基于改进YOLO v3的目标检测算法

Target Detection Algorithm Based on Improved YOLO v3

随着深度学习的不断发展与广泛运用,基于深度学习的目标检测算法已成为新的主流。为了进一步提高卷积神经网络YOLO v3(You only look once v3)的检测精度,在原算法的网络结构上添加卷积层模块对样本进行目标背景分类,并粗略调整特征图上的锚框大小。该模块输出目标背景概率后,过滤掉背景概率值低于设定阈值的样本,从而解决原算法中存在的正负样本比例失衡的问题。使用调整过的锚框替代原算法中直接由聚类生成固定大小的锚框,该过程为边界框的预测提供更优的初始值。在VOC数据集上的实验结果表明,相较于原算法,改进的YOLO v3具有更高的检测精度。

With the continuous development and wide applications of deep learning,target detection algorithms based on deep learning have become a new mainstream.To further improve the detection accuracy of the convolutional neural network YOLO v3(You only look once v3),a convolution layer module was added to the network structure of the original algorithm to classify the target background of the sample and the anchor frame size of the feature map was roughly adjusted.To resolve the challenge of unbalanced proportion of positive and negative samples in the original algorithm,samples with background probability value less than the set threshold value were filtered by the module after outputting the target background probability.The adjusted anchor box was used to replace the anchor box of fixed sizes directly generated by clustering in the original algorithm.This process provides a better initial value for bounding box prediction.Experimental results on VOC dataset indicate that the improved YOLO v3 shows higher detection accuracy than the original algorithm.