Document image layout detection from scientific literature using combined ConvNext and cascade mask RCNN networks

Scientific publications contain many types of information needed by researchers in various disciplines and are categorized into abstracts, bibliographies, and sections documenting related work, experimental methods, and results. However, due to their different layouts and the different scales of target regions present in them, existing extraction methods or models cannot accurately extract and recognize them across different layouts. To address this problem, this paper proposes a vision-based network, CNC-net, which consists of a lightweight convolutional feature extraction network ConvNext, and a Cascade Mask RCNN. ConvNext represents a markedly efficient convolutional network for feature extraction. Distinguished by its refined structural enhancements over conventional convolutional frameworks, ConvNext demonstrates heightened precision in extracting document image features. Concurrently, Cascade Mask RCNN exhibits a comparable proficiency in feature extraction from document images. ConvNext emerges as a proficient convolutional network tailored for feature extraction, showcasing an enhanced capacity for precise feature extraction from document images through structural refinements relative to conventional convolutional architectures. Meanwhile, Cascade Mask RCNN achieves heightened document image recognition proficiency at elevated Intersection over Union (IOU) thresholds, leveraging a multi-detector approach characterized by varying threshold configurations. We extend the document image dataset GROTOAP2 by a data enhancement method, and build a higher-quality dataset GROTOAP_Next. The proposed model is applied to the two publicly datasets and the extended-enhanced GROTOAP_Next. GROTOAP_Next dataset for experiments and compared with state-of-the-art methods algorithms, the mean accuracy attains 92.2%, accompanied by negligible supplementary computational overhead, thus exemplifying commendable scalability and practical utility. The average performance improvement is 1.6% on small target detection and 3.1% on medium target detection.