A Novel Shape Feature for Image Classification and Retrieval

Abstract

Abstract. In this paper a novel statistical shape feature called the edge co-occurrence matrix (ECM) is proposed for image classification and retrieval. The ECM indicates the joint probability of edge directions of two pixels at a certain displacement in an image. The ECM can be applied to various tasks since it does not require any segmentation information unlike most shape features. Comparisons are conducted between the ECM and several other feature descriptors with two defect image databases. Both the classification and retrieval performances are tested and discussed. The results show that the ECM is efficient and it provides noticeable improvement to the performance of our CBIR system.

Image Feature Extraction Using Gradient Local Auto-Correlations

“…ECM [14] also utilizes a joint histogram of orientation pairs, but it is a special case of GLAC: w ≡ 1 (no weighting) and the G-O vector consists of binary values (0 or 1) in ECM. It suffers from boundary effects of the magnitude and the orientation of image gradients.…”

“…The concept of correlation has also been adopted in self similarity [8], in which extracted edges in Shape Context [7] are substituted with cross-correlation values between local patches at a reference position and its local neighborhoods. Our work is most closely related to ECM [14] which utilizes joint histograms of orientations of gradient pairs. Differences in the details are described in Sec.3.2.…”

Higher-order Co-occurrence Features based on Discriminative Co-clusters for Image Classification

“…Recently, the standard co-occurrence features are founded on the qualitative data (symbols); gradient orientation bins [8,17], indexed colors [7] and visual words [10,24]. These methods first cluster (quantize) primitive quantitative data, e.g., gradient orientations, RGB colors and local features [11], into those symbols and then measure the co-occurrences among them.…”

“…The statistical features are robust to noises and they are fed into subsequent classification methods, such as SVM [19], for accomplishing the image classification. Beyond the histogram-based methods considering occurrences, co-occurrence feature extraction methods have also attracted keen attentions thanks to the superior performances [6,7,8,9,10,17,23,24]. The methods statistically describe the image by using c 2012.…”

“…For extracting the co-occurrences, it is common to transform the quantitative data into qualitative data (symbols) by means of quantization (clustering) at first, as is the case with the histogram-based methods; for example, continuous gradient orientation is coded into orientation bins [8,17], RGB colors are indexed [7] and local features are categorized into visual words [10,24], although a few works [1,13] attempt to directly code joint pairs of local features for BoF. The quantization process taken over from the histogram-based methods is given a priori in an ad-hoc manner based on prior knowledge regarding the point-wise statistics of the quantitative data, not pair-wise data.…”