AASNet: A Novel Image Instance Segmentation Framework for Fine-Grained Fish Recognition via Linear Correlation Attention and Dynamic Adaptive Focal Loss

Smart fisheries, integrating advanced technologies such as the Internet of Things (IoT), artificial intelligence (AI), and image processing, are pivotal in enhancing aquaculture efficiency, sustainability, and resource management by enabling real-time environmental monitoring, precision feeding, and...

Full description

Saved in:
Bibliographic Details
Main Authors: Jianlei Kong, Shunong Tang, Jiameng Feng, Lipo Mo, Xuebo Jin
Format: Article
Language:English
Published: MDPI AG 2025-04-01
Series:Applied Sciences
Subjects:
Online Access:https://www.mdpi.com/2076-3417/15/7/3986
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Smart fisheries, integrating advanced technologies such as the Internet of Things (IoT), artificial intelligence (AI), and image processing, are pivotal in enhancing aquaculture efficiency, sustainability, and resource management by enabling real-time environmental monitoring, precision feeding, and disease prevention. However, underwater fish recognition faces challenges in complex aquatic environments, which hinder accurate detection and behavioral analysis. To address these issues, we propose a novel image instance segmentation framework based on a deep learning neural network, defined as the AASNet (Agricultural Aqua Segmentation Network). In order to improve the accuracy and real-time availability of fine-grained fish recognition, we introduce a Linear Correlation Attention (LCA) mechanism, which uses Pearson correlation coefficients to capture linear correlations between features. This helps resolve inconsistencies caused by lighting changes and color variations, significantly improving the extraction of semantic information for similar objects. Additionally, Dynamic Adaptive Focal Loss (DAFL) is designed to improve classification under extreme data imbalance conditions. Abundant experiments on two underwater datasets demonstrated that the proposed AASNet obtains an optimal balance between segmentation performance and efficiency. Concretely, AASNet achieves mAP scores of 31.7 and 47.4, respectively, on the UIIS and USIS dataset, significantly outperforming existing state-of-the-art methods. Moreover, AASNet achieves an inference image recognition speed of up to 28.9 ms/per, which is suitable for practical agricultural applications of smart fish farming.
ISSN:2076-3417