Semi-Supervised Fish Counting Network Based on Generating Adversarial Learning

Automatic fish counting is essential for fishery management. However, the existing density map regression-based fish counting methods rely on large-scale labeled training data to assure counting accuracy. The labeling process of dense fish in images is time-consuming and laborious. To address the above issues, a semi-supervised fish counting network based on generating adversarial learning is proposed in this study. To fully utilize the information in unlabeled images, a pixel discriminator is designed to align the prediction of labeled and unlabeled images from pixel perspectives. The squeeze excitation mechanism (SE) is embedded into the discriminator to enhance its ability of discriminative feature extraction by modeling the channel dependency of the features. Comparison experiments demonstrate that the proposed semi-supervised counting method effectively improves the fish counting performance by fully utilizing the unlabeled data. The proposed method achieves superior fish counting performance compared to the state-of-the-art counting methods.