Interpol review of imaging and video 2016-2019

Abstract

This review paper covers the forensic-relevant literature in imaging and video analysis from 2016 to 2019 as a part of the 19th Interpol International Forensic Science Managers Symposium. The review papers are also available at the Interpol website at: https://www.interpol.int/content/download/14458/file/Interpol%20Review%20Papers%202019.pdf.

Keywords: Image analysis; Imaging; Interpol; Video.

Fig. 1

Example of image manipulation that appeared in press in July 2008. (a) The forged image displaying four missiles. Only three of them are real, two different sections (encircled in red and orange, respectively) are replicates of other image sections (b)The original image showing only three missiles [6]. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

Fig. 2

Left: A regular 3-layer (2 hidden and 1 output) neural network with one dimensional layers. Right: a convolutional neural network with the neurons arranged in three dimensions. Every layer transforms the 3D input volume to a 3D output volume of neuron activation’s. The red input layer holds the image, with its width and height equal to the spatial dimensions of the image, and a depth of 3 (the Colour channels Red, Green, Blue) [11]. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

Fig. 3

CNN architecture as proposed by Kim and Lee [15] consisting of 1 HPF, 2 convolutional layers, 2 max pooling layers, and 2 fully connected layers with softmax function for classification. The networks input dimension is a 256 × 256 sized grayscale image.

Fig. 4

CNN architecture as proposed by Bayar and Stamm [9] consisting of 1 constrained convolutional layer, 2 convolutional layers, 2 max pooling layers, and 3 fully connected layers with softmax function for classification. The networks input dimension is a 227 × 227 sized grayscale image.

Fig. 5

Baseline CNN architecture as proposed by Bayar and Stamm [10] consisting of 1 constrained convolutional layer, 3 convolutional layers with PReLU activation functions, 2 max pooling layers and 1 average pooling layer, and 3 fully connected layers with softmax function for classification. The networks input dimension is a 256 × 256 green layer image. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

Fig. 6

CNN architecture as proposed by Bayar and Stamm [12] consisting of 1 constrained convolutional layer, 4 convolutional layers, 3 max pooling layers, 1 average pooling layer and 3 fully connected layers with softmax function/extremely randomised tree for classification. The networks input dimension is a 256 × 256 sized grayscale image [12].

Fig. 7

The constrained weights of a 5 × 5 isotropic filter as proposed by Chen et al. [18].

Fig. 8

Rotation invariant CNN architecture as proposed by Chen et al. [18] consisting of one constrained isotropic conv layer, 4 isotropic conv layers, 4 pooling layers and three fully connected layers with softmax function.

Fig. 9

CNN architecture as proposed by Chen et al. [7] with 8 layer groups, the first being the isotropic convolutional layer and the second to eight the traditional convolutional layer. Additionally, it has 3 transition layers 3 max pooling layers and 1 fully connected layer.

Fig. 10

CNN architecture as proposed by Boroumand and Fridrich [28] consisting of 8 convolutional layers with ReLU activation and batch normalisation, 7 average pooling layers and one fully connected layers with softmax function.

Fig. 11

Graphical representation of proposed approach by Meyer et al. [32] (a) transfer learning and (b) multitask learning, both using an example share depth up to first fully connected layer (fc1).

Fig. 12

[2].

Fig. 13

CNN architecture as proposed by Yu et al. The networks input dimension is a 128 × 128 grayscale image (16.384 neurons). The architecture 5 convolutional layers, two pooling layers and one fully connected layer connected to the output layer through a soft max function [9].