Robust segmentation of underwater fish based on multi-level feature accumulation

Because fish are vital to marine ecosystems, monitoring and accurate detection crucial for assessing the potential fisheries in these environments. Conventionally, fish-related assessment is conducted manually, which makes it labor-intensive time-consuming. In addition, assessments challenging owing underwater visibility limitations, leads poor accuracy. To overcome problems, we propose two novel architectures automatic high-performance segmentation of populations. this study, efficient network (EFS-Net) multi-level feature accumulation-based (MFAS-Net) base final networks, respectively. deep convolutional neural initial layers usually contain spatial information. Therefore, EFS-Net employs a series convolution early stage optimal extraction. boost accuracy, MFAS-Net uses an refinement transfer block refine low-level information subsequently transfers stages network. Moreover, accumulation that improves pixel-wise prediction indistinct. The proposed networks evaluated using publicly available datasets, namely DeepFish semantic imagery (SUIM), both images. experimental results reveal mean intersection-over-unions 76.42% 92.0% attained by method SUIM respectively; values higher than those state-of-the-art methods such as A-LCFCN+PM DPANet. high performance achieved without compromising computational efficiency networks. requires only 3.57 million trainable parameters be fully trained. model complete code will made 1 .