This paper deals with smartphone model identification using image features. 64 image features are grouped into colour, wavelet and image and quality features are extracted from high resolution smart phone images. A multiclass support vector machine is used as a classifier. 1800 images were captured using 10 different smartphones and devices. Wavelet features and image quality metrics are shown to contain vital device and model information compared to colour features. Compared to colour features, quality and wavelet features are more sensitive to simple image modifications.
In investigating child pornography, blackmail and bullying there is a need to prove the originality of digital images and their source to help during investigations and this can provide evidence as well. Methods ranging from gathering information from the JPEG or EXIF header have been used to predict the source device. Images nowadays are captured using smartphones as they are portable and affordable, and images can be shared instantly. Crimes committed by images captured by smartphones is now recognised by the law. They include activities like web browsing, online social networking, email, MMS and video chat. Previous research on source devices identification using digital images focused on digital cameras, printers and scanners leaving little on cameras embedded in smartphones.
The paper uses the image feature based on Kharrazi et al approach to identify the source and model of an image. This includes evaluating the performance of the image feature category, investigating resizing and cropping and designing a new smartphone image database to evaluate existing techniques.
The second section of the paper describes image features and the classification model trained for source device identification. Section three explains image data collection and different image sizes. Section four presents experimental findings and discussion and conclusion in section five.
Colour, wavelet and image quality metrics are used to identify modern smartphone device linked to smartphone. This approach analysis the relationship between the image and its device. This is because different components and settings used have effect on the image produced and this can lead to source device. The denoised image for computing IQM features was obtained by adding Gaussian noise of zero mean and unit variance and a filter is used to remove any noise. The Haar wavelet transform was used to calculate the wavelet features. In this case the sum of coefficients in the three-corresponding high frequency wavelet sub bands in extracting relevant features were used and the results of these features were similar to using features extracted from each sub band and the total number of image features was reduced to 40 from 64.
Extracted features were fed to a SVM 11 to train a model to be used to classify new images. An SVM machine was used as a classifier. Images used in this experiment consisted of 600 of origin size ranging from buildings, sky and natural environment. 10 modern smartphones were used to capture images including two HTC devices of the same model.
A set of 600 images were created by resizing the original images to 800 by 600 pixels, the resized images of Nokia N8 were 800 by 450 pixels and a further set of 600 images were created by randomly cropping out a 700 by 700 region from each original image. This was done to evaluate the accuracy of smartphone identification using a cropped or resized version of the original image the data base had 1800 images. For performance of image features, several smartphone model identifiers were conducted. The effect of training and testing sample sizes on identification accuracy was evaluated. Percentages were split into three parts of 10%, 50%, and 90% of images were randomly selected for training and the remainder used for testing were used in each experiment. Original size images, used for training and testing showed an overall accuracy rates of 62%, 64% and 76% for using only colour features in 10%, 50% and 90% splits respectively. Using IQM only features, achieved accuracy rates of 77%, 88%, 96% respectively. Wavelet features attained accuracy rates of 82%, 88%, and 95% for the three percentage splits respectively. By using all image features the accuracy rates were 80%, 90%, and 97.5%. This shows an improvement in accuracy over all individual feature types.
Resized images of smartphone model results show a total identification correctness of 92% a slightly decrease in performance compared to the 97.5% accuracy attained on original size images when all features were used. Using half of the images in training and the other half in testing, combining all features resulted in 73% accuracy, IQM features achieved an accuracy of 72%, wavelet features of 70% whilst colour features reached an accuracy level of 65%. By using 10% of images from each device in training and 90% in testing, combined features achieved 63% accuracy when compared to 80% for the same situation with original images. IQM features attained 65% exactness, wavelet features 60% accuracy and colour features a correctness of 62% respectively.
On cropped images, the results have an accuracy of 93%, that shows a momentous drop in accuracy when associating with original images in training and testing. This accuracy is obtained a result of engaging all features and using 90% of images in training. Correspondingly, IQM features and wavelet features were substantial in this experiment reporting 85% and 84% accuracy respectively. However, these features were the most delicate, sensitive to image cropping. Colour features had 71% accuracy and remained relatively unaffected due to image cropping.
For varied images, some experiments were conducted using the different image sizes to train original images and test resized and cropped images for the model. This was done in direction of further testing the robustness of the model against modified images. Individual device and model accuracy rates is shown in figure 4. By using 90% of original images in training and testing against 90% cropped images, overall accuracy is 76.5%. The uppermost distinct accuracy of 91.85% comes from the Samsung Omnia 7 device and the lowest being iPhone 4 with a lower percentage of 48.33.
The potential of using previously proposed image features to classify images from modern smartphones was addressed. The use of colour, image quality and wavelet feature to classify the smartphone model for an unknown image using a build in smartphone image database is studied. This shows that it is manageable to trace the source of an image source without any knowledge of the model or device. It was found that quality and wavelet feature are the most significant image features.
Improvement in identification accuracy can be amplified by increasing the number of training images. When the unknown image is a modified version of an original image identification accuracy can drop significantly. In the near future there is hope to investigate the use of genetic algorithms for feature selection and feature weighting at feature level and decision level fusion.