|
| Method | Dataset (Image size) | Preprocessing | Methods | Result |
|
| LDA_DCT [18] | INRIA holiday (512 × 512)
ImageNet (128 × 128)
Caltech-256 (128 × 128) | Image block scanning DCT value | Generative-based; generate corresponding hash sequences based on image DCT values; both communication parties use the same feature | Low capacity and low robustness. It has the security risk of transmitting the same image repeatedly and amount of additional information |
|
| DenseNet_DWT [19] | INRIA holiday (512 × 512)
ImageNet (128 × 128)
Caltech-256 (128 × 128) | Image block scanning DWT value | Generative-based; generate corresponding hash sequences based on image DWT values; both communication parties use the same feature | Low capacity and low robustness. It has the security risk of transmitting the same image repeatedly and amount of additional information |
|
| DCMH-CNN [32] | INRIA holiday (512 × 512)
ImageNet (128 × 128)
Caltech-256 (128 × 128) | Extract high-dimensional features for sorting | Mapping-based; extract high-dimensional features using a network and use K-means clustering to form a CID; sort the high-dimensional features and construct mapping rules; both communication parties use the same feature | High capacity and strong robustness. But it has the security risk of transmitting the same image repeatedly and additional information |
|
| Our | INRIA holiday (128 × 128, 256 × 256, 512 × 512)
ImageNet (128 × 128, 256 × 256, 512 × 512)
Caltech-256 (128 × 128, 256 × 256, 512 × 512) | Calculate average pixel value for sorting | Mapping-based; generate a random sequence to form a CID; construct mapping rules; only the receiver uses SURF features | High capacity and strongest robustness. There is no security risk of repeatedly transmitting the same image and additional information |
|
