This project is related to learning recongnise attention point in an image data. Each pixel of an image is classified into two categories whether it belongs to the salient object or not. Different features of images are combined using Conditional Random Field to classify the image in the two categories.
salient object in an image or in a sequence of images is part of images which is the main attention point. In static object it can be described as different in color,pattern, orientation and in sequences it motion also creates attention. The salient object detected is detection of some familier object but the identity of the object is not knon. Salient objects can be used to crop images autometically, displaying on small screens and classifying the image data. In image sequence it may be used to cut video in parts, tracking objects. In our project we will first detect salient object in a static image later on we will do it in image sequences.
Most of the saliency detection method are bottom up are approach in which low level features of images such as intensity, colors are used. The previous well known method for salient object detection is Itti-koch[2] Method. Here three features of images were takem int account color, intensity and orientation. Using these some number of color, intensity and oreientation map are created and are then combined to give saliency map. saliency is computed by a center-surround operation, self-information , or graph-based random walk using image features.
In our project, in order to recognise salient object, we compute a binary mask for each pixel of an image
which describes the presence of that pixel in the salient object of the image.This mask is calculated on CRF
learning of the combinations of the features of salient object features and pairwise features of an image.
The salient object features of the image used used in CRF learning is :
a)
MultiScale Contrast : The major factor aused to detect saliency in image is contrast. We will
calculate the Gaussian image pyramid of the image and take linear combination
of the apropriate levels for contrast computation.
b)
Center-Surround Histogram :
Here we compute the histogram for a small window of different aspect ratio arround each pixel using some
funtion and the data used is the histogram data for the RGB colors of the image.
b)
Color Spatial Distribution :
This method is used to include global features of color in combination of salient features. Here we see
existance of any global color. MOre dominant a color is in an image less likely probability of the salient
object having it.
The pair wise features of an image is the relationship between a pixel and it's adjacent pixels.
[1] Tie Liu, Zejian Yuan, Jian Sun, Jingdong Wang, Nanning Zheng, Fellow, IEEE, Xiaoou Tang, Fellow, IEEE, and Heung-Yeung Shum, Fellow, IEEE, 2011.
Learning to Detect Salient Object
[2] Laurent Itti, Christof Koch and Ernst Niebur, "A Model of Saliency-based Visual Attention for Rapid Scene Analysis"
[3] Stéphane Nicolas1, Julien Dardenne2, Thierry Paquet1, Laurent Heutte1, "Document Image Segmentation Using a 2D Conditional Random Field Model"
[4] MSRA Salient Object Database
Database