Depth analysis of monocular natural scenes using gray level co-occurrence matrix

Representation of depth in a real world environment is an essential attribute of its semantic representation. A coarse estimate of image-depth (defined as mean distance between the object and the observer) is relevant for identifying the context of the scene and can be used to facilitate search and recognition of objects. In this paper, a GLCM based scheme is proposed to analyze the depth information of real world natural scenes. A distant image being smoother has a low value of dissimilarity. This quantization helps in the categorization of scenes into three classes viz. `near' (less than 5 meters), `not-so-near' (about 50 meters), and `far' (beyond 500 meters). In the proposed method, at each image pixel, a set of co-occurrence matrices is calculated for different orientations and inter-pixel distances. From these matrices, dissimilarity feature is extracted which characterizes the neighborhood of the concerned pixel. Image features thus extracted are used to classify natural scene images into `near', `not-so-near' and `far' categories with the help of a probabilistic neural network classifier.