二次特征融合的YOLO目标检测算法

YOLO Target Detection Algorithm Based on Secondary Feature Fusion

为解决YOLO目标检测算法精确度低的问题,提出一种二次特征融合的YOLO目标检测算法。使用改进的可分离卷积模块来对传统的卷积模块进行改进,从而减小网络的模型,保证目标检测的速度;将卷积块之间的残差连接改进为密集连接,从而增大感受野;对于输出的三个不同大小的特征图进行二次特征融合,将高层特征信息与下层特征信息结合,让下层信息充分融合上层信息,从而提高对图片特征提取的能力。在PASCAL VOC 2007、WiderFace数据集上实验的结果表明,该方法对于目标检测的精度有明显的提高。

In order to solve the problem of low accuracy of YOLO target detection algorithm,a YOLO target detection algo⁃rithm based on secondary feature fusion is proposed.An improved separable convolution module is used to improve the traditional convolution module,thereby reducing the network model and ensuring the speed of target detection.The residual connection be⁃tween the convolution blocks is improved to dense connections,thereby increasing the experience wild.The output feature maps of three different sizes are subjected to secondary feature fusion,the high-level feature information is combined with the lower-level feature information,so that the lower-level information fully integrates the upper-level information,thereby improving the ability to extract image features.The experimental results on the PASCAL VOC 2007 and WiderFace datasets show that this method can signif⁃icantly improve the accuracy of target detection.