Trademark Detection Based on Improved SSD Algorithm

Abstract: Trademarks are found everywhere in daily life. With the development of deep learning in the field of image recognition, trademark detection has become a hot research topic. Although trademark detection has made great progress, there are still some issues that need to be addressed. This paper proposes a trademark detection algorithm based on improved SSD model to address the challenge of low precision in detecting small-sized trademark images. In order to address the issue of excessive parameters in the original SSD network, the lightweight network Darknet is adopted to replace the feature extraction network in SSD. This reduces the number of network parameters and enhances the speed of network detection. To tackle the challenging detection of small objects, an improved FPN network is employed. It progressively fuses multi-scale feature maps from the deep to shallow layers of the network in SSD, aiming to enrich information across all prediction layers to maximize the detection accuracy of small objects. As for the mismatch between the prior box mechanism of the original SSD and the detection of small object datasets, the K-means++ algorithm is utilized to determine the sizes of anchor boxes and obtain appropriate aspect ratios for the current dataset. This enhances both the efficiency and accuracy of the model's detection capability. Experimental results show that compared with the original SSD algorithm, the improved SSD detection algorithm in this paper has better detection accuracy, and the average detection average is increased by 9.8% which proves the effectiveness of the proposed model.