A comparative approach on detecting multi-lingual and multi-oriented text in natural scene images

摘要

Text helps to convey the intended message to users very accurately. Detecting text from natural scene images for quadrilateral-type and polygon-type datasets is the primary scope of this work. A regression-based method using modified You Only Look Once YOLOv4 network is used for quadrilateral-type datasets. Hyperparameters for training the network are optimized using the Genetic Algorithm which proves to be a suitable candidate than traditional methods. The Pixels-IoU (PIoU) loss is introduced to derive an accurate bounding box and it seems to be productive under various challenging scenarios with high aspect ratios and complex background. This yielded quick results for quadrilateral-type datasets but did not scale for arbitrarily-shaped and curved scene text. So the approach is changed to segmentation based for enhancing the results. This introduces binarization operation in a segmentation network to boost its detection accuracy for polygon-type datasets. The introduction of a new module DiffBiSeg (Differentiable Binarization in Segmentation network) facilitates post-processing and text detection performance by setting the thresholds flexibly for binarization in the segmentation network. The efficacy of both approaches is clearly seen in their respective experimental results.