An integrated end-to-end deep neural network for automated detection of discarded fish species and their weight estimation

Sustainable management of aquatic resources requires efficient acquisition and processing of vast amounts of information to check the compliance of fishing activities with the regulations. Recent implementation of the European Common Fisheries Policy Landing Obligation implies the declaration of all listed species and sizes at the harbour. To comply with such regulation, fishers need to collect and store all discards onboard the vessel, which results in additional processing time, labour demands, and costs. In this study, we presented a system that allowed image-based documentation of discards on the conveyor belt. We presented a novel integrated end-to-end simultaneous detection and weight prediction pipeline based on the state-of-the-art deep convolutional neural network. The performance of the network was evaluated per species and under different occlusion levels. The resulting model was able to detect discards with a macro F1-score of $94.10\%$ and a weighted F1-score of $93.88\%$. Weight of the fish could be predicted with mean absolute error, mean absolute percentage error, and root squared error of 29.74 (g), $23.78\%$, and 44.69 (g), respectively. Additionally, we presented a new dataset containing images of fish, which, unlike common object detection datasets, also contains weight measurements and occlusion level per individual fish.

Automated summary

Efficient acquisition and processing of vast amounts of information are essential for sustainable management of aquatic resources and compliance monitoring of fishing activities. This study presented a system that utilized image-based documentation to detect and predict the weight of discards on the conveyor belt. The proposed integrated pipeline achieved high detection and weight prediction accuracy, with macro F1-score of 94.10% and weighted F1-score of 93.88%, respectively. The new dataset provided in this study contains fish images, weight measurements, and occlusion levels, which is valuable for further research in object detection.