DE202022104365U1 - A robust color image hashing system for image authentication - Google Patents

Abstract

A neuro-evolutionary detection system (100) for early detection of lung cancer, the system (100) comprising:
a plurality of receivers (102) for receiving an input image and a hash value, the received image being encapsulated with a blind rotational scaling transform (RST);
a geometric transform correction module (104) connected to the plurality of receivers (102) for eliminating features of the blind RST transform by using inherent properties of the geometric transform and evaluating the rotation condition of the received image, wherein the features of the blind RST transformation are eliminated by computing the coordinates of non-zero value boundary pixels to obtain an interpolated image;
an image authentication module (106) connected to the geometric transformation module (104) for generating a hash code by an L2 regularization constraint (L2RC) applied to activation of a termination block in an encoder (106a), an estimation module (108), embodied in the image authentication module (106), estimates a correlation coefficient between the received hash and the generated hash, wherein if the correlation coefficient is less than a first threshold, the interpolated image is either a manipulated image or a different image ; and
a tamper detection and localization module (110) coupled to the image segmentation module (106) for classifying the image as a tampered image or another image identification portion of the tampered images, wherein the tampered image is mapped to the generated hash code to provide a generate an image map and locate the manipulated image.

Description

FIELD OF THE INVENTION

The present invention relates to a field of image authentication systems. More particularly, the present invention relates to robust color image hashing using convolutional stacked denoising auto-encoders for image authentication.

BACKGROUND OF THE INVENTION

With the development of sophisticated image editing programs, manipulating or forging image content is quite easy. Tampering detection is very important to determine the validity of an image. This problem is best solved using perception-based image hashing techniques. These techniques extract salient features from an image to create a hash.

Conventional hash methods such as MD5 (Message Digest 5) or SHA-256 generate highly sensitive hash codes for data. Any small change will result in different hash codes. This behavior proves detrimental to digital images (original or transmitted images in image communications) that undergo CPOs such as compression, filtering, and scaling to keep the visual information close to the original. The goal of the perceptual hashing technique is to generate hash codes that only consider changes in the visual domain. There is a high degree of correlation between the hash codes of the original image and the counterparts processed by CPOs; H. Robustness, while the correlation between the hash codes of the original image and the distorted content image should be lower, i.e. H. discrimination. The performance of perceptual hashing techniques is measured by their robustness to CPOs and their sensitivity to malicious operations such as removing or adding content.

In general, hashing methods fall into two categories, location-based hashing (LSH) and learning-based hashing (LBH). LSH uses functions that are independent of the given data to generate hash codes, while LBH techniques learn hash functions based on the input data. With LSH, the retention of the semantic similarity, which is a basic requirement for the use of image processing systems, is not guaranteed. To overcome these problems, it is recommended to use learning-based hashing techniques.

The LBH techniques use input data to map similar images into hash codes with a high degree of correlation. These methods are scalable and preserve a semantic relationship between the input images. Therefore, the LBH technique based on Convolutional Stacked Denoising Auto-Encoders (CSDAEs) for image authentication was proposed in this work.

The Auto-Encoders (AEs) are able to learn the salient features of blank data (or images) and provide an efficient representation (short-length array) of the input data. This representation ensures minimal reconstruction losses during decoding. Most AEs are fully connected feedforward networks with a regularization constraint (RC) in the middle layer that produces a hash code. RC enables the model to learn data distributions instead of copying input to output. Convolutional neural networks (CNNs) are robust to overfitting due to their weight-sharing properties. CNNs work well for data with local spatial coherence, e.g. B. with pictures.

The first known image hashing was presented by Schneider and Chang, followed by the work of R. Venkatesan et al. After that, researchers paid much attention to image hashing to authenticate content.

According to another researcher, Tang et al. Introduced ring partitions and NMF (non-negative matrix factorization) to generate hash codes that are robust to rotation and have good discrimination ability, but are sensitive to some geometric operations.

According to another researcher, Davarzani et al. introduced singular value decomposition (SVD) and centrally symmetric local binary patterns (CSLBP) to create a perceptual image hash (PIH). len. This method can detect tampering if the tampered area is at least 10% of the reference image.

According to another researcher, Ding et al. proposed a DCT-based image authentication. This method cannot detect manipulations in the harmony.

Another researcher, Xue et al., developed a hashing method using SIFT and LBP. This method is limited in case of geometric operations.

Another researcher, Pun et al., proposed a feature point selection-based image hash technique for content authentication and manipulation localization. This method is limited when manipulations and composite RST occur simultaneously.

According to another researcher, Ouyang et al. a PIH technique that uses QZMs (quaternion-Zernike moments) for global features and local SIFT features, but is affected by luminance and geometric perturbations. The prior art used ring partitions and factorization of the non-negative gradient matrix, which rendered the system rotation insensitive. Local features are then derived from the salient regions to locate the fake regions in an image.

Another researcher, Karsh et al., proposed a hashing technique using DWT-SVD. However, these methods failed on color images, but offered hash robustness against high levels of rotation. The hashing methods are used for image authentication, but when composite RST and malicious tampering occur simultaneously, tampering cannot be detected.

Robustness to geometric changes is a major issue with the above image authentication methods. The effects of geometric distortions in image authentication have been mitigated using image alignment techniques.

According to another researcher, Battia et al. presented the voting procedure for restoring the original image. Image recovery performance is better, but requires longer length of hash, i.e. H. 1000 digits.

Another researcher, Yan et al., presented an approach using a quaternion Fourier-Mellin transform for image restoration, but this method is limited in case of manipulations.

According to another researcher, Karsh et al. introduced a geometric correction approach based on the inherent properties of the geometric transformation. This method requires a very small hash length but is limited at negative spin degrees. Autocoders (AEs) have been part of the historical landscape for decades. Traditionally, AEs are mainly used for feature extraction and learning.

According to another researcher, Makhzani et al. Autocoders with aggressive parsimony constraints. With this approach, only nb activations of each trait card are kept while the rest are made zero.

The previous work mentioned above either focused on making the hash codes robust or on locating manipulated regions.

Therefore, there is a need to develop an apparatus for training a single stacked convolutional denoising autocoder to generate hash codes that are robust against various geometric attacks while locating the manipulated regions.

The technical advance disclosed by the present invention overcomes the limitations and disadvantages of existing and conventional systems and methods.

SUMMARY OF THE INVENTION

The present invention relates generally to a neuro-evolutionary detection system for the early detection of lung cancer.

An aim of the present invention is to propose an image hashing method based on CSDAEs during training,

Another aim of the present invention is to propose a geometric correction approach, in particular for composite RST, to detect manipulated areas and

Another object of the present invention is to detect image manipulations.

• eine Vielzahl von Empfängern zum Empfangen eines Eingangsbildes und eines Hashwertes, wobei das empfangene Bild mit einer blinden Rotationsskalierungstransformation (RST) gekapselt ist;
• ein geometrisches Transformationskorrekturmodul, das mit der Vielzahl von Empfängern verbunden ist, um Merkmale der blinden RST-Transformation zu eliminieren, indem inhärente Eigenschaften der geometrischen Transformation verwendet werden und die Rotationsbedingung des empfangenen Bildes ausgewertet wird, wobei die Merkmale der blinden RST-Transformation eliminiert werden, indem die Koordinaten von Grenzpixeln mit Werten ungleich Null berechnet werden, um ein interpoliertes Bild zu erhalten;
• ein Bildauthentifizierungsmodul, das mit dem geometrischen Transformationsmodul verbunden ist, zum Erzeugen eines Hash-Codes durch eine L2-Regularisierungsbeschränkung (L2RC), die auf die Aktivierung eines Abschlussblocks in einem Codierer angewendet wird, wobei ein in dem Bildauthentifizierungsmodul enthaltenes Schätzmodul einen Korrelationskoeffizienten zwischen dem empfangenen Hash und dem erzeugten Hash schätzt, wobei, wenn der Korrelationskoeffizient kleiner als ein erster Schwellenwert ist, das interpolierte Bild entweder ein manipuliertes Bild oder ein anderes Bild ist; und
• ein Modul zur Erkennung und Lokalisierung von Manipulationen, das mit dem Bildsegmentierungsmodul verbunden ist, um das Bild als manipuliertes Bild oder einen anderen Bildidentifizierungsbereich der manipulierten Bilder zu klassifizieren, wobei das manipulierte Bild dem erzeugten Hash-Code zugeordnet wird, um eine Bildkarte zu erzeugen und das manipulierte Bild zu lokalisieren.

In one embodiment, a neuro-evolutionary detection device for the early detection of lung cancer, the device comprising:

• a plurality of receivers for receiving an input image and a hash value, the received image encapsulated with a blind rotational scaling transform (RST);
• a geometric transform correction module connected to the plurality of receivers to eliminate blind RST transform features by using inherent properties of the geometric transform and evaluating the rotation condition of the received image, thereby eliminating the blind RST transform features , by calculating the coordinates of non-zero boundary pixels to obtain an interpolated image;
• an image authentication module connected to the geometric transformation module for generating a hash code by an L2 regularization constraint (L2RC) applied to the activation of a termination block in an encoder, wherein an estimation module included in the image authentication module calculates a correlation coefficient between the received hash and the generated hash, wherein if the correlation coefficient is less than a first threshold, the interpolated image is either a manipulated image or a different image; and
• a tampering detection and localization module connected to the image segmentation module to classify the image as a tampered image or another image identification portion of the tampered images, wherein the tampered image is associated with the generated hash code to generate an image map and locate the manipulated image.

In one embodiment, the boundary pixels include the top pixel, the rightmost pixel, the leftmost pixel, and the bottom pixel.

In one embodiment, the geometric transformation correction module evaluates the direction of rotation, i. i.e., either clockwise or counterclockwise, and adjusts the rotated image where, if the received image is rotated counterclockwise by θ degrees, the top non-zero coordinate point is further to the right than the bottom point, regardless from its translation degree, and vice versa for a clockwise rotated image up to -θ degrees, where if the value of θ is substantially equal to θ' degrees, the received image is considered to be a rotated image counter-rotated by θ' degrees is.

In one embodiment, a binary interpolation module calculates a region of interest by manipulating the image to one dimension of the received image while adjusting the rotation angle.

In one embodiment, the encoder is a convolutional stacked denoising auto-encoder (CSDAE) having a plurality of blocks to generate the hash image, the CSDAE being a forward neural network, each of the plurality of blocks consisting of a convolution layer followed by stack normalization and a max pooling layer.

In one embodiment, the CSDAE hierarchically maps received interpolated images into a latent space that is re-mapped to the dimension of the received image, with the L2 RC configured to: Remove irrelevant components from the hash image by using a chooses smallest combination to solve learning problems; and suppressing the effects of static noise on the targets by preventing overfitting.

In one embodiment, the corrupted area is located by multiplying the corrupted area by the corrupted image, where if the corrupted area is less than 30%, the received image is judged as "corrupted image" and otherwise as "different image".

In one embodiment, the first threshold is 0.98 and a second temperament localization threshold is 0.5.

In one embodiment, each convolutional layer is initialized with a plurality of filters, except for the last layer, each having a kernel size of (3,3), where the CSDAE is trained layer by layer and then fine-tuned as a whole.

In one embodiment, a decoder block layer includes a 2D convolution layer and an up-sampling layer, followed by a batch norm layer.

In order to further clarify the advantages and features of the present invention, a more detailed description of the invention will be made by reference to specific embodiments thereof illustrated in the accompanying figures. It is understood that these figures show only typical embodiments of the invention and therefore should not be considered as limiting its scope. The invention will be described and illustrated with additional specificity and detail with the accompanying figures.

character list

• 1 ein Blockdiagramm einer neuro-evolutionären Erkennungsvorrichtung zur Früherkennung von Lungenkrebs zeigt, und
• 2 eine grafische Darstellung der TPR-Raten für verschiedene Rotationsgrade zeigt.

These and other features, aspects and advantages of the present invention will be better understood when the following detailed description is read with reference to the accompanying figures, in which like characters represent like parts throughout the figures, wherein:

• 1 shows a block diagram of a neuro-evolutionary detection device for the early detection of lung cancer, and
• 2 shows a graphical representation of the TPR rates for different degrees of rotation.

Those skilled in the art will understand that the elements in the figures are presented for simplicity and are not necessarily drawn to scale. For example, the flow charts illustrate the method of key steps to enhance understanding of aspects of the present disclosure. In addition, one or more components of the device may be represented in the figures by conventional symbols, and the figures only show the specific details relevant to understanding the embodiments of the present disclosure, not to encircle the figures with details to overload, which are easily recognizable to those skilled in the art familiar with the present description.

DETAILED DESCRIPTION

For the purposes of promoting an understanding of the invention, reference will now be made to the embodiment illustrated in the figures and specific language will be used to describe the same. It should be understood, however, that no limitation on the scope of the invention is intended, and such alterations and further modifications to the illustrated system and such further applications of the principles of the invention set forth therein are contemplated as would occur to those skilled in the art invention would normally come to mind.

It will be understood by those skilled in the art that the foregoing general description and the following detailed description are exemplary and explanatory of the invention and are not intended to be limiting.

When this specification refers to "an aspect," "another aspect," or the like, it means that a particular feature, structure, or characteristic described in connection with the embodiment is present in at least one embodiment of the present invention. Therefore, the phrases "in one embodiment," "in another embodiment," and similar phrases throughout this specification may or may not all refer to the same embodiment.

The terms "comprises,""including," or other variations thereof are intended to cover non-exclusive inclusion, such that a method or method that includes a list of steps includes not only those steps, but may also include other steps that are not expressly stated or pertaining to any such process or method. Likewise, any device or subsystem or element or structure or component preceded by "comprises...a" does not, without further limitation, exclude the existence of other devices or other subsystem or other element or other structure or other component or additional device or additional subsystems or additional elements or additional structures or additional components.

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one skilled in the art to which this invention pertains. The system, methods, and examples provided herein are for purposes of illustration only and are not intended to be limiting.

Embodiments of the present invention are described in detail below with reference to the attached figures.

1 shows a block diagram of a neuro-evolutionary detection system (100) for early detection of lung cancer, the system (100) comprising: a plurality of receivers (102), a module for correcting geometric transformations (104), an image authentication module (106), a encoder (106a), a decoder (106b), an estimation module (108) and a module for detecting and locating manipulations (110).

The plurality of receivers (102) for receiving an input image and a hash value, the received image encapsulated with a blind rotational scaling transform (RST).

The geometric transformation correction module (104) is connected to the plurality of receivers (102) to eliminate features of the blind RST transformation by using inherent properties of the geometric transformation and evaluating the rotation condition of the received image, where the characteristics of the blind RST transform can be eliminated by computing the coordinates of boundary pixels with non-zero values to obtain an interpolated image. The edge pixels include the top pixel, the rightmost pixel, the leftmost pixel, and the bottom pixel.

The geometric transformation correction module evaluates the rotation direction, i.e., either clockwise or counterclockwise, and adjusts the rotated image, where if the received image is rotated counterclockwise by up to θ degrees, the top non- zero value coordinate point is further to the right than the bottom point, regardless of its degree of translation, and vice versa for a clockwise rotated image by up to -θ degrees, where if the value of θ is substantially equal to θ' degrees, the received image as a rotated image is viewed that is counter-rotated by the θ' degree.

andernfalls wird der Drehwinkel wie folgt berechnet. $$\mathrm{\theta }=-\left(90-\text{arctan arctan}\left(\frac{\Delta y}{\Delta x}\right)\right);$$

$$\mathrm{\theta }\text{'}=-\left(90-\text{arctan arctan}\frac{\Delta y\text{'}}{\Delta x\text{'}}\right));$$

wobei Δy=Y_r-Y_b, Δx=X_r-X_b, Where Δy'=Y_t-Y_l, und
Δx'=X_t-X_l, worin (Xr,Yr), (Xt,Yt), und (Xl; Yl) die Koordinaten der Pixelwerte sind, die nicht Null sind, für den rechten, oberen bzw. unteren, äußersten linken Punkt. Um die Skalierungs- und Translationsfaktoren zu berücksichtigen, werden die Koordinaten der obersten, untersten, rechten und linken Pixelwerte, die nicht Null sind, berechnet. Anhand dieser Informationen wird dann der interessierende Bereich beschnitten und durch bilineare Interpolation an die Abmessungen des Abfragebildes angepasst.When an image is rotated counterclockwise up to 45 degrees, the top non-zero coordinate point is further to the right than its bottom point, regardless of its degree of translation, ie Xt>Xb, and vice versa for a clockwise rotated image up to - 45 degrees. If the image is judged to be rotated counterclockwise, the rotation angle is calculated as follows: $$\mathrm{\theta }=\text{arctan}\frac{\Delta y}{\Delta x\text{'}}\mathrm{\theta }\text{'}=\text{arctan}\frac{\Delta y\text{'}}{\Delta x\text{'}}$$

otherwise the angle of rotation is calculated as follows. $$\mathrm{\theta }=-\left(90-\text{arctan arctan}\left(\frac{\Delta y}{\Delta x}\right)\right);$$

$$\mathrm{\theta }\text{'}=-\left(90-\text{arctan arctan}\frac{\Delta y\text{'}}{\Delta x\text{'}}\right));$$

where Δy=Y _r -Y _b , Δx=X _r -X _b , Where Δy'=Y _t -Y _l , and
Δx'=X _t -X _l , where (Xr,Yr), (Xt,Yt), and (Xl;Yl) are the coordinates of the non-zero pixel values for the right, top, and bottom, leftmost, respectively Point. To account for the scaling and translation factors, the coordinates of the top, bottom, right, and left non-zero pixel values are calculated. Using this information, the region of interest is then cropped and fitted to the dimensions of the query image using bilinear interpolation.

The binary interpolation module (104) calculates a region of interest by manipulating the image to one dimension of the received image while adjusting the rotation angle.

The image authentication module (106) is connected to the geometric transformation module (104) to generate a hash code by an L2 regularization constraint (L2RC) applied upon activation of a termination block in an encoder (106a), using a in the image authentication module (106) included estimation module (108) estimates a correlation coefficient between the received hash and the generated hash, wherein if the correlation coefficient is less than a first threshold, the interpolated image is either a manipulated image or a different image.

The encoder (106a) is a convolutional stacked denoising auto-encoder (CSDAE) having a plurality of blocks to generate the hash image, the CSDAE being a forward neural network, each of the plurality of blocks consisting of a convolution layer followed by a stack normalization and a max pooling layer. The CSDAE hierarchically maps received interpolated images into a latent space that is re-mapped to the dimension of the received image, with the L2 RC configured to: Remove irrelevant components from the hash image by choosing a smallest combination to solve learning problems; and suppresses effects of static noise on targets by preventing overfitting.

Each convolutional layer is initialized with a multitude of filters, except for the last layer, each having a kernel size of (3,3), where the CSDAE is trained layer by layer and then fine-tuned as a whole.

The block layer of the decoder (106b) consists of a 2D convolution layer and an up-sampling layer, followed by a batch norm layer.

The tampering detection and localization module (110) is connected to the image segmentation module (106) to classify the image as a tampered image or another image identification portion of the tampered images, wherein the tampered image is mapped to the generated hash code, to generate an image map and locate the manipulated image.

The corrupted area is located by multiplying the corrupted area by the corrupted image, where if the corrupted area is less than 30%, the received image is judged as "corrupted image" and otherwise as "different image".

The first threshold is 0.98 and a second temperament localization threshold is 0.5.

• T (x,y)= {Ĩ(x,y) - I̅(x,y)} ≥ δ₂ wobei δ2 ein Schwellenwert für die Manipulationslokalisierung ist und x,y Pixelkoordinaten sind. Schließlich wird T mit dem manipulierten Bild multipliziert, um die wahrscheinlich manipulierten Bereiche des empfangenen Bildes anzuzeigen.

Schichten Encoder-Blockschicht Decoder-Blockschicht Kernel_Größe 3*3 3*3 Anzah_der_Karten 16 16 Pool Größe 2*2 2*2 Pool_Typ max max The manipulated regions (T) are located as follows:

• T (x,y)= {Ĩ(x,y) - I̅(x,y)} ≥ δ ₂ where δ2 is a threshold for manipulation localization and x,y are pixel coordinates. Finally, T is multiplied by the manipulated image to indicate the likely manipulated areas of the received image.

layers encoder block layer Decoder Block Layer kernel_size 3*3 3*3 number_of_cards 16 16 pool size 2*2 2*2 pool_type Max Max

The optimal parameter values for the proposed system are chosen as follows: P × Q × R = 128 × 128 × 3, and the first threshold δ1 is 0.98, δ2 is 0.5. The Tensor Flow Deep Learning Framework was used to implement the CSDAE. All training and test images are scaled to [0,1] before being fed into the CSDAE. During the training phase, the CSDAE receives images that have undergone CPOs and corresponding reference images as output. Each convolutional layer is initialized with 64 filters, except for the last layer, each of which has a kernel size of (3,3). The CSDAE is trained in layers and then fine-tuned as a whole. The training is performed for a total of 43,328 images. The source images are from the USC SIPI database and are trained for 100 epochs at a batch size of 400, where 320 is the training batch and the 80 samples are the validation batch or the testing batch. The USC-SIPI database consists of different volumes such as textures and aerial photographs using different characters of the images. The size of the images in each band varies from 256x256 pixels to 1024x1024 pixels. The proposed system will be evaluated and tested in terms of blind RST correction, robustness and discrimination performance, and tampering detection and localization.

The system is trained on input images of dimensions (512 X 512 X 3) and (128 X 128 X 3) for 100 epochs. The accuracy and loss curves appear to be oscillating for the model trained for 512 X 512 X 3 images, while they are relatively smooth for the model trained for 128 X 128 X 3 images. The model for 512 inputs does not converge, i.e. "TA 512 input."

It is concluded that the accuracy performance is better for a 128 x 128 x 3-dimensional input image than for a 512 x 512 x 3-dimensional one.

2 shows a graphical representation of the TPR rates for different degrees of rotation. It can be observed that the CSDAE is robust to rotations from -5° to +5°. This gives the proposed RST correction an error margin of (-5,+5) degrees. Combining this with the proposed RST correction, the CSDAE produces hash codes that are robust for the entire range of (-45,+45) degrees.

According to one embodiment, the robustness and discrimination of the proposed image hashing is evaluated. The experiment is performed on 10,094 similar image pairs generated from 103 reference images by applying content-preserving operations (CPOs). The system is also evaluated with 19,900 dissimilar images generated from 200 images as 200 (200-1)/2. The images come from different sources, namely 100 images from the reference database, 50 images from the Nikon D3200 and 50 images from the internet. The experiment is also performed on 763 manipulated image pairs, of which 365 "large manipulations" come from CASIA V2.0 and 398 "small manipulations" from the NITS database. The manipulated NITS dataset was created by varying the manipulated range from 1 to 30%. In addition, the images consist of different foregrounds and backgrounds. This makes the data set a challenge. The PMCC (Pearson Product Momentum Correlation Coefficient) is calculated between the hash codes of the reference image and the corresponding semantically similar images, dissimilar image pairs and manipulated image pairs.

It was experimentally evaluated that by choosing the threshold value δ1 =0.98 only 1.10% of the semantically similar images are classified as different and only 1.57% of the manipulated images are falsely recognized as similar. However, the FPR for different pairs is zero. Therefore, the chosen threshold offers a better compromise between FPR and TPR as shown in the table below: hash correlation TPR FPR 0.97 0.9979 0.0157 0.98 0.9879 0.0130 0.99 0.9626 0.0130

However, the false acceptance rate is further improved by varying δ1, which depends on the specific application. Here the ability of the proposed system to detect and locate malicious additions or deletions of content on images is evaluated. The classification is done depending on the percentage of manipulated content on the image. Images with more than 29% manipulated content are classified as "heavily manipulated" and as "different image" considered. Images with a fake area between 1 and 29% are classified as "small". 1.57% of total fake images are misclassified as semantically similar.

According to one embodiment, the Receiver Operating Characteristics (ROC) curve is a suitable approach to measure the performance of a classifier at different thresholds. If two methods have the same FPR value, then the method with the higher TPR value will outperform the other. The performance comparison is made from four points of view: the trade-off between robustness and discrimination, robustness against CPOs, robustness against virulent operations, and the overall performance of the module. The proposed system has a higher AUC than the other systems. This may be due to the geometric correction. Therefore, the proposed system offers better trade-off performance between robustness and discrimination.

The Karsh et al. and Yan et al. The proposed approach also has good trade-offs that the proposed system follows. Both methods are also robust to some geometric operations. Therefore, the performance is comparable to that of the proposed system.

The table below shows the comparison of TPR rates when determining the optimal threshold for all compared systems. operation Lv et al Ouyang et al Tang et al. Yan et al. Karsh et al. Proposed system Big manipulations 0.4042 0.2383 0.180 0 0.342 5 0.042 5 0.0082 Small manipulations 0.6123 0.4692 0.451 8 0.620 0 0.120 7 0.0226

The hash codes are found to be robust to most parameters, with the exception of gamma correction. The proposed system has a higher TPR rate than comparable systems. So, the proposed system is robust in terms of brightness, contrast adjustment, salt and pepper noise addition, JPEG compression, embedding, watermarking, etc., but is somewhat limited in gamma correction.

The robustness of Yan et al. and Karsh et al. is greater than that of the other methods, followed by the proposed method. The systems have a geometric correction on the receiver side. The technique of Ouyang et al. offers very good robustness to rotation and scaling, but is sensitive to other factors. In contrast, the method of Lv et al. robust to watermark embedding due to the SIFT feature, but sensitive to other factors. The technique of Tang et al. is robust to rotation due to the ring partition approach, but sensitive to gamma correction, translation, and composite RST.

The proposed system has the lowest false acceptance rate among the systems compared. Only 3 out of 365 "large distorted" images are incorrectly classified as similar, while 9 out of 398 "small distorted" images are incorrectly accepted. However, this is still less than the compared systems. This could be due to the presence of a stacked convolution layer representing the deep feature image. As a result, small manipulations are adequately detected. The FPR for small-scale manipulations is higher than for large-scale because small content may not be reflected well in the hash. Therefore, the FPR for small manipulations is 0.0226, while for large manipulations it is 0.0082. However, when two or more ROC curves from compared systems are connected, it can be difficult to distinguish which system is superior. In such a case, the performance of the systems can be compared using the area under the ROC curve (AUC). The AUC indicates the area bounded by the ROC curve and the coordinate axis. It is evident that the AUC of the proposed system is 0.9973, which is the highest value among the compared systems. It is noted that the performance of the proposed system using only global features is better than the performance of the prior art (Tang et al.) which uses the features, i. H. global and local, combined. The proposed system detects tampering in a small area, which could be due to the deep learning paradigm, which is a major limitation of the existing systems.

With simultaneous manipulation and rotation, the proposed systems identify the area of \u200b\u200bthe manipulation. In addition, the proposed system restores the rotated image on the receiver side without requiring any information from the sender side. In contrast, need the previous procedures information from the sender side. Although the hash length of the proposed system is larger than that of the systems compared, it offers a higher AUC and a lower false acceptance rate. Furthermore, the proposed system can localize the manipulated regions even when manipulations and composite RST occur simultaneously, and it is also more robust to CPOs.

According to an alternative embodiment, the weight matrices of the CSDAE can be pre-trained with an RBM (Restricted Boltzmann Machines) to achieve faster convergence and lower MSE losses.

According to another alternative embodiment, a binary relaxation term can be added to the loss function to make the hash codes sparse.

According to another alternative embodiment, the residual connections can be added in the CSDAE to avoid the zero gradient problems encountered during model training.

The figures and the preceding description give examples of embodiments. Those skilled in the art will understand that one or more of the elements described may well be combined into a single functional element. Alternatively, certain elements can be broken down into multiple functional elements. Elements from one embodiment may be added to another embodiment. For example, the order of the processes described herein may be changed and is not limited to the manner described herein. Also, the acts of a flowchart need not be performed in the order presented; also, not all actions have to be performed. Also, the actions that are not dependent on other actions can be performed in parallel with the other actions. The scope of the embodiments is in no way limited by these specific examples. Numerous variations are possible, regardless of whether they are explicitly mentioned in the description or not, e.g. B. Differences in structure, dimensions and use of materials. The scope of the embodiments is at least as broad as indicated in the following claims.

Advantages, other benefits, and solutions to problems have been described above with respect to particular embodiments. However, the benefits, advantages, problem solutions, and components that can cause an advantage, benefit, or solution to occur or become more pronounced are not to be construed as a critical, required, or essential feature or component of any or all claims.

Reference List

100

A neuro-evolutionary detection system for early detection of lung cancer.

102

recipient

104

Geometry transformation correction module

106a

encoder

106b

decoder

106

Image authentication module

108

estimation module

110

Module for detection and localization of manipulations

Claims (10)

A neuro-evolutionary detection system (100) for early detection of lung cancer, the system (100) comprising: a plurality of receivers (102) for receiving an input image and a hash value, wherein the received image is transformed with a blind rotational scaling transform (RST) is encapsulated; a geometric transform correction module (104) connected to the plurality of receivers (102) for eliminating features of the blind RST transform by using inherent properties of the geometric transform and evaluating the rotation condition of the received image, wherein the features of the blind RST transformation are eliminated by the Koordi nats of boundary pixels having non-zero values are calculated to obtain an interpolated image; an image authentication module (106) connected to the geometric transformation module (104) for generating a hash code by an L2 regularization constraint (L2RC) applied to activation of a termination block in an encoder (106a), an estimation module (108), embodied in the image authentication module (106), estimates a correlation coefficient between the received hash and the generated hash, wherein if the correlation coefficient is less than a first threshold, the interpolated image is either a manipulated image or a different image ; and a tamper detection and localization module (110) coupled to the image segmentation module (106) for classifying the image as a tampered image or another image identification portion of the tampered images, wherein the tampered image is mapped to the generated hash code to generate an image map and locate the manipulated image. system after claim 1 , where the boundary pixels include the top pixel, the rightmost pixel, the leftmost pixel, and the bottom pixel. system after claim 1 , where the geometric transformation correction module evaluates the direction of rotation, i.e., either clockwise or counterclockwise, and adjusts the rotated image, where if the received image is rotated counterclockwise by up to θ degrees, the top not -zero coordinate point is further to the right than the bottom point, regardless of its degree of translation, and vice versa for a clockwise rotated image up to -θ degrees, where if the value of θ is substantially equal to θ' degrees, the received image as a rotated image is viewed that is counter-rotated by the θ' degree. system after claim 1 wherein a binary interpolation module (104) calculates a region of interest by manipulating the image to one dimension of the received image while adjusting the angle of rotation. system after claim 1 , wherein the encoder (106a) is a convolutional stacked denoising auto-encoder (CSDAE) having a plurality of blocks for generating the hash image, the CSDAE being a forward neural network, each of the plurality of blocks consisting of a convolutional layer followed by a stack normalization and a max pooling layer. system after claim 5 , where the CSDAE hierarchically maps received interpolated images into a latent space reduced to the dimension of the received image, where the L2 RC is configured to: Remove irrelevant components from the hash image by choosing a smallest combination to solve learning problems; and to suppress the effects of static noise on targets by preventing overfitting. system after claim 1 , where the corrupted area is located by multiplying the corrupted area by the corrupted image, where if the corrupted area is less than 30%, the received image is judged as "corrupted image" and otherwise as "different image". system after claim 1 , where the first threshold is 0.98 and a second threshold for tempering localization is 0.5. system after claim 1 , where each convolutional layer is initialized with a plurality of filters, except for a last layer, each having a kernel size of (3,3), where the CSDAE is trained layer by layer and then fine-tuned as a whole. system after claim 1 , wherein a decoder (106b) block layer comprises a 2D convolution layer and an upsampling layer followed by a batch norm layer.

Images