基于双流卷积神经网络的RGB-D图像联合检测

Joint Detection of RGB-D Images Based on Double Flow Convolutional Neural Network

当前卷积神经网络结构未能充分考虑RGB图像和深度图像的独立性和相关性,针对其联合检测效率不高的问题,提出了一种新的双流卷积网络。将RGB图像和深度图像分别输入到两个卷积网络中,两个卷积网络结构相同且权值共享,经过数次卷积提取各自独立的特征后,在卷积层根据最优权值对两个卷积网络进行融合;继续使用卷积核提取融合后的特征,最后通过全连接层得到输出。相比于以往卷积网络对RGB-D图像采用的早期融合和后期融合方法,在检测时间相近的情况下,双流卷积网络检测的准确率和成功率分别提高了4.1%和3.5%。

The convolutional neural network structure fails to consider the independence and correlation between RGB images and depth images fully, so its detection is not high. A new double flow convolution network is proposed for the joint detection of RGB-D images. The RGB image and depth image are inputted to the two convolutional networks and the two networks have the same structure and weight sharing. After several convolutions, the independent features are extracted. According to the optimal weights in the convolution layer, the two convolutional networks are fused. The fused features are extracted continuously using convolution kernels, and the output is obtained by full connection layer finally. When the detection time is similar, the detection accuracy and the success rate are increased by 4.1% and 3.5% respectively, compared with the previous early and late fusion methods.