Contour segmentation of fish body with neural network model and characteristic parameter analysis of zebrafish swimming

Abstract

Research on the high maneuverability of fish swimming primarily involves addressing the batch processing of large experimental data, specifically how to simultaneously capture and rapidly process deformation-displacement information of fish bodies and related flow fields. The primary objective of this study is to integrate high-speed photography technology with deep learning methods to propose a set of data processing methods suitable for extracting fish swimming characteristic parameters. For the rapid movements of zebrafish (millisecond-level motion), this study utilized a high-speed camera for image acquisition, obtaining batches of swimming fish images and fluorescence particle information in the flow field. The geometric reconstruction of zebrafish under high-speed swimming was achieved by introducing deep learning algorithms and refining the U-Net model. To tackle the challenges of complex fish swimming scenes, we utilized a novel residual connection approach (path modification) and constructed a hybrid function model (module enhancement), resulting in a new neural network model tailored for zebrafish swimming image processing: Mod-UNet. Through testing, the improved Mod-UNet model effectively eliminated interference from fluorescence particles in the flow field on the extraction of fish body contours, achieving an overall IoU coefficient of 93%. The processing results demonstrated a kind of consistency compared to results obtained with traditional methods by previous researchers. By calculating the geometric morphology of zebrafish, we further derived the kinematic characteristics of zebrafish. Simultaneously, by applying cross-correlation algorithms to calculate the positions of fluorescence particles, the velocity characteristics of the flow field were obtained. The λ_ci method and the Ω method were used to identify vortex structures, providing the evolution patterns of corresponding flow field characteristic parameters. The experimental data processing method proposed in this paper provides technical support for establishing a zebrafish swimming information database.