摘要
为实现较小目标低分辨率的精确实时测量,提出了基于Faster-RCNN目标检测的改进算法,通过结合特征提取阶段的较浅层卷积神经网络的小感受野目标特征,实现目标检测的精细化。同时将模型的全连接层替换为卷积层,结合FPGA优良的并行处理性能,实现算法的加速处理,并在Small Object Dataset[13]数据集上进行实验验证,取得了较优的性能,与改进前的算法相比,准确度和速度都有较大提升,将提出的目标检测方法应用到实际的中小目标低分辨率识别定位场景是可行的。
In order to achieve accurate real-time measurement of small targets with low resolution,an improved target detection algorithm based on Faster-RCNN is proposed.By combining the small receptive field target features of the shallow convolution neural network in the feature extraction stage,the target detection is refined.At the same time,the full connection layer of the model is re⁃placed by the convolution layer,Combined with the excellent parallel processing performance of FPGA,the accelerated processing of the algorithm is realized,and the experimental verification is carried out on the Small Object Dataset[13]data set,and better perfor⁃mance is achieved.Compared with the algorithm before improvement,the accuracy and speed are greatly improved.It is feasible to ap⁃ply the proposed target detection method to the actual low-resolution recognition and positioning scene of small and medium targets.
作者
胡晶晶
Hu Jingjing(College of Computer Science and Engineering,Northeastern University,Shenyang 110819)
出处
《现代计算机》
2021年第30期82-87,共6页
Modern Computer
