstructions and occlusions of fingerlings, low sharpness from digital images with higher fish density, and variation in the number of fry in the dataset were examples of such limitations. All these factors are challenging in network learning and can interfere with accuracy and other metrics (Kuswantori et al., 2023; Barreiros et al., 2021; Babu et al., 2023). The evaluated model partially overcame all these obstacles.
Regarding the evaluated batch normalization technique, it was found that the precision, recall, and mAP values of the model trained without these parameters were lower than those with batch normalization. Nevertheless, the difference between models with different batches were not significant. When the batch size is larger, the detection algorithm tends to perform better in the proposed task and consequently increase metrics such as precision. However, higher batch numbers end up increasing computational costs (Sung & Girdhar, 2017; Knausgard et al., 2022).
A batch size of 10 achieved a precision of 94.23%, and with batch size of 20, the value found was 96.28% (Iqbal et al., 2021) for fish detection and classification using an improved CNN. These values were very similar to those found in this investigation, confirming that as the batch size increases, precision will also increase. In studies focused on the recognition and counting of fingerlings by detection algorithms, small batch size values up to 20 achieve excellent efficiency and performance without significant computational costs (Lin et al., 2022; Radiuk, 2017).
Good quality datasets are crucial for achieving excellent results in CNN learning. When creating them, the labelling work is time-consuming and arduous (Krell & Kim, 2017; Zhang et al., 2024). In this study, the responsible team monitored meticulously to prevent faulty labeling. The formed dataset presented some variations to increase the robustness of the trained models. In total, a thousand images with varying fingerling density between 10 to 50 individuals per tank were used to form the training bank. It can be noticed in the images that the water was slightly turbid, and no lighting system was used to facilitate visualization. Evaluating neural networks, Fernandes et al. (2024) used a lighting system to facilitate visualization, obtaining metrics above 99% with a much smaller number of images.
Robust data is essential to make trained models efficient and practical in real life situations (Lyu et al., 2022). Although the data collected in the investigation were from a controlled environment, the adverse variations presented were sufficient for the robustness of the trained networks (Seo et al., 2020). Thus, the experimental conditions and the models trained for counting fingerlings of pirapitinga through images can be applied in retail businesses as a replacement for the manual methods usually used, such as sampling through sieves. Establishments where fingerlings are commercialized usually count the animals first and then distribute them in packages for transport. In these establishments they use less turbid water with a better lighting system to facilitate the fingerling visualization.