基于改进Mosaic数据增强和特征融合的Logo检测

近年来,Logo检测在知识产权保护和产品品牌管理等领域得到了广泛应用。针对Logo检测中的复杂背景和多尺度问题,提出了一种改进Mosaic数据增强和特征融合的Logo检测算法。将六张原始图片随机翻转、缩放和拼接构成合成图像,与单张图像和由四张原始图片合成的图像一起作为YOLOv4模型的训练输入,并确定三种输入形式的最优比例,同时使用一种新的训练策略,改进的Mosaic数据增强方法丰富了Logo对象的尺度和背景,使模型更好地学习全局和局部特征；在路径整合网络(PANet)的基础上引入跨层连接、重复堆叠、直接连接和加权特征融合等操作,改进的PANet扩大了模型感受野,增强了模型的多尺度特征表达能力。实验结果表明,提出的MP-YOLOv4算法在减小21.7%模型大小的同时, IoU(Intersection of Union)等于0.5时的平均精度上达到了67.4%,较YOLOv4提高了2.4%,同时在多尺度目标上的检测性能得到了改善。

Logo Detection Based on Improved Mosaic Data Enhancement and Feature Fusion

Logo detection has been widely used in intellectual property protection and product brand management in recent years. Aiming at the complex background and multi-scale problems in Logo detection, a Logo detection algorithm based on improved Mosaic data enhancement and feature fusion was proposed. Six original images were randomly flipped, scaled and combined to form a composite image, which was used as the training input of YOLOv4 model together with single image and a composite of four original images, and the optimal proportion of the three input was determined. Meanwhile, a new training strategy was used. The improved Mosaic data enhancement further enriched the scale and context of Logo objects, enabling the model to learn the global and local features better. Based on the path integration network (PANet), some operations such as cross-layer connection, repeated stacking, direct connection and weighted feature fusion were introduced. The improved PANet enlarged the receptive field of the model and enhanced the multi-scale feature expression ability of the model. Experimental results show that the proposed MP-YOLOv4 algorithm can reduce the model size by 21.7% and reach 67.4% average precision when IoU(Intersection of Union) equals 0.5, which is 2.4% higher than YOLOv4. At the same time, the detection performance of multi-scale targets is improved.