一种优化的深度学习立体匹配算法

Optimized Deep Learning Stereo Matching Algorithm

现如今用于立体匹配的深度学习算法都存在网络结构复杂、消耗大的问题。为解决此类问题,提出了一种参数量只有参考网络PSMNet一半的立体匹配端到端网络结构。所提结构在特征提取模块保留大致框架的同时,减少多余卷积层,并融合空间注意力机制和通道注意力机制来汇聚上下文信息;在代价计算模块,通过加大偏移步长减少视差计算输入的视差维度,使视差计算的参数量和消耗大幅度减少;在视差计算中,对匹配成本特征体的输出进行多视差预测;在L1损失函数的基础上加入交叉熵损失函数,这样可在降低消耗的同时保证了模型匹配精度。在KITTI数据集和SceneFlow数据集上对所提模型进行测试,实验结果表明:与基准方法相比,所提模型的参数量减少了58%,精度提升24%。

Nowadays, deep learning algorithms used for stereo matching have the problems of complex network structure and high consumption. In order to solve such problems, an end to end stereo matching network structure with only half the parameters of the reference network PSMNet is proposed. In the feature extraction module of the proposed network, the general framework is retained, its redundant convolutional layers are reduced, and meanwhile the spatial attention mechanism and channel attention mechanism are integrated to gather contextual information. In the cost calculation module, the input disparity dimension of the disparity calculation is reduced by increasing the offset, and therefore, the parameter amount and consumption of disparity calculation are greatly reduced. In the disparity calculation, the multi-disparity prediction is performed for the output of the matching cost feature body. And the cross-entropy loss function is added to the L1 loss function, which ensures the matching accuracy when reducing the consumption of the model. The proposed algorithm is tested on the KITTI dataset and SceneFlow dataset. The experimental results show that compared with the benchmark method, the parameter amount of the proposed model is reduced by 58% while the accuracy is increased by 24%.