HIDANet: a lightweight deep learning framework for Vannamei post-larval stage classification and morphometric...

Area of Science:

• Aquaculture
• Machine Learning
• Image Analysis

Background:

• Hatchery production of Pacific white shrimp (Litopenaeus vannamei) requires accurate post-larvae (PL) developmental staging for quality control.
• Current manual visual evaluation methods are subjective, leading to observer bias and inconsistency.

Purpose of the Study:

• To introduce a novel, lightweight convolutional neural network, the Hierarchical Isotropic Dense Attention Network (HIDANet), for automated developmental staging of L. vannamei PL.
• To develop an end-to-end workflow for simultaneous stage classification and population assessment (counting, morphometrics, density).

Main Methods:

• Developed HIDANet, a 0.033M parameter CNN utilizing multibranch isotropic depthwise separable convolution and CBAM for feature recalibration.
• Employed aggressive data augmentation to mitigate background color bias and address class imbalance.
• Integrated CLAHE image enhancement, Gaussian adaptive thresholding, and skeleton-based morphometric filtering for automated population analysis.

Main Results:

• HIDANet achieved 98.44% test accuracy with color images and 96.89% with grayscale images, with a macro-averaged F1-score of 0.99.
• The model demonstrated strong morphological learning independent of chromatic background signals.
• The integrated workflow enabled automated larva counting, morphometric measurements, and density categorization.

Conclusions:

• The HIDANet framework provides a scalable and low-cost solution for real-time quality monitoring in commercial aquaculture.
• Simultaneous classification of PL stages and larval population assessment enhances efficiency and objectivity in hatchery operations.