An intelligent generated image detection method based on multi-granularity artifact feature fusion

Abstract

The present application relates to the field of generated image detection, and provides an intelligent generated image detection method based on multi-granularity artifact feature fusion, comprising: collecting generated images and real images as a basic data set; extracting multi-granularity artifact features of images in the basic data set; training a multi-granularity artifact feature fusion detection model using the multi-granularity artifact features; and using the trained multi-granularity artifact feature fusion detection model to detect generated images. In one aspect, the present application extracts multi-granularity artifact features covering local-global and time-frequency. In another aspect, the multi-granularity artifact features are input as embedding sequences of a ViT model, the ability of a Transformer encoder to directly model global relationships between multi-granularity artifact features is utilized, and the multi-head self-attention mechanism is used to realize automatic association and fusion of multi-granularity artifact features. The method of the present application comprehensively utilizes the detection advantages of different granularity artifact features, and improves the generalization and robustness of generated image detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of generated image detection, in particular to an intelligent generated image detection method based on multi-granularity artifact feature fusion. BACKGROUND

[0002] With the rapid development of generative artificial intelligence, such as GAN (Generative Adversarial Network) and diffusion models, high-quality images can be synthesized in seconds, and it is difficult for human observers to distinguish them from real images. Although generative artificial intelligence brings great benefits, the misuse of generative technology is increasingly worrying. They can be used to create fake news and spread widely on the Internet. Therefore, it is urgent to develop effective detection methods for artificial intelligence generated images to reduce the misuse of generative technology and enhance the credibility of these technologies.

[0003] With the generated model image generation effect getting better and better, the spatial domain artifacts in the generated image are less and less obvious, but there are still more advanced semantic errors, such as the number of fingers not conforming to common sense. And frequency artifacts always exist in generated images, the main reason is that general image generation models all use up-sampling components to convert low-resolution hidden space into high-resolution, such as the up-sampling component in the GAN image generation model is ubiquitous, and the up-sampling layer in the diffusion model (such as DDPM (Denoising Diffusion Probabilistic Models), ADM (Ablated Diffusion Model), and LDM (Latent Diffusion Model)) widely using U-Net model is also a key module. Studies have shown that up-sampling operations change the spectral characteristics of signals, leading to high-frequency distortion in the output, which is also the case in artificial intelligence generated images, so frequency artifacts are often used as a typical feature to detect generated images. Based on this, the current methods to distinguish real images and artificial intelligence generated images are mainly two categories: one is a data-driven detection model, which mainly detects generated images through semantic features; the other is a detection model based on artifact features, which mainly realizes it through identifying texture-level artifacts generated in image generation, which can be spatial and temporal artifacts, or even a combination of the two. But overall, the current detection methods still have weak generalization ability and are easily affected by post-processing techniques such as JPEG compression (Joint Photographic Experts Group, an international standardization organization for static image compression standards). SUMMARY

[0004] The present application aims to provide an intelligent generated image detection method based on multi-granularity artifact feature fusion, to solve the problem of weak generalization ability and being easily affected by post-processing operation of the existing generated image detection method.

[0005] The present application provides an intelligent generated image detection method based on multi-granularity artifact feature fusion, comprising:

[0006] Collecting generated images and real images as a basic data set;

[0007] Extracting multi-granularity artifact features of images in the basic data set;

[0008] Training a multi-granularity artifact feature fusion detection model using the multi-granularity artifact features;

[0009] Using the trained multi-granularity artifact feature fusion detection model to detect generated images.

[0010] In a preferred embodiment, all images in the basic data set need to be subjected to data enhancement including image compression and image blurring; wherein the image compression is achieved by performing JPEG compression operation on the image, and the image blurring is achieved by performing Gaussian blurring on the image.

[0011] In a preferred embodiment, the multi-granularity artifact features include near-neighbor pixel relationship features, decoding artifact features, high-frequency noise features, and image semantic features.

[0012] In a preferred embodiment, the method for extracting the near-neighbor pixel relationship features comprises:

[0013] Regarding the image as a number of image blocks;

[0014] The near-neighbor pixel relationship features of the image are the pixel values of each image block of the image minus the pixel value of the top-left corner of the image block.

[0015] In a preferred embodiment, the method for extracting the decoding artifact features comprises:

[0016] Inputting the image into a pre-trained variational autoencoder;

[0017] Subtracting the input image from the output image of the variational autoencoder to obtain the decoding artifact features.

[0018] In a preferred embodiment, the method for extracting the high-frequency noise features comprises:

[0019] Using a steganalysis enrichment model to extract the high-frequency noise features from the image.

[0020] In a preferred embodiment, the method for extracting the image semantic features comprises:

[0021] An image encoder using the CLIP model extracts the image semantic features.

[0022] In a preferred embodiment, the near-neighbor pixel relationship features, the decoding artifact features and the high-frequency features need to be extracted by a feature extraction network to maintain the same dimension as the image semantic features; wherein the feature extraction network is constructed using residual blocks and convolutional layers.

[0023] In a preferred embodiment, the multi-granularity artifact feature fusion detection model uses a ViT model.

[0024] In a preferred embodiment, the loss function for training the multi-granularity artifact feature fusion detection model uses binary cross-entropy loss.

[0025] In summary, due to the use of the above technical solutions, the beneficial effects of the present application are:

[0026] In view of the current situation that the existing mainstream generated image detection features have advantages in generalization or robustness but cannot be considered together, the present application designs a multi-granularity artifact feature extraction method covering local-global and time-frequency, and realizes feature standardization; on the other hand, the multi-granularity artifact features are input as the embedding sequence of the ViT model, and the ability of the Transformer encoder to directly model the global relationship between image blocks is ingeniously utilized to realize the automatic association and fusion of multi-granularity artifact features by using the multi-head self-attention mechanism. Thus, the present application method comprehensively utilizes the detection advantages of different granularity artifact features to improve the generalization and robustness of generated image detection. At the same time, the method has strong scalability, and if effective new detection features appear, they can be easily used to enhance the detection ability of the model. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 A flowchart of an intelligent generated image detection method based on multi-granularity artifact feature fusion provided for an embodiment of the present application.

[0028] Figure 2 A structural diagram of a feature extraction network in an embodiment of the present application.

[0029] Figure 3 A generated image detection network architecture diagram in an embodiment of the present application.

[0030] Figure 4 A structural diagram of an electronic device provided for an embodiment of the present application. DETAILED DESCRIPTION

[0031] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will be combined with the accompanying drawings of the embodiments of the present application to make a clear and complete description of the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the accompanying drawings can be arranged and designed in various different configurations.

[0032] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0033] As shown in Figure 1 , the embodiments of the present application provide an intelligent generated image detection method based on multi-granularity artifact feature fusion, comprising:

[0034] S100, collecting generated images and real images as a basic data set;

[0035] S200, extracting multi-granularity artifact features of images in the basic data set;

[0036] S300, training a multi-granularity artifact feature fusion detection model using the multi-granularity artifact features;

[0037] S400, using the trained multi-granularity artifact feature fusion detection model to detect generated images.

[0038] The method of the present application comprehensively utilizes the detection advantages of different granularity artifact features to improve the generalization and robustness of generated image detection.

[0039] Set the dimension of the input image to , and the specific implementation of the above intelligent generated image detection method based on multi-granularity artifact feature fusion will be described in detail.

[0040] S100, collecting generated images and real images as a basic data set:

[0041] In the embodiments of the present application, from a typical generated image data set, generated images and real images generated by two categories of methods based on a generative adversarial network and a diffusion model are collected as a basic data set. All images in the basic data set need to be data enhanced including image compression and image blurring to help the detection model learn different attentions under different input conditions. Among them, the image compression is realized by JPEG compression operation on the image, and the image blurring is realized by Gaussian blurring on the image.

[0042] S200, Extract multi-granularity artifact features from the images in the basic dataset:

[0043] This invention presents a multi-granularity artifact feature extraction method covering local-global and time-frequency domains, and implements feature standardization. The multi-granularity artifact features include nearest-neighbor pixel relationship features, decoding artifact features, high-frequency noise features, and image semantic features.

[0044] (1) Nearest neighbor pixel relationship features:

[0045] Directly utilizing the nearest-neighbor pixel relationship features left by the upsampling operation as detection features has good generalization ability, but it is sensitive to post-processing operations. Because upsampling results in strong similarity among nearest-neighbor pixels in the image, the following method is used to extract nearest-neighbor pixel relationship features:

[0046] Image See as When using nearest-neighbor interpolation as an upsampling layer, each channel of an image patch of a certain size shares the same value. Although there are generally subsequent convolutional layers, the convolutional layers are fixed and translation-invariant when generating the image, so the similarity of neighboring pixels still exists.

[0047] image Nearest neighbor pixel relationship features For the image Each The pixel values ​​of each image patch are reduced by the pixel value of its top-left corner. This determines the nearest neighbor pixel relationship feature. Dimensions and images Same dimensions. Assume , Image Will be seen as indivual Image patch size, nearest neighbor pixel relationship features The calculation method is as follows:

[0048]

[0049] in, For image The a Line 1 b Columns The size of the image patch c 1 pixel.

[0050] (2) Decoding artifact features:

[0051] The decoding artifact feature refers to an artifact feature caused by a common component decoder of a generative model, and compared with a near-neighbor pixel relationship feature, the decoding artifact feature is a higher-level feature and is more robust in the face of various post-processing transformations. The decoding artifact feature is extracted by the following method:

[0052] An image is input into a pre-trained variational autoencoder (VAE);

[0053] An output image of the variational autoencoder is obtained. The decoding artifact feature is obtained by subtracting the input image from the output image , and is denoted as . The decoding artifact feature has the same dimension as the image

[0054] .

[0055] The high-frequency noise feature refers to a high-frequency signal feature of a generated image that is retained by a high-pass filter. Research shows that the high-frequency noise feature has a greater effect on generated image detection, but the high-frequency noise features of images generated by each model are not quite the same. Therefore, only using the high-frequency noise feature for generated image detection can easily overfit and cause poor generalization.

[0056] Inspired by the research on a steganalysis rich model (SRM) in image forensics, the method for extracting the high-frequency noise feature is to extract the high-frequency noise feature from an image by using the steganalysis rich model. In the embodiments of the present application, the high-frequency noise feature is extracted from an RGB image (Red-Green-Blue) by using the steganalysis rich model. Three filter kernels in the steganalysis rich model are used to extract the high-frequency noise feature, and the weights of the three filter kernels are as follows:

[0057]

[0058]

[0059]

[0060] The filter kernel size of the steganalysis rich model is defined as 5*5*3, and the output channel size of the steganalysis rich model is 3. The high-frequency noise feature has the same dimension as the image .

[0061] (4) Image semantic feature:

[0062] In the embodiment of the present application, an image encoder of a CLIP (Contrastive Language-Image Pre-training) model is used to extract image semantic features. Because the CLIP model forces the image encoder to capture semantic-related visual features through contrastive learning, it covers a wider range of visual concepts and noise patterns, and its feature extraction process not only depends on pixel-level information, but also implicitly includes cross-modal consistency constraints. Semantic features have good robustness for post-processing, but if the training data is small, they are prone to overfitting to irrelevant features in the detection data set, such as character identity and object type. Specifically, the image encoder CLIP ViT-B / 16 of the CLIP model can be selected as the basic feature extraction model, and the output of the last layer of the network is taken as the image semantic feature, which has a total of 512 dimensions. The parameters of the image encoder remain unchanged during the training process.

[0063] In particular, in order to perform multi-granularity artifact feature fusion detection later, the near neighbor pixel relationship feature, the decoded artifact feature and the high frequency feature need to be extracted by a feature extraction network to maintain the same dimension as the image semantic feature. In the embodiment of the present application, each granular feature is encoded into a 512-dimensional vector to realize feature standardization and ensure that each granular feature can be processed equally when performing multi-granularity artifact feature fusion detection. The feature extraction network is constructed by using a residual block (Resnet block) and a convolutional layer, and the network structure of the feature extraction network is as shown in Figure 2

[0064] S300, training a multi-granularity artifact feature fusion detection model using multi-granularity artifact features:

[0065] In the embodiment of the present application, the multi-granularity artifact feature fusion detection model is implemented by using a ViT model (Vision Transformer).

[0066] ​The ViT model is capable of effectively processing long-distance dependencies by directly modeling the global relationship between image blocks through a multi-head self-attention mechanism by dividing an image into a patch sequence. Therefore, it is expected that the effective fusion of multi-granularity artifact features can be achieved through the ability of the ViT model, so as to realize the generation image detection. Specifically, after the artifact features of different granularities are uniformly mapped to the same dimensional vector space, the multi-granularity artifact features are input as the patch embedding sequence of the ViT model, and the weight relationship between different artifact feature embeddings is dynamically adjusted through the multi-head self-attention mechanism, so as to automatically capture the hierarchical correlation of the multi-granularity artifact features. For example, fine-grained features (such as local texture abnormalities) and coarse-grained features (such as global structure distortion) can be respectively modeled and fused in different subspaces through the multi-head attention mechanism. Here, a ViT-Base model is used to achieve this, and the input patch embedding is a 512-dimensional vector sequence of each granularity feature. Its structure is as shown in FIG. 2. Figure 3

[0067] S300, training a multi-granularity artifact feature fusion detection model using the multi-granularity artifact features:

[0068] When training the multi-granularity artifact feature fusion detection model using the multi-granularity artifact features, a stochastic gradient descent method is used to optimize the multi-granularity artifact feature fusion detection model. Since the multi-granularity artifact feature fusion detection model is to identify generated images and real images, i.e., a two-dimensional real-fake classification problem, in the present embodiment, the loss function for training the multi-granularity artifact feature fusion detection model uses a binary cross-entropy loss , which is expressed as:

[0069]

[0070] wherein, N is the number of input samples, is the true label (0 or 1, representing generated images and real images, respectively) of the i-th sample, is the predicted probability of the i-th sample.

[0071] S400, using the trained multi-granularity artifact feature fusion detection model to perform generation image detection:

[0072] ​For the image to be detected, the multi-granularity artifact features are also extracted, the feature standardization is input into the multi-granularity artifact feature fusion detection model, and the result that the image to be detected is a generated image or a real image and the corresponding prediction probability can be identified.

[0073] Based on the same technical concept, the embodiment of the present application also provides an electronic device which can implement the intelligent generated image detection method based on multi-granularity artifact feature fusion provided by the above-mentioned embodiment of the present application. In an embodiment, the electronic device can be a server, a terminal device or other electronic device. As shown in the figure, Figure 4 The electronic device can include:

[0074] At least one processor and a memory connected with the at least one processor, and the specific connection medium between the processor and the memory in the embodiment of the present application is not limited, Figure 4 In the embodiment of the present application, the connection between the processor and the memory is taken as an example by bus connection. The bus is represented by a thick line, Figure 4 The connection mode between other components is only schematically illustrated and is not limited. The bus can be divided into address bus, data bus, control bus, etc. For convenience of representation, Figure 4 In the embodiment of the present application, only one thick line is used to represent the bus, but it does not mean that there is only one bus or one type of bus. Alternatively, the processor can also be called a controller, and the name is not limited.

[0075] In the embodiment of the present application, the memory stores instructions executable by the at least one processor, and the at least one processor can execute the foregoing intelligent generated image detection method based on multi-granularity artifact feature fusion by executing the instructions stored in the memory.

[0076] The processor is the control center of the device, can utilize various interfaces and lines to connect each part of the whole control device, and through running or executing the instructions stored in the memory and calling the data stored in the memory, the device can process data and various functions, thereby overall monitoring the device.

[0077] In an optional design, the processor can include one or more processing units, and the processor can integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface and application program, etc., and the modem processor mainly processes wireless communication. It can be understood that the above-mentioned modem processor can also not be integrated into the processor. In some embodiments, the processor and the memory can be realized on the same chip, and in some embodiments, they can also be realized on independent chips respectively.

[0078] The processor can be a general purpose processor, such as a CPU, a digital signal processor, an application specific integrated circuit, a field programmable gate array, or other programmable logic device, discrete gate or transistor logic, discrete hardware components, can implement or execute the methods, steps and logic block diagrams disclosed in the embodiments of the present application. The general purpose processor can be a microprocessor or any conventional processor, etc. The steps of the method for intelligently generating image detection based on multi-granularity artifact feature fusion disclosed in the embodiments of the present application can be directly embodied as hardware processor execution, or executed by a combination of hardware and software modules in the processor.

[0079] The memory is a non-volatile computer readable storage medium, which can be used to store non-volatile software programs, non-volatile computer executable programs and modules. The memory can include at least one type of storage medium, such as flash memory, hard disk, multimedia card, card type memory, random access memory (RAM), static random access memory (SRAM), programmable read only memory (PROM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), magnetic memory, magnetic disk, optical disk, etc. The memory is any other medium capable of carrying or storing desired program code in the form of instructions or data structures and capable of being accessed by a computer, but is not limited to this. The memory in the embodiments of the present application can also be a circuit or any other device capable of realizing the storage function, used for storing program instructions and / or data.

[0080] By designing and programming the processor, the code corresponding to the method for intelligently generating image detection based on multi-granularity artifact feature fusion introduced in the foregoing embodiments can be fixed into the chip, so that the chip can execute the steps of the method of the foregoing embodiments when running. How to design and program the processor is a technology known to those skilled in the art, which will not be described here.

[0081] Based on the same inventive concept, the embodiments of the present application also provide a storage medium storing computer instructions, when the computer instructions run on a computer, the computer executes the foregoing method for intelligently generating image detection based on multi-granularity artifact feature fusion.

[0082] In some alternative embodiments, each aspect of the intelligent generated image detection method based on multi-granularity artifact feature fusion can also be implemented in the form of a program product, which includes program codes for causing the control device to perform the steps of the intelligent generated image detection method based on multi-granularity artifact feature fusion according to various exemplary embodiments of the present application described above in the specification when the program product is run on the device.

[0083] It should be noted that although several units or sub-units of the apparatus are mentioned in the foregoing detailed description, such division is merely exemplary and not mandatory. Indeed, features and functions of two or more units described above can be embodied in one unit according to embodiments of the present application. Conversely, a unit described above can be further divided into plural units to embody the features and functions thereof. Moreover, although the operations of the method of the present application are described in a particular order in the accompanying drawings, this does not require or imply that the operations must be performed in that particular order, or that all of the illustrated operations must be performed to achieve the desired results. Additionally or alternatively, certain steps can be omitted, combined, performed in a different order, and / or split into multiple steps to achieve the desired results.

[0084] Those skilled in the art will appreciate that embodiments of the present application can be provided as methods, systems, or computer program products. Accordingly, the present application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can be embodied in the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk memory, CD-ROMs, optical memory, etc.) having computer usable program code embodied thereon.

[0085] The present application is described with reference to the flowcharts and / or block diagrams of the methods, apparatus (systems) and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, a special purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowcharts and / or block diagrams block or blocks. Figure 1 The flow or flows and / or block or blocks Figure 1 The apparatus that carries out the functions specified in the flow or flows and / or block or blocks.

[0086] Program code to implement the application can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C++ or the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computing device, partly on the user's computing device, as a stand-alone software package, partly on the user's computing device and partly on a remote computing device or entirely on the remote computing device or server. The embodiments of the application covered herein are to be

[0087] In situations in which the remote computing device utilizes a network, the remote computing device can be connected to the user's computing device through any kind of network, including a local area network (LAN) or a wide area network (WAN). Such networks are well known to those having ordinary skill in the art and therefore will not be discussed herein in more detail.

[0088] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the Figure 1 function specified in the flow or flows and / or blocks Figure 1 of the multiple flows and / or blocks.

[0089] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the Figure 1 function specified in the flow or flows and / or blocks Figure 1 of the multiple flows and / or blocks.

[0090] The embodiments of the application described herein are to be used as illustration only, and are not intended to limit the scope of the application. Those skilled in the art will be able to devise numerous alterations and variations to the embodiments of the application without departing from the scope of the application. Accordingly, the application is intended to encompass all such alterations and variations as fall within the scope of the application.

Claims

1. An intelligent generated image detection method based on multi-granularity artifact feature fusion, characterized in that, The method comprises the following steps: Collecting generated images and real images as a basic dataset; Extracting multi-granularity artifact features of images in the basic dataset; The multi-granularity artifact features include near-neighbor pixel relationship features, decoding artifact features, high-frequency noise features, and image semantic features; A multi-granularity artifact feature fusion detection model is trained using the multi-granularity artifact features; the near-neighbor pixel relationship features, decoding artifact features, and high-frequency features need to be extracted by a feature extraction network to maintain the same dimension as the image semantic features; the multi-granularity artifact feature fusion detection model adopts a ViT model; after the multi-granularity artifact features are uniformly mapped to the same dimension of the vector space, the multi-granularity artifact features are input as the patch embedding sequence of the ViT model, and the weight relationship between different artifact feature embeddings is dynamically adjusted through the multi-head self-attention mechanism of the ViT model, so as to automatically capture the hierarchical correlation of the multi-granularity artifact features; The generated images are detected using the trained multi-granularity artifact feature fusion detection model. 2.The multi-granularity artifact feature fusion based intelligent generated image detection method according to claim 1, characterized in that, All images in the basic dataset need to be subjected to data enhancement including image compression and image blurring; wherein the image compression is realized by performing a JPEG compression operation on the image, and the image blurring is realized by performing a Gaussian blurring operation on the image. 3.The intelligent generated image detection method based on multi-granularity artifact feature fusion according to claim 1, characterized in that, The method for extracting the near-neighbor pixel relationship features comprises the following steps: The image is regarded as a plurality of image blocks; The near-neighbor pixel relationship features of the image are obtained by subtracting the pixel value at the top left corner of each image block from the pixel value of each image block. 4.The intelligent image generation method based on multi-granularity artifact feature fusion according to claim 1, characterized in that, The method for extracting the decoding artifact features comprises the following steps: The image is input into a pre-trained variational autoencoder; The output image of the variational autoencoder is subtracted from the input image to obtain the decoding artifact features. 5.The multi-granularity artifact feature fusion based intelligent generated image detection method according to claim 1, characterized in that, The method for extracting the high-frequency noise features comprises the following steps: The steganalysis rich model is used to extract the high-frequency noise features from the image. 6.The multi-granularity artifact feature fusion based intelligent generated image detection method according to claim 1, characterized in that, The method for extracting the image semantic features comprises the following steps: The image encoder of the CLIP model is used to extract the image semantic features. 7.The multi-granularity artifact feature fusion based intelligent generated image detection method according to claim 1, characterized in that, The feature extraction network is constructed using residual blocks and convolutional layers. 8.The multi-granularity artifact feature fusion based intelligent generated image detection method according to claim 1, characterized in that, The multi-granularity artifact feature fusion detection model adopts a ViT model. 9.The multi-granularity artifact feature fusion based intelligent generated image detection method according to claim 1, characterized in that, The loss function for training the multi-granularity artifact feature fusion detection model adopts a binary cross-entropy loss.

Images