Abstract:

To realize a lightweight and high precision binocular stereo matching network,we propose a progressive dilated residual depth binocular stereo matching network:PDR_Net.In the feature extraction network module,a progressive dilated residual network structure is proposed.The dilated convolution network replaces the pooling down-sampling method to obtain the multi-scale feature information of the image,which can reduce image feature information loss caused by scale transformation in pooling down-sampling.At the same time,the residual network is introduced to alleviate the loss of image feature information from the characteristics of the dilated convolution network.The progressive cascade method is used to fuse the feature information between the branches of each scale,which promotes the fusion of the feature information from each image scale,namely,the strategy can reduce the network complexity and retain more image features.Finally,in the 3D convolutional network module,the stacked sand drain coding and decoding network structure is adopted,and the feature map can be effectively combined by the jump connection.The channel attention mechanism model is introduced,which enhances the aggregation learning ability of the network between the feature information of different disparities from different channels,and deepens the connection of the feature points from different disparities.Compared with the existing network,our proposed PDR_Net network has the advantages of less parameters,faster speed and higher accuracy.

Key words: stereo matching, progressive network, residual network, channel attention mechanism