No abstract

Many mobile applications running on smartphones and wearable devices would potentially benefit from the accuracy and scalability of deep CNN-based machine learning algorithms. However, performance and energy consumption limitations make the execution of such computationally intensive algorithms on mobile devices prohibitive. We present a GPUaccelerated library, dubbed CNNdroid [1], for execution of trained deep CNNs on Android-based mobile devices. Empirical evaluations show that CNNdroid achieves up to 60X speedup and 130X energy saving on current mobile devices. The CNNdroid open source library is available for download at https://github.com/ENCP/CNNdroid

No abstract

No abstract

Abstract-This paper investigates the use of Structural Similaritys (SSIM) index on the minimized side effect to image watermarking. For fast implementation and more compatibility with the standard DCT based codecs, watermark insertion is carried out on the DCT coefficients and hence a SSIM model for DCT based watermarking is developed. For faster implementation, the SSIM index is maximized over independent 4x4 non-overlapped blocks but the disparity between the adjacent blocks reduces the overall image quality. This problem is resolved through optimization of overlapped blocks, but, the higher image quality is achieved at a cost of high computational complexity. To reduce the computational complexity while preserving the good quality, optimization of semi-overlapped blocks is introduced. We show that while SSIM-based optimization over overlapped blocks has as high as 64 times the complexity of the 4x4 nonoverlapped method, with semi-overlapped optimization the high quality of overlapped method is preserved only at a cost of less than 8 times the non-overlapped method.Index Terms-discrete cosine transform, semi-overlapped blocks, structural similarity, subjective quality, watermarking. I. INTRODUCTIONDigital watermarking is a technique of inserting a signature in a cover signal, such that the Human Visual System (HVS) is not able to recognize it, while, the inserted signature can be extracted using watermark detectors, or at least the existence of the watermark can be confirmed. Due to the widespread use of multimedia products, the watermarking technology is used in various applications. In the watermarking design process, there is a trade-off between the three parameters of: robustness, imperceptibility and data payload. Robustness is the ability of watermark restoration after attacks. Imperceptibility, which is the main subject of this paper, addresses the quality degradation of watermarked image. Finally, after determining the acceptable requirements for robustness and imperceptibility; data-payload or storage capacity is defined as the number of bits that can be inserted into certain parts of the digital product. Depending on considered application of watermarking, a reasonable compromise among the above parameters should be maintained.Watermarking systems try to insert data in such a way that the signal quality is preserved. This issue is crucial in some applications such as Medical Imaging [6]. However, controlling the side effect of inserting data on signal quality requires an objective metric for optimization. It is normally believed Mean Squared Error (MSE) objective quality is not a good measure for such a case, since it does not well correlate to the subjective quality measures [7]. In the past decade the Video Quality Experts Group (VQEG) [8] have recommended a set of quality metrics, with or without the reference to non-distorted images or even to partial references, where the objective image quality well correlates to the subjective measures. However, these tools are most suited for measurements of ...