Основные понятия
MolNexTR is a novel deep learning model that accurately predicts molecular structures from diverse image styles, achieving superior performance in molecular structure recognition.
Аннотация
MolNexTR is a deep learning model designed to recognize molecular structures from various drawing styles prevalent in chemical literature. It combines ConvNext and Vision-Transformer to extract local and global features, predict atoms and bonds, and understand layout rules. The model incorporates advanced algorithms for data augmentation, image contamination simulation, and post-processing to enhance robustness against diverse imagery styles. MolNexTR outperforms previous models with an accuracy rate of 81-97% on test sets, marking significant progress in the field of molecular structure recognition.
Статистика
MolNexTRは、テストセットで81〜97%の精度を達成し、分子構造認識の分野で重要な進展を示しています。
モデルは、局所的およびグローバルな特徴を抽出し、原子と結合を予測し、レイアウト規則を理解するためにConvNextとVision-Transformerを組み合わせています。
モデルは、データ拡張、画像汚染シミュレーション、および後処理のための高度なアルゴリズムを組み込んでおり、さまざまな画像スタイルに対する堅牢性が向上しています。
Цитаты
"MoIVec [40] achieves good performance on CLEF, UOB, and USPTO datasets but declines on ACS due to diverse drawing styles."
"MolNexTR combines ConvNext and Vision-Transformer to extract local and global features for accurate prediction of atoms and bonds."
"MolNexTR demonstrates exceptional performance on multiple challenging datasets including Indigo, ChemDraw, RDKit, CLEF, UOB, USPTO, Staker, and ACS."