Performance analysis of the Edge Pixel Orientations Histogram

A study on the performance of the Edge Pixel Orientations Histogram is presented in this paper. Edges are detected with the Canny algorithm without and with hysteresis thresholding. The resulting edges are described in No different orientations using the gradient orientation. The two edge images are divided into N × N sub-images. Counting the number of edge pixels with each orientation for each sub-image results in a histogram with 2N
_o
N
²
bins. The image descriptor is tested with a subset of the TRECVID 2008 development database k-frames. The resulting Edge Pixel Orientation Histograms will be classified with the K Nearest Neighbour Algorithm and a high level description based on the image semantics is extracted. In particular the concept “images with at least one building” is tested. This descriptor is also used for the JPSearch Alinari image database annotation. Alinari database is composed by a set of 971 high resolution images (3888×2592) shot by professional photographers.