Design and Development of Medicine Text Identification System

Abstract

Recent studies in the field of computer vision and pattern recognition show a great amount of interest in content retrieval from images and videos. This content can be in the form of objects, colours, texture, shape as well as relationship between them. Text in images are important source of information for several advanced applications; such as, video and image retrieval, web image search, multilingual translation, content based automatic annotation of image databases and assisting visually impaired to read labels in map applications . The aim of text localization and detection is to find the regions in the medicine considered text by humans, mark text boundaries (usually by rectangular bounding boxes) and produce the associated characters. Text Extraction and identification from medicine strips and bottles is a valuable application that can help pharmacies to create their own medicine databases and assist the patients with the medicine information i.e., the salts present in the medicine and the possible substitutes to the medicine. The problem of text extraction is a challenging one due to variety of text variations on medicines such as; font, size, colour, alignment, illumination and reflection. In this dissertation, we put forth an accurate and robust medicine text detection algorithm. With the present technique, images of complete, partial, distorted or occluded medicine strip or bottled medicine can be used to identify the medicine name. The text extraction and identification is performed with the help of edge enhanced Maximally Stable Extremal Regions (MSERs) followed by geometric filtering and Stroke Width Transform to remove the non-text regions. Next, the OCR system uses Stroke Width transformed image and the region of interest in order to recognize and display the text string. To deal with missing and extraneous characters during recognition novel string editing is applied to extract the correct medicine name. The algorithm is evaluated on a dataset containing both of medicines strips and bottled medicines. Since there is no existing database, we created the database with a wide variety of medicines from different pharmacies using a regular camera under natural lighting conditions. The experimental results exhibit excellent performance for the proposed technique. The system gives an efficiency of approximately 95% on medicine strips and bottled medicines images captured under varied conditions; for example, reflection, bad illumination and skew-ness. The method is capable of detecting highly blurred text in low resolution medicine images as well as rotated text for the bottled medicines.