一种迭代机制改进基于卷积神经网络的目标检测

Improved Object Detection Based on Convolutional Neural Network with Iterative Mechanism

为提升目标检测模型的检测精度,提出一种迭代机制改进基于卷积神经网络的目标检测方法。首先采用与标准更快速区域卷积神经网络(Faster RCNN)一致的设定,在提取了区域候选框后,引入迭代机制来改进Faster RCNN。通过多次迭代优化候选框,使检测框近似于真实框。为发挥迭代机制优点,实现目标精确检测,改进了迭代版Faster RCNN的训练方式,一种改进是所有迭代步骤都定义了损失函数,另一种改进是使用ε-greedy策略。最后,在PascalVOC数据集和自制飞机数据集上进行的测试表明,改进后的迭代版Faster RCNN的检测精度高于标准Faster RCNN约8个百分比。

An improved object detection method based on convolutional neural network with iterative mechanism is proposed to improve the detection accuracy of object detection model.Firstly,the proposed method adopts the same settings as standard faster region convolutional neural network(Faster RCNN).Iterative mechanism is introduced to improve Faster RCCN after extracting the region candidate box.Then,the candidate box is optimized by iteration for many times,so that the final detection box can be similar to the real box.In order to take advantage of the iterative mechanism and achieve accurate object detection,the training method of the iterative version of Faster RCNN is improved.One improvement is to define the loss function in all iteration steps,and the other is to use the strategy ofε-greedy.Finally,tests on the data sets of Pascal VOC and aircraft show that the detection accuracy of the iterative version of Faster RCNN is about 8 percentage higher than that of the standard Faster RCNN.