BEGV2-UNet: A method for automatic segmentation and calculation of backfat and eye muscle region in pigs

Abstract

Rapid and accurate measurements of eye muscle area and backfat thickness in breeding pigs is crucial for improving breeding traits. Within reasonable ranges, these traits significantly influence the number of piglets born, their birth weights, and survival rates. Traditional detection methods are time-consuming and heavily reliant on operational expertise. While B-mode ultrasound is widely used as a non-invasive tool for measuring backfat thickness and eye muscle area, its efficiency and precision are limited by dependence on the operator.

To address these issues, this study introduces the BEGV2-UNet model, an innovative UNet network based on reconstructing down-sampling and up-sampling paths, incorporating GhostModuleV2, and incorporating a large kernel attention mechanism to better capture the boundaries and positions of backfat and eye muscle regions. The model can be used to segment these regions in breeding pigs and improve the loss function for accelerate convergence while remedying the low precision caused by class imbalance. Using a dataset of ultrasound images, the BEGV2-UNet model achieved an MIoU of 96.18 % and MPA of 98.12 %, with model size reduced to 18.69 MB and strong inference accuracy. We calculated the backfat thickness and eye muscle area using the model to achieve R² values of 0.98 and 0.96, respectively.

This study highlights the significant advantages of BEGV2-UNet in terms of image segmentation accuracy and lightweight design.