<p>Trademarks are an indispensable element of intellectual property and play a critical role in the progression of social and economic development. Nonetheless, with the exponential rise in their quantities, the accurate identification of trademarks to prevent infringement has grown increasingly arduous. In response to the limitations inherent in current trademark retrieval approaches, including the challenge of capturing global context information during feature extraction and the inadequate accuracy in detailing feature extraction, this paper introduces a novel trademark retrieval model termed PAFI-Net, which stands for Parameter-Free Attention Convolutional Self-Attention and Feature View Integration. Initially, a self-attention mechanism is employed in conjunction with a parameter-free attention convolutional block to enhance the ability to grasp global dependency relationships. Subsequently, utilizing multiple branches to learn the fine-grained details of feature maps mitigates the impacts of individual limitations, thereby augmenting the overall learning capacity and performance of the model. Ultimately, experimental outcomes from two trademark datasets, namely Logo-2K+ and Logo-627, reveal that PAFI-Net demonstrates outstanding performance in the task of trademark retrieval, thus validating the efficacy of the proposed model.</p>