关键词: Logo检测, Yolov5网络模型, 注意力机制, 小目标检测, 深度学习

Abstract: In recent years, logo detection and recognition have witnessed continuous innovation and rapid development in the field of computer science. However, challenges such as complex image backgrounds and varying sizes of logo targets still remain in the domain of logo detection. To address these issues, a modified network model based on Yolov5 is proposed. By integrating the CBAM attention mechanism, compression is applied to both the image channels and spatial directions, enabling the extraction of crucial information and important regions within the image. The use of SAC allows the network to adaptively adjust the receptive field size in feature maps at different scales, facilitating the detection of objects across multiple scales and enhancing the network's effectiveness. The NWD metric is embedded into the loss function to reduce the sensitivity of the network to small target sizes, thereby improving model robustness and stability. Experimental evaluations on the publicly available FlickrLogos-32 and QMULOpenLogo datasets demonstrate promising results. In the smaller FlickrLogos-32 dataset, the proposed model achieved a 1% im-provement in mAP compared to the baseline Yolov5. On the larger QMULOpenLogo dataset, the mAP reached 62.7%, showing a 2.3% enhancement. The detection performance surpasses that of both the original network model and classical object detection models.

Key words: logo detection, yolov5 network model, attention module, small object detection, deep learning