一种基于深度学习的商标检测方法

A Trademark Detection Method Based on Deep Learning

网购时代,五花八门的产品品牌让消费者选择困难,同时也滋生了大量线上商标侵权行为。为此,提出一种基于深度学习的商标检测方法。方法以YOLOv7-tiny网络模型为基础,首先,为提升多尺度检测能力,Neck部分的PAnet模块改为简化的自适应学习权重、多尺度特征融合网络SimBiFPN;接着,为关注关键语义信息,引入注意力机制,将Neck与Head间的卷积层Conv改为全维度动态卷积ODConv;最后,为了使算法具有像素级建模能力,将激活函数改为FRelu。选择天池平台数据集经Mosaic和Mixup数据增强后完成模型的训练与验证。结果表明:改进模型的mAP达到85.84%,较原始模型提升了近2个百分点,优于其他YOLO(you only look once)模型,且模型的参数量下降41%。所提方法有助于提高用户的在线购物效率,同时可加强商标侵权的打击力度。

In the era of online shopping,a variety of product brands make it difficult for consumers to choose,and also breed a large number of online trademark infringement.Therefore,a trademark detection method based on deep learning was proposed.Method was based on YOLOv7-tiny,first of all,in order to improve the multi-scale detection ability,PAnet module in Neck part was changed to a simplified adaptive learning weights,multi-scale feature fusion network SimBiFPN.Then,in order to focus on the key semantic information,the attention mechanism was introduced to change the Conv convolution layer between Neck and Head into full-dimensional dynamic convolution ODConv.Finally,in order to make the algorithm get pixel-level modeling capability,the activation function was changed to FRelu.The dataset of Tianchi platform was selected and then enhanced by Mosaic and Mixup to complete the training and verification of the model.The results show that the mAP of the improved model is raised to 85.84%,nearly 2 percentage points higher than that of the original model,better than other you only look once(YOLO)models,and the number of parameters of the model is reduced by 41%.The proposed method can be used to improve the efficiency of online shopping and strengthen the fight against trademark infringement.