Simulation of Saccadic Eye Movements
Output of Simulator showing low and high quality text
Abstract
In recent times digital cameras have become affordable, leading to the desire to use digital cameras for text detection too. The pictures recorded by a camera usually have a resolution too low for classical OCR-applications. Also they can be rotated, perspective distorted and lie on a curved surface.
Up to now either cameras with a very high-resolution were used (expensive and not always practicable) or a mosaicing technique was used. In the latter case many low-resolution but high-zoom pictures were taken and then combined by stitching into one large picture. This work now presents a third idea. Just like in the mosaicing technique, many low-resolution, high-zoom pictures are taken but they are not combined into one large picture. Instead the high-zoom pictures are chosen in such a way that each contains a complete word using stitching if the word is too long to be recorded at once. The combination takes place after the words have been processed by the OCR.
The work presented here detects the lines of text and the words therein in a low-resolution picture of a text. It can be either rotated and perspective distorted or no more than slightly rotated but lie on a curved surface. It then chooses the optimal camera positions. Taking extracts from a high-resolution picture of the text simulates the camera.