Abstract: In aquaculture, precise fish image segmentation is crucial for growth management. However, the intricate underwater environment, plagued by image blurriness and low quality, poses significant challenges to existing segmentation methods, often leading to reduced accuracy and limited generalization capabilities. To address these issues, we propose an underwater fish image segmentation approach based on an improved Segformer model, designated as FT-Segformer (SegFT for brevity). Our methodology meticulously extracts multi-scale features, spanning from fine-grained high resolutions to coarse-grained low resolutions, utilizing a sophisticated four-layered transformer block structure. Within the decoder, a feature pyramid fusion mechanism seamlessly integrates these features, bolstering contextual understanding. Subsequently, transposed convolutions refine the feature maps, restoring their dimensions and amplifying feature learning capabilities. To evaluate the model, we constructed the UAGF (Underwater Aquaculture Goldfish Fishes) dataset, a genuine underwater aquaculture environment dataset featuring ornamental goldfish, and conducted extensive validation experiments thereon. The experimental results demonstrate that SegFT outperforms existing methods across evaluation metrics such as mIoU, mPA, and mRecall, achieving improvements of 1.76%, 0.39%, and 0.19%, respectively. Notably, in terms of mIoU, SegFT surpasses U-Net, PSPNet, HRNet, and Deeplabv3+ by impressive margins of 1.92%, 3.73%, 3.07%, and 3.58%, respectively. This study underscores the remarkable effectiveness and robustness of our proposed method in complex underwater settings, outperforming existing supervised image segmentation techniques in terms of segmentation performance.

Key words: smart aquaculture, underwater fish image segmentation, feature fusion, transpose convolution, deep learning

摘要： 水产养殖管理中，精准分割图像中的鱼类生长管理至关重要，但水下环境复杂，图像质量低，现有分割方法面临精度低、泛化能力弱等挑战。提出了一种改进Segformer模型（FT-Segformer，简称SegFT）的水下鱼类图像分割方法。首先，利用四层transformer block提取输入图像高分辨率到低分辨率的不同尺度特征。在解码器部分，借助特征金字塔融合机制增强上下文感知；然后，利用转置卷积还原特征图维度，进一步提升特征学习的效果；最后，构建了一个用于模型评估的真实水下养殖环境的锦鲤数据集UAGF，并在该数据集上进行相关验证试验。结果显示：该模型在mIoU、mPA和mRecall等评估指标上均优于现有方法，分别提升了1.76%、0.39%和0.19%，在mIoU指标上， SegFT分别超越了U-Net、PSPNet、HRNet、Deeplabv3+模型1.92、3.73、3.07和3.58个百分点。研究表明，所提出的方法在复杂的水下环境下，具有显著的有效性和鲁棒性。分割性能上优于现有的监督图像分割方法。

关键词: 智慧水产养殖, 水下鱼类图像分割, 特征融合, 转置卷积, 深度学习