Perceptual Hashing With Deep and Texture Features

Image hashing is a useful technique for many multimedia systems, such as image authentication, image copy detection, tampering detection, and image quality assessment (IQA). However, most of the image hashing schemes do not make desirable performance of IQA. To tackle this, a new hashing scheme with deep and texture features is proposed for reduced-reference (RR)-IQA. In the proposed hashing, deep features are calculated from the discrete cosine transform coefficients of the three-order tensor stacked by the feature maps generated by the pretrained ResNet18. Texture features are extracted by the gray-level co-occurrence matrix in the nonsubsampled shearlet transform domain. Hash is determined by combining the quantized versions of the deep and texture features. Extensive experiments performed on open datasets indicate that the proposed perceptual hashing is superior to some baseline schemes in the performances of RR-IQA and classification.