Object Detection in Natural Scene Images Using Thresholding Techniques

Abstract

A lot of attention has emerged regarding the aspects of text detection and identification as OCR has generated a lot of prominence over the years. There has been a number of experiments conducted in this field to make the results more and more accurate. Most of the experiments carried out have paid attention to only a few attributes and not a lot of trails have been done for unusual scenarios, like a lot of techniques produces accurate results only for horizontal textual orientation. So there should different techniques for analyzing such images which have a complex background, different font styles, colors, textual orientations. Text detection on images containing texts of different orientations, different font types, and images with complex backgrounds is taken for the proposed work. There are mainly 3 steps in the algorithm proposed, the Canny edge detection approach for gradient filtering is applied in the first stage to detect the skeletal structure of various objects in the image. In the next stage textual threshold-based object filtering is carried out using the convolution technique with a heuristic thresholding model. The textual object filtering after convolution is subject to the last stage called post enhancement technique. In this stage, partial non-textual objects being filtered out are employed for removal based on geometrical properties of gradients of images, thus retaining only the textual objects. Finally, the textual object filtered gradient image is considered as a mask image for mapping it to the original image for text detection. Experimentations are conducted on Google Street View Datasets for which a subjective evaluation procedure is adapted to validate the results resulting in promising outcomes for more than 50% of images.