As one of the computer vision applications in ocean exploration, marine animal segmentation (MAS) plays an important role in understanding species distribution, behavior, and interactions. However, the poor image quality resulting from underwater light absorption, scattering, and the camouflage behavior of marine animals pose challenges to MAS. Existing segmentation methods either perform poorly in underwater environments or have excessive computational demands for underwater systems, which is not conducive to the development of MAS. This paper proposes a foreground transformation contrast network (FTCNet) to address the challenge of degraded underwater image quality and biological camouflage. Specifically, we first design a foreground transformation contrast learning strategy (FTCLS) to extract shared semantic features between the original image and foreground-transformed images, thereby mitigating the impact of poor image quality and animal camouflage with color-invariant information. Additionally, a progressive multi-scale feature fusion module (PMFFM) is devised to fuse features across different layers progressively to address the scale variations caused by animal sizes and capture distances. Experimental results on the Real-World Marine Animal Segmentation (RMAS) dataset demonstrate that FTCNet outperforms existing MAS methods with mIoU of 0.75.