Resource-constrained stereo matching for greenhouse tomatoes and napa cabbages: An efficient and accurate approach with optimized memory usage

Abstract

This study introduces a novel stereo matching algorithm designed for agricultural applications in greenhouses. Addressing the limitations of existing convolutional neural network (CNN)-based stereo matching models, the proposed architecture enhances efficiency by converting RGB images to grayscale and applying histogram equalization. This method preserves essential visual information while reducing data size and computational complexity. Rectangle filter kernels are also used to prioritize horizontal information, aligning with the typical arrangements of stereo camera pairs. The disparity prediction model is initially trained on a subset of 29,204 image pairs from the Scene Flow dataset and subsequently retrained and evaluated using 2,470 tomato image pairs and 486 napa cabbage image pairs from Greenhouse. This further refines its performance in real agricultural settings. The proposed model surpasses existing models such as geometry and context network (GC-Net), pyramid stereo matching network (PSMNet), 2D-Mobilestereonet and 3D-Mobilestereonet in terms of disparity prediction accuracy and computational speed, consuming less than one-third of the memory. Given the unique challenges of greenhouse environments, this approach demonstrates a robust method for developing stereo matching algorithms suited for stereo vision applications in such settings.