A Comparative Study of Contrastive Learning-Based Few-Shot Unsupervised Algorithms for Efficient Deep Learning

Abstract

The rapid development in computer vision algorithms, particularly in target detection algorithms based on deep neural networks, has led to significant advancements in various fields. However, recent research on target detection algorithms has shifted towards small sample scenarios with long-tailed distribution of categories, where achieving high accuracy target detection with limited data has become a crucial research topic. Deep neural networks often face several challenges when dealing with limited data, including convergence problems, overfitting, and poor generalization performance. In such scenarios, categories with little data are easily overwhelmed by the negative gradients generated by other categories during network training, which affects the final detection results. To address this issue, an unsupervised contrast learning algorithm has been proposed that can achieve accurate results on a small number of reference datasets without requiring a large number of datasets. One recently proposed contrast learning algorithm is PixelCL, which combines information from image depth in pretraining to improve results. The results of this algorithm show that it can obtain more accurate results through contrast learning with limited sample data. These findings highlight the potential of contrast learning algorithms in addressing negative gradient issues and improving the performance of deep neural networks in target detection tasks with limited data.