Line-labelling enhanced CNNs for transparent juvenile fish crowd counting

Abstract

Counting juvenile fish in aquaculture is challenging due to their small, fragile, and often transparent bodies, especially under high-density conditions. To address this, we propose a novel line-labeling annotation method specifically designed for transparent juvenile fish counting, which enhances supervision quality and provides both positional and morphological cues. We also introduce an improved CSRNet-based convolutional neural network, optimized for high-density fish scenarios. A dataset of 9000 annotated images of Silver Carp and Tilapia, categorized into four density ranges (0–10, 10–20, 20–30 and 30–40 fish/cm²), was used to train and evaluate our method. To determine the optimal approach, four combinations of labeling and image enhancement methods were tested: Point Labeling + Original Image (P + O), Line Labeling + Original Image (L + O), Point Labeling + Image Enhancement (P + I) and Line Labeling + Image Enhancement (L + I). Counting accuracy was assessed using heatmap-based visualizations. Experimental results demonstrate that the line-labeling method significantly improves counting accuracy, achieving 97.73 % for Silver Carp and 98.04 % for Tilapia, outperforming conventional point-based annotations in high-density contexts. This study highlights the potential of structured annotations and tailored network designs for advancing precision in fish counting tasks.