A New Stereo Matching Function by a Hybrid Convolutional Neural Network

Abstract

This research introduces a novel method for creating stereo matching algorithms, combining elements of learned and custom methods. The purpose of this strategy is to provide more reliable outcomes than conventional approaches. The primary goal of the technique for stereo corresponding remains to harvest a difference map. Three-dimensional (3D) reconstruction is only one of many uses for this map. Convolutional neural network (CNN) produced raw disparity maps may still contain mistakes in low-resolution textures. The programme will utilise a hybrid CNN-based approach and a faster directional intensity computation to enhance the accuracy of the matching cost calculation phase. Use of the modified truncated value of directional intensity has the potential to greatly reduce radiometric errors. To further enhance the accuracy attained by the approach, a cost aggregation phase is used in which the bilateral filter (BF) is applied to the raw matching cost. Building a gap map with the whole cost of costs is the first step in the WTA optimisation process. The final disparity map is an improvement above the basic map thanks to several rounds of refinement. By means of the Middlebury Online Stereophonic Benchmarking Scheme, the article verified the algorithm efficacy.