Abstract: In beauty product retrieval, there exists much interference information in the product images, such as complicated backgrounds, camera shooting angles and product poses. Meanwhile, the retrieval dataset often contains large images with very similar products within them. To deal with these factors, introducing salient attention mechanism to loss, dropout and the similarity calculation, the novel SA Loss, SA Dropout and SA Similarity are designed which can effectively reduce the interference information. Then, the effective feature extractor is chosen to resist similar interference. Finally, combining these four modules, a salient attention based beauty product retrieval neural network (SA-Net) is proposed to enhance the retrieval accuracy. SA-Net is trained on Pro-10k dataset. The comparative experiment and ablation learning on Per-500k dataset show that the retrieval accuracy of SA-Net is 3% higher than that of the best state-of-the-art.