基于区域激活策略的Tiny YOLOv3目标检测算法

Tiny YOLOv3 Target Detection Algorithm Based on Region Activation Strategy

针对Tiny YOLOv3模型检测精度低的问题,提出一种将分割信息引入深度卷积神经网络结构中的方法。模型训练期间,将目标真实的位置信息加入网络层中,并手动激活这些目标区域,激励的大小随着训练的进行逐渐减小直至降为零。测试结果表明,在VOC2007数据集上,改进后的Tiny YOLOv3模型的平均准确率提升至58.9%,并且在检测速度上与原模型保持一致,满足实时检测的需要。

Aiming at the problem of low detection accuracy of Tiny YOLOv3 model,a method to introduce segmentation information into deep convolutional neural network structure is proposed.During the model training,the real position information of the target is added to the network layer,and these target areas are manually activated.The size of the excitation gradually decreases as the training proceeds until it drops to zero.The test results show that on the VOC2007 data set,the average accuracy of the improved Tiny YOLOv3 model is increased to 58.9%,and the detection speed is consistent with the original model to meet the needs of real-time detection.