Scene Based Text Recognition From Natural Images and Classification Based on Hybrid CNN Models with Performance Evaluation

Abstract

Similar to the recognition of captions, pictures, or overlapped text that typically appears horizontally, multi-oriented text recognition in video frames is challenging since it has high contrast related to its background. Multi-oriented form of text normally denotes scene text which makes text recognition further stimulating and remarkable owing to the disparaging features of scene text. Hence, predictable text detection approaches might not give virtuous outcomes for multi-oriented scene text detection. Text detection from any such natural image has been challenging since earlier times, and significant enhancement has been made recently to execute this task. While coming to blurred, low-resolution, and small-sized images, most of the previous research conducted doesn’t work well; hence, there is a research gap in that area. Scene-based text detection is a key area due to its adverse applications. One such primary reason for the failure of earlier methods is that the existing methods could not generate precise alignments across feature areas and targets for those images. This research focuses on scene-based text detection with the aid of YOLO based object detector and a CNN-based classification approach. The experiments were conducted in MATLAB 2019A, and the packages used were RESNET50, INCEPTIONRESNETV2, and DENSENET201. The efficiency of the proposed methodology - Hybrid resnet -YOLO procured maximum accuracy of 91%, Hybrid inceptionresnetv2 -YOLO of 81.2%, and Hybrid densenet201 -YOLO of 83.1% and was verified by comparing it with the existing research works Resnet50 of 76.9%, ResNet-101 of 79.5%, and ResNet-152 of 82%.