I want to know if features extracted from a co-occurrence matrix of an image scale, orientation, and shift invariant.
If I understand the question correctly, no, they would not, because of the way the co-occurrence matrices are computed. But I guess this would all depend on the type of images used (sensor type, resolution, features observed). Co-occurrence matrices are calculating the number of times similar levels are found, at a certain distance and (often) at a certain angle from each other, within a particular window. So scale invariance would not work, as taking a smaller window would look at smaller-scale processes (even if normalising the matrix, we would still look at something different, like a satellite image at full and medium resolution). Orientation invariance: this will depend on the way they are calculated (GLCMs often use only one direction). Shift invariance: if this is the shift of the computation window, there would be no invariance as we would observe different objects. If, which seems most likely, this is the shift in image levels, the co-occurrence matrix would have its peaks shifted one way or the other, but correctly-defined textural measures would account for that.
Co-occurence features are computed for assigned values of distance and orientation. As a consequence, they are both scale and orientation-dependent. Invariance against rotation can be obtained in various ways - most commonly by averaging the features computed at the different orientations (for a fixed distance value). As for shift invariance, I would say co-occurrence features are quite robust if the input image is 'stationary' enough.
Te feature extracted from the co-occurrence matrix are not invariant. Analyze the feature you need and looking for characteristics that can made them invariant. Let me know what's kind of images are you processing and the features you want to extract in order to give you an idea if you want.
The features extracted from the co-occurence matrix are not rotation invariant, but it is actually translation and scale invariant. Averaging by using 0, 45, 90 and 135 degrees can be used to overcome rotation problems. I have applied this approach on plant identificaton. See my paper "Neural Network Application on Foliage Plant Identification".
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing