Frontiers in Marine Science (Oct 2022)
Robust segmentation of underwater fish based on multi-level feature accumulation
Abstract
Because fish are vital to marine ecosystems, monitoring and accurate detection are crucial for assessing the potential for fisheries in these environments. Conventionally, fish-related assessment is conducted manually, which makes it labor-intensive and time-consuming. In addition, the assessments are challenging owing to underwater visibility limitations, which leads to poor detection accuracy. To overcome these problems, we propose two novel architectures for the automatic and high-performance segmentation of fish populations. In this study, the efficient fish segmentation network (EFS-Net) and multi-level feature accumulation-based segmentation network (MFAS-Net) are the base and final networks, respectively. In deep convolutional neural networks, the initial layers usually contain potential spatial information. Therefore, the EFS-Net employs a series of convolution layers in the early stage of the network for optimal feature extraction. To boost segmentation accuracy, the MFAS-Net uses an initial feature refinement and transfer block to refine potential low-level information and subsequently transfers it to the deep stages of the network. Moreover, the MFAS-Net employs multi-level feature accumulation that improves pixel-wise prediction for fish that are indistinct. The proposed networks are evaluated using two publicly available datasets, namely DeepFish and semantic segmentation of underwater imagery (SUIM), both of which contain challenging underwater fish segmentation images. The experimental results reveal that mean intersection-over-unions of 76.42% and 92.0% are attained by the proposed method for the DeepFish and SUIM datasets, respectively; these values are higher than those by the state-of-the-art methods such as A-LCFCN+PM and DPANet. In addition, high segmentation performance is achieved without compromising the computational efficiency of the networks. The MFAS-Net requires only 3.57 million trainable parameters to be fully trained. The proposed model and the complete code will be made available1.
Keywords
