Shrimp Detection Method Based on Dense Unbiased Teacher

With the development of computer vision, agricultural technology is also moving towards intelligence. Deep learning-based object detection algorithms have gained widespread adoption in the shrimp farming industry, providing superior accuracy and efficiency compared to traditional machine learning algorithms. However, a large amount of labeled data are required in previous detector training procedures. In some intensive scenarios, labeling a single image can even take a few minutes, while model training requires hundreds or thousands of labeled data to obtain good performance. Therefore, we propose the Dense Unbiased Teacher (DUT) algorithm for shrimp detection in this work, an semi-supervised object detection (SS-OD) pipeline based on dense pseudo-labels, which can learn rich semantic information from unlabeled data more efficiently. Compared to the original Unbiased Teacher (UT) algorithm, our method can more efficiently utilize the supervision of the teacher model. We utilized the high-definition camera to capture shrimp image data and used this dataset to establish a Shrimp-SSOD benchmark for evaluating the performance of our model. The experimental results show that our method outperforms the original UT in all Shrimp-SSOD settings. Specifically, DUT achieves 2.21 mAP improvements compared to UT and 9.21 mAP improvements compared to the supervised baseline when using 50% of labeled data. It is worth noting that our method achieves comparable performance to fully supervised learning despite using only 50% labeled data, which means we can save half of the labeling time.