基于感兴趣区域聚集层的强化学习算法的目标检测

Target detection based on reinforcement learning algorithm of region of interest clustering layer

在传统深度强化学习的目标检测基础上,提出了一种基于感兴趣区域聚集层的策略,改进传统深度强化学习中使用RoIPooling层将感兴趣区域池化为固定尺寸时造成的像素偏差.首先遍历整个感兴趣区域,保持感兴趣区域的浮点数边界;然后将感兴趣区域平均分割成7×7个矩阵单元,保持每个矩阵单元的边界也是浮点数;最后在每个矩阵单元中使用双三次插值法计算固定的4×4个坐标位置,进行最大池化操作.结果表明:在Pascal VOC2012数据集上,改进方法提升了模型的检测精度;智能体(agent)可以在较少的步数里检测出更多图像中的物体,效率更高;在精准率与召回率上优于传统深度强化学习策略,平均精度比传统目标检测方法更有优势.

On the problem of target detection in traditional deep reinforcement learning,a strategy based on region of interest clustering layer is proposed to improve the pixel deviation caused by pooling the region of interest into a fixed size by using RoIPooling layer in traditional deep reinforcement learning.The strategy will first traverse the entire region of interest,keep the floating-point boundary of the region of interest.And then divide the region of interest into 7×7 matrix cells averagely,keep the boundary of each matrix cell is also a floating-point number.Finally,the bicubic interpolation method is used to calculate the fixed 4×4 coordinate positions in each matrix cell to maximize the pool operation.The experimental results show that:on the Pascal VOC2012 dataset,this improvement helps to improve the detection accuracy of the model;the agent can detect more objects in the image in less steps,with higher efficiency;the accuracy and recall rate are better than the traditional deep reinforcement learning strategy,and the accuracy and average accuracy rate are better than the traditional target detection method.