KEYWORDS: Education and training, Image classification, Deep learning, Machine learning, Data modeling, Ear, Cameras, Feature extraction, Visualization, Global Positioning System
Growth stage (GS) is an important crop growth metric commonly used in commercial farms. We focus on wheat and barley GS classification based on in-field proximal images using convolutional neural networks (ConvNets). For comparison purposes, use of a conventional machine learning algorithm was also investigated. The research includes extensive data collection of images of wheat and barley crops over a 3-year period. During data collection, videos were recorded during field walks at two camera views: downward looking and 45 deg angled. The resulting dataset contains 110,000 images of wheat and 106,000 of barley taken over 34 and 33 GS classes, respectively. Three methods were investigated as candidate technologies for the problem of GS classification. These methods were: (I) feature extraction and support vector machine, (II) ConvNet with learning from scratch, and (III) ConvNet with transfer learning. The methods were assessed for classification accuracy using test images taken (a) in fields on days imagined in the training data (i.e., seen field-days GS classification) and (b) in fields on days not imagined in the training data (i.e., unseen field-days principal GS classification). Of the three methods investigated, method III achieved the best accuracy for both classification tasks. The model achieved 97.3% and 97.5% GS classification accuracy for seen field-day test data for wheat and barley, respectively. The model also achieved accuracies of 93.5% and 92.2% for the principal GS classification task for wheat and barley, respectively. We provide a number of key research contributions: the collection curation and exposure of a unique GS labeled proximal image dataset of wheat and barley crops, GS classification, and principal GS classification of cereal crops using three different machine learning methods as well as a comprehensive evaluation and comparison of the obtained results.
Steganographic embedding is generally guided by two performance constraints at the encoder. Firstly, as is typical in the field of watermarking, all the transmission codewords must conform to an average power constraint. Secondly, for the embedding to be statistically undetectable (secure), it is required that the density of the watermarked signal must be equal to the density of the host signal. Assuming that this is not the case, statistical steganalysis will have a probability of detection error less than 1/2 and the communication may be terminated. Recent work has shown that some common watermarking algorithms can be modified such that both constraints are met. In particular, spread spectrum (SS) communication can be secured by a specific scaling of the host before embedding. Also, a side informed scheme called stochastic quantization index modulation (SQIM), maintains security with the use of an additive stochastic element during the embedding. In this work the performance of both techniques is analysed under the AWGN channel assumption. It will be seen that the robustness of both schemes is lessened by the steganographic constraints, when compared to the standard algorithms on which they are based. Specifically, the probability of decoding error in the SS technique increases when security is required, and the achievable rate of SQIM is shown to be lower than that of dither modulation (on which the scheme is based) for a finite alphabet size.
KEYWORDS: Quantization, Modulation, Optical spheres, Digital watermarking, Data hiding, Signal to noise ratio, Error analysis, Chemical elements, Matrices, Binary data
Spread-Transform Dither Modulation (STDM) is a side-informed data hiding method based on the quantization of a linear projection of the host signal. This projection affords a signal to noise ratio gain which is exploited by Dither Modulation (DM) in the projected domain. Similarly, it is possible to use to the same end the signal to noise ratio gain afforded by the so-called sphere-hardening effect on the norm of a vector. In this paper we describe the Sphere-hardening Dither Modulation (SHDM) data hiding method, which is based on the application of DM to the magnitude of a host signal vector, and we give an analysis of its characteristics. It shown that, in the same sense as STDM can be deemed to be the side-informed counterpart of additive spread spectrum (SS) with repetition coding, SHDM is the side-informed counterpart of multiplicative SS with repetition. Indeed, we demonstrate that SHDM performs similarly as STDM in front of additive independent distortions, but with the particularity that this is achieved through different quantization regions. The quantization hyperplanes which characterize STDM are replaced by quantization spheres in SHDM. The issue of securing SHDM is also studied.
In this paper we present a statistical analysis of a particular audio fingerprinting method proposed by Haitsma et al.1 Due to the excellent robustness and synchronisation properties of this particular fingerprinting method, we would like to examine its performance for varying values of the parameters involved in the computation and ascertain its capabilities. For this reason, we pursue a statistical model of the fingerprint (also known as a hash, message digest or label). Initially we follow the work of a previous attempt made by Doets and Lagendijk2-4 to obtain such a statistical model. By reformulating the representation of the fingerprint as a quadratic form, we present a model in which the parameters derived by Doets and Lagendijk may be obtained more easily. Furthermore, our model allows further insight into certain aspects of the behaviour of the fingerprinting algorithm not previously examined. Using our model, we then analyse the probability of error (Pe) of the hash. We identify two particular error scenarios and obtain an expression for the probability of error in each case. We present three methods of varying accuracy to approximate Pe following Gaussian noise addition to the signal of interest. We then analyse the probability of error following desynchronisation of the signal at the input of the hashing system and provide an approximation to Pe for different parameters of the algorithm under varying degrees of desynchronisation.
Digital steganography is the art of hiding information in multimedia
content, such that it remains perceptually and statistically unchanged. The detection of such covert communication is referred to as steganalysis. To date, steganalysis research has focused primarily on either, the extraction of features from a document that are sensitive to the embedding, or the inference of some statistical difference between marked and unmarked objects. In this work, we evaluate the statistical limits of such techniques by developing asymptotically optimal tests (Maximum Likelihood) for a number of side informed embedding schemes. The required probability density functions (pdf) are derived for Dither Modulation (DM) and Distortion-Compensated Dither Modulation (DC-DM/SCS) from an steganalyst's point of view. For both embedding techniques, the pdfs are derived in the presence and absence of a secret dither key. The resulting tests are then compared to a robust blind steganalytic test based on feature extraction. The performance of the tests is evaluated using an integral measure and receiver operating characteristic (ROC) curves.
In spread-spectrum watermarking, the watermarked document is obtained from the addition of an attenuated watermark signal to a cover multimedia document. A traditional strategy consists of optimising the detector for a given embedding function. In general, this leads to sub-optimal detection and much improvement can be obtained by exploiting side-information available at the embedder. In some prior art, the authors showed that for blind detection of low-power signals, maximum detection power is obtained to first order by setting the watermark signal to the gradient of the detector. In the first part of the paper, we develop this idea further and extend Costa's decoding theory to the problem of watermarking detection. In the second part, we propose a practical implementation of this work using non-linear detectors based on our family of polynomial functions which demonstrate some
improved performance of the technique. Finally, the robustness of our side-informed scheme is assessed by computer simulations on real audio signals in the presence of additive, multiplicative and coloured noise resulting from perceptual coding for a low watermark to content power ratio.
Compact representation of perceptually relevant parts of multimedia data, referred to as robust hashing or fingerprinting, is often used for efficient retrieval from databases and authentication. In previous work, we introduced a framework for robust hashing which improves the performance of any particular feature extraction method. The hash generation was achieved from a feature vector in three distinct stages, namely: quantization, bit assignment and application of the decoding stage of an error correcting code. Results were obtained for unidimensional quantization and bit assignment, on one code only. In this work, we provide a generalisation of those techniques to higher dimensions. Our framework is analysed under different conditions at each stage. For the quantization, we consider both the case where the codevectors are uniformly and nonuniformly distributed. For multidimensional quantizers, bit assignment to the resulting indexes is a non-trivial task and a number of techniques are evaluated. We show that judicious assignment of binary indices to the codevectors of the quantizer improves the performance of the hashing method. Finally, the robustness provided by a number of different channel codes is evaluated.
The vulnerability of quantization-based data hiding schemes to amplitude scaling has required the formulation of countermeasures to this relatively simple attack. Parameter estimation is one approach, where the applied scaling is estimated from the received signal at the decoder. As scaling of the watermarked signal creates a mismatch with respect to the quantization step assumed by the decoder, this estimate can be used to correct the mismatch prior to decoding. In this work we first review previous approaches utilizing parameter estimation as a means of combating the scaling attack on DC-DM. We then present a method for maximum likelihood estimation of the scaling factor for this quantization-based method. Using iteratively decodable codes in conjunction with DC-DM, the estimation method exploits the reliabilities provided by the near-optimal decoding process in order to iteratively refine the estimate of the applied scaling. By performing estimation in cooperation with the decoding process, the complexity of which is tackled using the expectation maximization algorithm, reliable estimation is possible at very low watermark-to-noise power ratios by using sufficiently low rate codes.
Steganalysis is the art of detecting and/or decoding secret messages embedded in multimedia contents. The topic
has received considerable attention in recent years due to the malicious use of multimedia documents for covert
communication. Steganalysis algorithms can be classified as either blind or non-blind depending on whether or
not the method assumes knowledge of the embedding algorithm. In general, blind methods involve the extraction
of a feature vector that is sensitive to embedding and is subsequently used to train a classifier. This classifier can
then be used to determine the presence of a stego-object, subject to an acceptable probability of false alarm. In
this work, the performance of three classifiers, namely Fisher linear discriminant (FLD), neural network (NN)
and support vector machines (SVM), is compared using a recently proposed feature extraction technique. It
is shown that the NN and SVM classifiers exhibit similar performance exceeding that of the FLD. However,
steganographers may be able to circumvent such steganalysis algorithms by preserving the statistical transparency
of the feature vector at the embedding. This motivates the use of classification algorithms based on the entire
document. Such a strategy is applied using SVM classification for DCT, FFT and DWT representations of an
image. The performance is compared to a feature extraction technique.
KEYWORDS: Digital watermarking, Sensors, Distortion, Signal detection, Forward error correction, Computer programming, Error control coding, Modulation, Signal to noise ratio, Reliability
The application of error correction coding to side-informed watermarking utilizing polynomial detectors is investigated.
The overall system is viewed as a code concatenation in which the outer code is a powerful channel
code and the inner code is a low rate repetition code. For the inner code we adopt our previously proposed
side-informed embedding scheme in which the watermark direction is set to the gradient of the detection function
in order to reduce the effect of host signal interference. Turbo codes are employed as the outer code due
to their near capacity performance. The overall rate of the concatenation is kept constant while parameters
of the constituent codes are varied. For the inner code, the degree of non-linearity of the detector along with
repetition rate is varied. For a given embedding and attack strength, we determine empirically the best rate
combinations for constituent codes. The performance of the scheme is evaluated in terms of bit error rate when
subjected to various attacks such as additive/multiplicative noise and scaling by a constant factor. We compare
the performance of the proposed scheme to the Spread Transform Scalar Costa Scheme using the same rates
when subjected to the same attacks.
In spread-spectrum watermarking, the watermarked document is obtained from the addition of an attenuated watermark signal to a
cover multimedia document. A traditional strategy consists of
optimizing the detector for a given embedding function. In
general, this leads to sub-optimal detection and much improvement
can be obtained by exploiting side-information available at the
embedder. In some prior art, the authors showed that for blind
detection of small signals, maximum detection power is obtained to
first order by setting the watermark signal to the gradient of the
detector. Recently, Malvar et al. improved the performance of direct-sequence spread-spectrum watermarking by using a signal dependent modulation. In the first part of the paper, we develop this idea further and extend Costa's decoding theory to the problem of watermarking detection. In the second part, we propose a practical implementation of this work using non-linear detectors based on our family of polynomial functions. We show some improved performance of the technique.
KEYWORDS: Digital watermarking, Sensors, Signal to noise ratio, Signal detection, Optical correlators, Stereolithography, Video, Distortion, Feature extraction, Receivers
This paper deals with some detection issues of watermark signals. We propose an easy way to implement an informed watermarking embedder whatever the detection function. This method shows that a linear detection function is not suitable for side information. This is the reason why we build a family of non-linear functions based on nth-order statistics. Used with a side-informed embedder, its performance is much better than the classical direct sequence spread spectrum method.
Second generation watermarking schemes differ from first generation symmetric schemes in that the detection process does not require the use of the same private key in both the embedder and the detector. An advantage of such schemes is that estimation of the watermark by an averaging attack is rendered impossible, so that the overall system is more secure. Almost all second generation schemes to date are also second order; that is, they are based on the computation of a quadratic form in the detector. Recently, Furon presented a unified description of second order schemes. Furon showed that O(m2) attacks were required to estimate the quadratic form where m is the spreading period. This presents a significant improvement over O(m) attacks required to estimate the watermarking key in symmetric schemes. In this work, the authors propose an audio watermarking scheme which employs an n-th order detection process. The scheme is based on a generalized differential modulation scheme and provides increased security over second order schemes. The cost of such increased security is a loss of efficiency, so that the watermark must be spread over more content. The paper presents an efficiency and security analysis for the third- and fourth-order scheme.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.