为了提高基于双目视觉中立体匹配在弱纹理场景下的精准性,提出了一种基于注意力机制特征提取的三维重建算法。利用卷积神经网络(CNN)训练左右图像的特征表示,计算出立体匹配的匹配代价。在CNN特征提取阶段,加入图像注意力机制模块和通道注意力机制模块,得到特征图各个像素点之间的联系,使网络可以更好地捕获图像上下文信息,进而在重建过程中能够更加精确地重建出弱纹理区域。对于网络损失函数,集成了语义编码损失,最终将损失函数定义为语义编码损失和重建损失的加权和,有效提升了弱纹理区域下的重建精度。使用KITTI和Sceneflow数据集对算法进行验证,实验结果证明,相比目前国内外先进方法,本文算法在精度方面有较大提升,尤其体现在弱纹理区域。

To improve the accuracy of stereo matching based on binocular vision applied to weak texture scenes, this study proposes a 3D reconstruction algorithm based on feature extraction using an attention mechanism. The proposed model uses convolutional neural network (CNN) to train feature representation of left and right images and calculates the matching cost of stereo matching. First, during the CNN feature extraction stage, attention mechanism module and channel attention mechanism module are summed to obtain the connection of each pixel in the feature image, enabling the network to capture the context information better and reconstruct weak texture areas more accurately in the reconstruction process. Second, we integrate the semantic coding loss in our neural network. The final loss function is defined as the weighted sum of the semantic coding loss and the reconstruction loss, which can effectively improve the reconstruction accuracy of a region with weak texture. We use KITTI and Sceneflow datasets to validate the algorithm. Experimental results show that the proposed method yields good improvements in accuracy, particularly in areas with weak textures.