A Deep Learning Framework for Corner Point Localization and Shape Recognition in 2D Drawings Using Long Short-Term Memory Networks

Abstract

Corner point detection and shape recognition in 2D drawings represent fundamental tasks in sketch-based modeling, human–computer interaction, and intelligent CAD systems. Traditional geometric and curvature-based methods often struggle with noise sensitivity, stroke variability, and ambiguous sketch structures. This research proposes a deep learning framework leveraging Long Short-Term Memory (LSTM) networks for robust corner point localization and shape recognition in hand-drawn 2D sketches. The model treats stroke sequences as temporal data, enabling contextual learning of geometric transitions and structural dependencies.
The proposed approach integrates sequential stroke representation with bidirectional LSTM encoding to capture both forward and backward dependencies in drawing trajectories. The framework is conceptually aligned with prior machine learning-based segmentation techniques such as RankFrag and ClassySeg, while extending their capability through deep temporal modeling (Costagliola et al., 2015; Herold & Stahovich, 2011). Classical corner detection algorithms and curvature-based approaches provide baseline theoretical grounding, particularly in comparison to discrete geometry and scale-space methods (Nguyen & Debled-Rennesson, 2011; Zhong & Liao, 2007).
Experimental insights suggest that sequence-aware modeling improves robustness in detecting corner points under stroke distortion, variability, and incomplete sketch inputs. The study further highlights the role of LSTM-based architectures in bridging the gap between rule-based geometric detection and data-driven learning systems. The framework demonstrates potential applicability in CAD automation, sketch recognition systems, and intelligent drawing interpretation tools.