Abstract: To reduce the manual annotation cost of large-scale image dataset construction and to solve the problem that traditional active learning methods tended to ignore long-tail scenarios, which limited model performance, a three-stage progressive intelligent annotation method for image datasets was constructed. First, in the cold start stage, DINOv2 was introduced for feature pre-extraction, and an adaptive minority cluster boosting algorithm (AMCB) was proposed; then, in the active learning cycle stage, a two-stage hybrid sampling algorithm was designed, which performed diversity coarse screening through MiniBatch K-means and combined prediction uncertainty with local feature density for fine selection, so as to efficiently drive model iteration; finally, after the model performance reached the standard, a confidence-based stratified filtering mechanism was used to perform full automatic annotation on the remaining data. The experimental results on the cotton pests and diseases image dataset showed that compared with the random sampling algorithm, the AMCB + two-stage hybrid sampling algorithm could cover rare classes more evenly and significantly reduced the manual annotation workload while achieving the same detection accuracy.

Key words: active learning, cotton pests and diseases, image dataset, annotation method