.png){kind=icon}
.png){kind=icon}
Return
A Scale-Aware local Context aggregation network for Multi-Domain shrimp counting
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2024.126179.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Shrimp counting is crucial in industrial aquaculture for controlling feed, adjusting density, and evaluating economic performance. However, existing shrimp counting studies are still in the early stages and have not considered the practical domains in intensive farming, such as bubbles, excreta, and yellow-green turbid backgrounds. Moreover, the challenge of scale variation further complicates accurate counting. It is essential to study a model that performs well in multiple domains. We propose a Scale-aware Local Context Aggregation Network (SLCA-Net) to address scale variation, domain bias, and noise interference, and we publicly release a multi-domain shrimp dataset. SLCA-Net consists of a Base Extractor, a Lightweight Local Scale Perception (LLSP) module, and an Attention-Guided Aggregation (AGA) module. The LLSP module captures rich scale information, utilizing a lightweight, multi-branch, multi-level fusion architecture. Meanwhile, the designed Local Perceptual Attention (LPA) mechanism enables fine-grained perception, highlighting domain-invariant information of shrimp while mitigating shallow noise interference. The AGA module enhances the model's learning capacity through a strategy of intra-group cooperation and inter-group competition to intelligently select key information. For network optimization, we develop a Ring Counting (RC) loss that emphasizes local information, addressing the shortcoming of the counting loss ignoring spatial distribution. Experiments show that SLCA-Net achieves a Mean Absolute Error (MAE) of 2.377 and a Root Mean Square Error (RMSE) of 3.003, reducing these errors by 16.6 % and 17.2 % respectively compared to state-of-the-art methods. Meanwhile, our method has a smaller size (9.76 M), indicating its potential feasibility for practical applications.
Keywords:
Density map estimation
Scale-aware network
Attention
Counting loss
Scale aggregation
Shrimp counting
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W
Organization
No organization information available