Impact evaluation of deep learning on image segmentation for automatic bluefin tuna sizing

• DOI: 10.1016/j.aquaeng.2022.102299
• Publication Date: 2022-10-21T15:43:27Z
• Authors: P. Muñoz-Benavent; J. Martínez-Peiró; G. Andreu-García; V. Puig-Pons; V. Espinosa; I. Pérez-Arjona; F. De la Gándara; A. Ortega
• Keywords: Underwater stereo vision; Fishery management; Automatic fish sizing; Biomass estimation; 14.- Conservar y utilizar de forma sostenible los océanos, mares y recursos marinos para lograr el desarrollo sostenible; ARQUITECTURA Y TECNOLOGIA DE COMPUTADORES; 03 medical and health sciences; 0302 clinical medicine; FISICA APLICADA; Computer vision; Convolutional neural networks; 14. Life underwater

This paper evaluates the impact of using deep learning techniques in an automatic fish sizing process. Automatic with a non-invasive approach involves working different views fish's body and changing environments, being stage extraction individuals image quality segmentation essential to obtain good measurements. The goal this work is improve results functionality achieved our previous studies conventional methods based on local thresholding, where limitations were observed, mainly necessity parameters tuning high computational cost. number detections must also increase significantly reliability statistical results. An convolutional neural networks proposed for detection videos acquired under real conditions, which eliminates engineering procedure parameter adjustment generalises solution deal environmental conditions (illumination water turbidity) background variability. show that enhanced thanks improvement instance segmentation. In particular, measurements increases by up 2.45 times when Mask R-CNN PointRend module, thus increasing accuracy length estimation, per minute computing time 3.5 times. Our proposal obtains highly accurate estimations juvenile bluefin tuna stereoscopic vision system deformable model silhouette, both from ventral dorsal perspectives. important applying CNN, as demonstrated segmented instances, required segment instance, achieved.