A mathematical model is constructed for the global algorithm to build an energy function, utilizing methods such as Newton’s method and gradient descent to minimize the energy function and find the optimal matching solution. Shahbazi et al. [7] from the University of Calgary employed the non-parametric census transform as the data term and used the geometric features of intrinsic curves as the smoothness term to minimize the energy function, thereby improving the problem of a high mismatch rate caused by image occlusion. Zhou Jiali et al. [8] from Zhejiang University of Technology based their approach on conventional graph cut algorithms. They corrected the matching region based on the labels of the matching blocks and spatial geometric information, continuously updating the selected labels to find the labels that minimize the global energy. By applying the left–right consistency criterion and mean filtering to refine the disparity map, they achieved higher sub-pixel-level accuracy in the matching disparity. The matching accuracy in the edge and occlusion regions of the image pair was significantly improved.

The matching cost convolutional neural network (MC-CNN) was proposed by Zbontar et al. [9] of the University of Ljubljana. In the cost-computation stage, a convolutional neural network is used to train on input image patches and annotated images, obtaining matching cost values. These values are then optimized using a cross-cost aggregation method, incorporating a left–right consistency check and bilateral filtering for refining the disparity map. Seki et al. [10] from Toshiba Corporation introduced the SGMNet algorithm, which utilizes a convolutional neural network to automatically learn penalty parameters, eliminating the manual adjustment process in traditional SGM algorithms. This algorithm divides the disparity transition along the scanline into positive and negative disparity transitions based on different occlusion relationships between objects, ensuring good disparity prediction even in pathological regions. However, non-end-to-end stereo matching algorithms often fail to meet practical requirements due to their high time complexity. They remain within the traditional stereo matching framework and require significant time and effort for parameter adjustments, such as filter size and matching window size.

The end-to-end stereo matching algorithm learns the features in the input images adaptively through a deep learning model, eliminating the need for manual feature design and selection. The model is more adaptable to different scenes and details. Mayer et al. [11] constructed a virtual synthetic dataset and proposed the end-to-end network DispNetC. This network introduced an autoencoder–decoder structure, taking in left and right images and outputting a disparity map without any post-processing steps. Xu et al. [12] from the University of Science and Technology of China extracted features of different resolutions using a shared feature pyramid network. They designed three same-scale aggregation modules to optimize different resolution features and proposed the adaptive aggregation network (AANet), which fuses features through a cross-scale aggregation module to address the issues of large parameter and computational requirements in deep stereo matching networks, thereby improving algorithm efficiency. Vladimir et al. [13] from Google introduced HITNet, a convolutional neural network for real-time stereo matching. It infers disparity through fast multi-resolution initialization and transformation, without incurring significant costs. By propagating information across levels, it reduces algorithm complexity while improving matching accuracy. Tang Haifeng et al. [14] from Inner Mongolia University proposed DFFNet, an end-to-end stereo matching network that incorporates dense feature fusion. They utilized multiple residual modules to construct a feature pyramid network, capturing multi-scale contextual information with a low parameter count. The network enhanced its matching capability in complex regions such as weak texture areas and edges through dense fusion modules and mixed attention modules, thereby improving the extraction of useful information.