Table 1 Compact detail of competitive methods of CBIR
Method | Problem addressed | Extracted features | Dimension reduction | Clustering | Classification | Similarity measure | Shortcomings |
|---|---|---|---|---|---|---|---|
SIFT-FREAK [25] | Semantic gap | SIFT-FREAK/color and texture features | Not applied | k-means | SVM | L2-norm | Hand-crafted features, computationally expensive |
Optimized TPTSSR [26] | The computational expense of sparse classifiers, semantic gap | TPTSSR features | Not applied | Not applied | TPTSSR/NN classifier | L2-norm | Unification of parameters obtained through proposed strategies is unfeasible |
MO-BoF [27] | Optimizing the BoF model by exploiting the manifold structure of the histogram spaces, semantic gap | HoG features | PCA | Spectral clustering | Not applied | X2 distance | Hand-crafted features, computationally expensive |
EODH-color SIFT [28] | Semantic gap | EODH features and color-SIFT descriptor | Not applied | k-means/weighted distribution/unweighted distribution | Not applied | L2-norm | Computationally expensive |
Modified VLAD [29] | Unequal contribution of residual vectors/insufficient burst patterns because of power-law normalization, semantic gap | VLAD/local coordinate system/residual normalization | PCA/product quantization | k-means | Not applied | L2-norm | Computationally expensive |
Attribute features+ Fisher vectors [30] | Enhancing image retrieval by incorporating attribute features along with FV, semantic gap | Fisher vectors/attributes/text | Random selection of attributes/selection with cross-validation/PCA/product quantization | Nearest neighbor search | SVM | L2-norm | Limited impact of semantic attributes on image retrieval |
Fisher kernel-GMM [31] | The high computational expense of Fisher vectors, semantic gap | SIFT/Fisher vectors | PCA/simple binarization, local sensitivity hashing/spectral hashing | GMM | Maximum likelihood estimation | Cosine similarity | Exhaustive database search |
Spatial L2 method [32] | Lack of spatial information in the BoF model, semantic gap | Dense SIFT | Not applied | k-means | RBF-NN/SVM/DBN | L2-norm | Hand-crafted equal weights for all triangular histograms |
RSHD method [33] | Semantic gap | Quantized RGB/local neighboring structure pattern | Not applied | Not applied | Not applied | Canberra distance | Hand-crafted feature area, non-robust performance |
WATH method [34] | Lack of spatial information in BoVW model, semantic gap | Dense SIFT, weighted triangular histograms | Not applied | Hard clustering based on k-means++ | SVM | L2-norm | Overfitting problem, computationally expensive |
Hybrid [35] | Scene categorization, semantic gap | GIST/HOG/LBP/dense SIFT | Not applied | k-means | SVM | X2 distance | A smaller overlap threshold results in overlapped annotations for objects within images. |