Washroom Sign Detection Using Convolutional Neural Network in Natural Scene Images

Due to disabilities, visually impaired or blind people face difficulties to recognize washroom sign in public places by themselves. In natural scene images, so many objects are present that are similar to human shaped male or female washroom sign making it more difficult to detect and classify between male and female washroom sign. Moreover, at a certain distance, a human body is also look like washroom sign, where system might get confused to classify between a real human figure and a human shaped washroom sign. Focusing on this issue, deep learning- based methods are proposed to detect common patterns of washroom signs in natural images. In this proposed method, MSER algorithm is used for object detection, Geometrical properties algorithm is used for text part and unwanted part removal and then region of interest has been detected by bounding box algorithm, at last CNN is used to classify washroom sign images into three different classes, i.e. `washroom sign', `female washroom sign' and `men washroom sign'. Our CNN classifier gives an accuracy with 96%- 99%. The experimental results were compared with other methods such as SVM, HOG, AdaBoost, MCT to compare the accuracy results with our proposed method which is described in Proposed Method.