Vision-Based Counting of Stacked MDF and Particleboard Panels: Detection, Density Regression, and Structure-Aligned 1D Modeling

Counting stacked wood-based panels from side-view images is difficult because adjacent layers have similar appearances and weak boundaries. In this study, we introduce an annotated medium-density fiberboard (MDF) and particleboard image dataset and Stack1D, a structure-aligned density-regression architecture. The dataset comprises 409 images of single-thickness stacks and 24 images of mixed-thickness stacks. Stack1D combines two-dimensional feature extraction with width-wise feature aggregation and one-dimensional density decoding along the stacking axis. We compare Stack1D with six YOLO detectors and CSRNet-based density models across three progressive training-count settings, using higher-count validation data for model selection. Evaluation covers maximum-count images of single-thickness stacks and a separate mixed-thickness test set. The best-performing density architecture varies with material, training stage, and test condition. In the maximum-count tests, Stack1D variants achieve the lowest density-model errors in several settings, while channel-reduced two-dimensional models remain competitive, particularly when both materials are used for training. On the evaluated mixed-thickness stacks, the best-performing YOLO detectors achieve lower counting errors than the density models. These findings characterize the relative strengths of detection and density regression for panel counting under shared acquisition conditions.