Trademark retrieval is a frequently used task in intellectual property protection. Utilizing efficient trademark retrieval methods can improve retrieval efficiency, reduce manual review costs, and effectively prevent trademark infringement. While the diversity and complexity of trademark, as well as the scarcity of labeled data, challenge existing retrieval methods, we propose a revised trademark retrieval system based on self-supervised learning. Our method revises MoCoV2 self-supervised learning framework by introducing a hard sample selection strategy to enhance the model’s performance and its capacity of feature representation. We also integrate attention mechanisms and multi-stage feature fusion to improve the model’s ability to capture significant visual elements in trademark and multi-scale features. We conducted evaluation and comparison experiments on the METU dataset. The experimental results indicate that our method achieves better performance on the metrics of NAR (Normalized Average Rank) and MAP@ 100 (Mean Average Precision at 100) compared to the state-of-the-art method, proving the effectiveness of the proposed trademark retrieval method.