Deconvolution Enhancement Keypoint Network for Efficient Fish Fry Counting.

Area of Science:

• Computer Vision
• Aquaculture Technology
• Machine Learning

Background:

• Accurate fish fry counting is crucial for fish farming but challenging due to occlusion, density, and small size.
• Existing computer-based methods struggle with efficiency and accuracy in large-scale fry enumeration.

Purpose of the Study:

• To develop an accurate and efficient method for fish fry counting using a novel single-keypoint approach.
• To introduce the deconvolution enhancement keypoint network (DEKNet) for improved fish fry enumeration.

Main Methods:

• DEKNet employs a fish fry feature extractor (FFE) with parallel dual branches for high-resolution representation.
• Two identical deconvolution modules (TDMs) generate a high-resolution keypoint heatmap.
• Fish fry are identified by local peaks in the heatmap, enabling precise counting and localization.

Main Results:

• DEKNet achieved 98.59% accuracy in fish fry counting on the FishFry-2023 dataset.
• The method demonstrated high accuracy (98.51%) on the Penaeus dataset and low MAE (13.32) on Adipocyte Cells.
• DEKNet shows superior performance in accuracy, parameter count, and computational effort compared to existing methods.

Conclusions:

• DEKNet offers a robust and effective solution for accurate fish fry counting in challenging conditions.
• The single-keypoint approach and deconvolution enhancement significantly improve counting precision.
• This research advances automated counting techniques in aquaculture and related biological applications.