A 3D point cloud and deep learning based automated process for quantifying multi-scale phenotypes in sliced bread

In response to the current focus on two-dimensional phenotypic analysis of sliced bread evaluation parameters, this study developed an innovative technology using an inexpensive 3D laser scanner to accurately capture and analyze the three-dimensional structure of both the entire bread slice and the pore surface of the slice. A 3D line-scan laser profiling sensor integrated with a three-axis motion platform was used to efficiently acquire 3D point cloud data of bread surfaces. For the pre-processed point cloud data, a segmentation model called 3D-PoreSegNet was proposed to separate the bread surface and pore regions within the point cloud. By using 2D projection, contour extraction and inverse transformation techniques, accurate 3D structural information of the pore edges was obtained. Based on the pore edges and the entire point cloud data, we successfully reconstructed the 3D pore structures and extracted detailed phenotypic parameters for both the bread and its pores. Furthermore, two novel phenotypic parameters, pore distribution and pore layer distribution, were proposed to provide more scientific support for comprehensive quality assessment of bread. To facilitate rapid phenotypic analysis, we developed the Bread3D-Measure software. Experimental results showed high accuracy in extracting 13 phenotypic parameters, including seven total traits - height (97.3 %), length (95.2 %), width (95.6 %), surface area (86.6 %), volume (82.8 %), symmetry index (91.5 %) and uniformity index (93.4 %) and six pore-related parameters - maximum pore diameter (84.7 %), maximum pore area (82.6 %), pore count (87.1 %), maximum pore depth (90.3 %), maximum pore volume (83.7 %) and pore elongation (78.5 %). These results confirm the effectiveness of the proposed method in meeting the requirements of various phenotypic applications.