Probably too much normalization which can flatten the result, try using some statistical functions such as correlation or cross correlation to link the result with and without weighting the descriptors.

Pierre De Handschutter FLAN or BF matchers with SIFT, SURF or ORB like key point descriptors would work reasonably only when the same or part of the query image is appeared (in different scale, orientation, illumination and with other variations) in the matching images. But that is not the case in Corel dataset hence those method would not work as expected. If you really need to use them, try using Bag-of-Feature models of SIFT, SURF and ORB with SVM like classifier. You may refer the following code.

https://www.codeproject.com/Articles/619039/Bag-of-Features-Descriptor-on-SIFT-Features-with-O

Color histograms (including all the color spaces such as RGB, HSV and Lab etc) have their limits and its obvious they do not hold the necessary information about the shapes of the objects appeared in the image. Therefore no wonder about the poor results in color histograms based CBIR.

LBP is better at discriminating objects in a particular domain (like faces, vehicles etc) and it is vulnerable in general image retrieval applications where the images having highly cluttered background.

GLCM along has the same problem where it only characterizes the texture details which may not be the discriminating feature for a pair of images. Hu in the other hand should not use for natural image classification where it performs reasonably with single object having a well defined shape.

The best non deep learning based approach I found so far is the Fused Color Gabor Pyramidal Histogram of Oriented Gradients (FC-GPHOG) descriptor classified with nearest neighbor on Enhanced Fisher Model (EFM). You can find my implementation of FC-GPHOG from:

https://github.com/ravimalb/fc-gphog

The code related to the work mentioned in the paper :

Sinha, Atreyee, Sugata Banerji, and Chengjun Liu. "New color GPHOG descriptors for object and scene image classification." Machine vision and applications 25, no. 2 (2014): 361-375.

EFM is nothing but the applying LDA on truncated PCA projection of the FC-GPHOG descriptor.

Ravimal Bandara Thank you very much for your detailed explanations ! Have you tested "FC-GPHOG" on some queries from Corel so far ?

Moreover, for comparing descriptors, is compareHist method (with all norms ?) still relevant (I guess yes but prefer to be sure) ?

When you call the function `computFCGPHOG(input,output, 16, 2); `, how did you choose your parameters ? Thank you in advance

Regards

Pierre De Handschutter As I remember I tested it with Caltech datasets.

Descriptors can be compared with a particular norm only if those descriptors are in a space where that norm fairly quantify the similarity of the descriptors. For an example, binary descriptors typically cannot be evaluated with L2 norm. Otherwise, compareHist should give a fair result.

in computFCGPHOG(input,output, 16, 2), 16 is number of bins and 2 is the pyramidal levels. Higher the number of bins, higher the descriminative power but lower robustness and wise versa. Higher the levels would grab more details over different scales yet lower the robustness. You may try (16,2), (8,2), (16,4) and (8,4).

Eduardo Daniel Di Santi