Space imaging interference method based on neural network
By adopting a neural network-based method in spatial imaging technology, combining environmental perception and multimodal feature fusion, adaptive interference strategies are generated and interference information is embedded in the frequency domain, the problem of interference methods in the prior art lack of adaptability and multimodal processing capabilities is solved, and a more natural, interpretable and stable interference effect is achieved.
Patent Information
- Application Number
- CN202411703746.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-26
- Publication Date
- 2025-04-22
Smart Images
Figure CN119862525A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of space imaging, and in particular to a space imaging interference method based on a neural network. Background Art
[0002] Space imaging interference refers to the use of technical means to process images in space to change their content, thereby affecting the recognition, analysis or understanding of the image. This interference can be used in a variety of scenarios, such as protecting privacy, preventing illegal use of images, and improving military defense. Space imaging interference can be achieved through a variety of methods such as adding noise, modifying image content, and adjusting image parameters. Its goal is to make the processed image difficult to accurately identify or interpret under certain conditions.
[0003] The existing methods for implementing spatial imaging interference mainly include adding noise, image synthesis, parameter adjustment, frequency domain processing and physical interference. Adding noise is to reduce the clarity and recognizability of the image by adding random or specific types of noise to the image. Common noise types include Gaussian noise, salt and pepper noise, etc. Image synthesis is to add false objects or content to the image to disrupt the real information of the image. Generative adversarial networks (GAN) or variational autoencoders (VAE) are usually used to generate false content. Parameter adjustment is to change the visual effect of the image by adjusting parameters such as contrast, brightness, saturation, etc. of the image. It can be achieved using traditional image processing technology or machine learning models. Frequency domain processing is to convert the image to the frequency domain, select a specific frequency band to embed interference information, and then convert it back to the air domain to prevent the interference information from being directly detected in the air domain. Physical interference is to directly affect the image acquisition process by placing interference objects in the real environment or adjusting environmental parameters. It is often used in the military and security fields. However, the existing technology has some shortcomings in implementing spatial imaging interference. First, many existing technologies use fixed interference methods and lack adaptive capabilities. For example, the method of adding noise may have different effects in different environments and scenarios, and the interference strategy cannot be adjusted dynamically. Secondly, the interference effects generated by some interference methods are difficult to reverse, which is an important limitation in some application scenarios where the original image needs to be restored. In addition, the interference effects generated by some interference methods are unnatural and easy to be identified and detected, which reduces the effectiveness of the interference. In addition, existing technologies usually only target images of a single modality and lack the ability to comprehensively process multimodal images. Finally, the interference effects of existing technologies often do not rely on specific environmental characteristics, resulting in unstable effects in different environments. Summary of the invention
[0004] In view of the defects in the prior art, the present invention provides a method for spatial imaging interference based on a neural network, which specifically includes the following steps:
[0005] S1. Environmental perception and feature extraction
[0006] The environmental information of the target area is collected in real time through environmental sensors, and key features are extracted from the environmental information and multimodal images using feature extraction methods;
[0007] S2. Multimodal feature fusion and adaptive interference strategy generation
[0008] The extracted environmental features and multimodal image features are fused to generate a comprehensive feature vector, and an adaptive interference strategy is generated through a reinforcement learning model;
[0009] S3. Dynamic interference generation
[0010] According to the generated adaptive interference strategy, the interference information is embedded in the frequency domain using a multi-scale generation model and a time series generation model, and superimposed with the original image;
[0011] S4. Environmental dependence and anti-interference resistance
[0012] The environmental features are integrated with the interference information to ensure that the interference effect is highly consistent with the environment, and anti-interference training is performed by building a generative adversarial network;
[0013] S5. Interpretability and reversibility design
[0014] The attention mechanism is introduced to generate interpretable interference effects, and the interference model is designed using a reversible convolutional network to ensure that the original image can be restored under certain conditions;
[0015] S6. Distributed Interference
[0016] The image data is divided into multiple subtasks and assigned to different computing nodes for parallel processing, and the interference results of each node are integrated in the control center.
[0017] Preferably, the step S1 of environmental perception and feature extraction further includes: using environmental sensors to collect temperature, humidity, and light intensity environmental information of the target area, using convolutional neural networks to extract key features from the environmental information, and image feature extraction using convolutional neural networks to extract key features from multimodal images.
[0018] Preferably, the step S2 of multimodal feature fusion and adaptive interference strategy generation further includes: using an attention mechanism or a feature splicing method to fuse environmental features and multimodal image features, adaptive strategy generation defines states (environmental features + current image), actions (interference actions) and rewards (image clarity, detection errors), and uses a deep reinforcement learning model DQN or PPO to train a decision model.
[0019] Preferably, the step S3 of dynamic interference generation further includes: using a multi-scale generation model (such as Progressive GAN) and a time series generation model (such as LSTM or GRU+GAN) to generate interference, embedding interference information through frequency domain processing (such as Fourier transform), and generating corresponding interference according to the material classification result, and finally superimposing the interference information with the original image.
[0020] Preferably, the step S4 of environmental dependence and anti-interference confrontation further includes: fusing environmental features with interference information to ensure that the interference effect is highly consistent with the specific environment, constructing a generative adversarial network (GAN), including an interference model and an anti-interference model, and optimizing the interference model through adversarial training.
[0021] Preferably, the interpretability and reversibility design of step S5 further includes: introducing an attention mechanism to generate an attention map, and using a natural language processing model to generate text or graphic explanations, generating a key, and using a reversible convolutional network (such as i-RevNet) to design an interference generation model to ensure that it is difficult to restore the original image without the key.
[0022] Preferably, the S6 distributed interference further includes: dividing the image data into multiple subtasks and assigning them to different computing nodes, deploying the same interference generation model on each node, performing parallel processing, synchronizing the model parameters of each node, and finally integrating the results.
[0023] The beneficial effects of the present invention are as follows: the present invention automatically generates the optimal interference strategy according to the environmental features and image content by introducing a reinforcement learning model, thereby improving the adaptability and dynamics of the interference method; introducing an attention mechanism in the interference generation process to generate an interpretable attention map, so that the interference effect has stronger interpretability; at the same time, a reversible convolutional network is used to design an interference model to ensure that the original image can be restored under specific conditions; through multimodal fusion and frequency domain processing, the generated interference effect is more natural and difficult to be detected by existing anti-interference technology, thereby improving the effectiveness of interference; the present method closely combines environmental features with the interference generation process to ensure that the interference effect is highly dependent on specific environmental features, thereby being more stable and effective in different environments; the present method can process multimodal image data, generate interference effects by combining features of different modalities, and improve the comprehensive processing capability of the interference method; through a distributed computing framework, the present method can generate interference effects in parallel on multiple nodes, thereby improving computing efficiency and flexibility, and being suitable for large-scale applications; the present method has obvious advantages in adaptability, interpretability, reversibility, naturalness, environmental dependence, and multimodal processing capability, and can more effectively achieve the goal of spatial imaging interference. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for the specific embodiments or the description of the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.
[0025] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION
[0026] The following embodiments of the technical solution of the present invention are described in detail in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and are therefore only used as examples, and cannot be used to limit the protection scope of the present invention.
[0027] It should be noted that, unless otherwise specified, the technical terms or scientific terms used in this application should have the common meanings understood by those skilled in the art to which the present invention belongs.
[0028] like Figure 1 As shown, an embodiment of the present invention provides a method for spatial imaging interference based on a neural network, which specifically includes the following steps:
[0029] S1. Environmental perception and feature extraction
[0030] Environmental information of the target area is collected in real time through environmental sensors, and key features are extracted from environmental information and multimodal images using feature extraction methods.
[0031] Environmental perception and feature extraction further include:
[0032] Environmental sensors are used to collect environmental information such as temperature, humidity, and light intensity in the target area. Convolutional neural networks are used to extract key features from the environmental information. Image feature extraction uses convolutional neural networks to extract key features from multimodal images.
[0033] S2. Multimodal feature fusion and adaptive interference strategy generation
[0034] The extracted environmental features and multimodal image features are fused to generate a comprehensive feature vector, and an adaptive interference strategy is generated through a reinforcement learning model.
[0035] The multimodal feature fusion and adaptive interference strategy generation further include:
[0036] The environmental features and multimodal image features are fused using the attention mechanism or feature splicing method, and the adaptive strategy generates the defined state (environmental features + current image), action (interference action) and reward (image clarity, detection error), and the decision model is trained using the deep reinforcement learning model DQN or PPO.
[0037] S3. Dynamic interference generation
[0038] According to the generated adaptive interference strategy, the interference information is embedded in the frequency domain using a multi-scale generation model and a time series generation model, and superimposed with the original image.
[0039] Dynamic interference generation further includes:
[0040] Use multi-scale generation models (such as Progressive GAN) and time series generation models (such as LSTM or GRU+GAN) to generate interference, embed interference information through frequency domain processing (such as Fourier transform), generate corresponding interference according to the material classification results, and finally superimpose the interference information with the original image.
[0041] S4. Environmental dependence and anti-interference resistance
[0042] The environmental features are integrated with the interference information to ensure that the interference effect is highly consistent with the environment, and anti-interference adversarial training is performed by constructing a generative adversarial network.
[0043] Environmental dependence and anti-interference countermeasures further include:
[0044] The environmental characteristics are integrated with the interference information to ensure that the interference effect is highly consistent with the specific environment. A generative adversarial network (GAN) is constructed, including an interference model and an anti-interference model, and the interference model is optimized through adversarial training.
[0045] S5. Interpretability and reversibility design
[0046] An attention mechanism is introduced to generate explainable interference effects, and a reversible convolutional network is used to design an interference model to ensure that the original image can be restored under certain conditions.
[0047] Interpretability and reversibility design further includes:
[0048] An attention mechanism is introduced to generate an attention map, and a natural language processing model is used to generate text or graphic explanations to generate keys. A reversible convolutional network (such as i-RevNet) is used to design an interference generation model to ensure that it is difficult to restore the original image without the key.
[0049] S6. Distributed Interference
[0050] The image data is divided into multiple subtasks and assigned to different computing nodes for parallel processing, and the interference results of each node are integrated in the control center.
[0051] Distributed interference further includes:
[0052] The image data is divided into multiple subtasks and assigned to different computing nodes. The same interference generation model is deployed on each node for parallel processing, and the model parameters of each node are synchronized. Finally, the results are integrated.
[0053] Use environmental sensors, meteorological sensors, and light sensors to collect environmental information of the target area in real time. Collect multimodal image data, such as visible light images, infrared images, and SAR images. Use convolutional neural networks (CNNs) to extract key features from the collected environmental information and multimodal images. Environmental features include weather, light, and geographic features. Image features include texture, shape, and color.
[0054] The extracted environmental features and multimodal image features are input into a fusion module and fused using multimodal fusion methods such as attention mechanism and feature splicing. A comprehensive feature vector is generated to represent the multimodal information of the target area. The comprehensive feature vector and the current image are used as state inputs. A series of possible interference actions are defined, such as adding noise, adjusting contrast, and adding false objects. A reward function is designed to evaluate the interference effect. For example, image clarity and detection error indicators can be used. A decision model is trained using a reinforcement learning algorithm such as DQN or PPO, which selects the optimal interference action based on the current state. Through interaction with the environment, the strategy is continuously optimized to improve the interference effect.
[0055] Use Progressive GAN to build a multi-scale generative model, gradually increasing the depth and resolution of the generator and discriminator. Combine LSTM or GRU to build a generative adversarial network (GAN) to generate dynamic interference. Input the interference action selected by the adaptive strategy generation module and the current image into the generation module. Convert the image to the frequency domain, select a specific frequency band to embed the interference information, and then convert it back to the spatial domain. Generate interference information that matches the material characteristics based on the material type output by the material classification module. Overlay the generated interference information with the original image to form the final interference image.
[0056] The environmental features are fused with the output of the interference generation module to ensure that the interference information is highly consistent with the environment. The frequency of the interference information is adjusted in the frequency domain to match the changes in the environment. A generative adversarial network (GAN) is designed, which includes an interference model and an anti-interference model. The interference model generates an interference image. The anti-interference model attempts to restore the original image. Through adversarial training, the interference model is optimized so that the interference it generates is difficult to be restored by the anti-interference model.
[0057] Introduce an attention mechanism into the interference generation model to generate an attention map that highlights the key areas of interference. Use a natural language processing model (such as BERT, Transformer) to generate a text explanation, or use a graph generation model to generate a graphical explanation to explain the specific method and impact of the interference. Generate a key for encoding and decoding interference information. Use a reversible convolutional network (such as i-RevNet) to build an interference generation model. Input the original image and the key into the encoding module to generate the interference image. Use the decoding module to restore the original image based on the key. Design a loss function to ensure that the interference effect is difficult to recover without the key.
[0058] Divide the image data to be interfered into multiple subtasks and assign them to different computing nodes. Deploy the same interference generation model on each node. Each node processes the assigned subtasks in parallel to generate interference images. Use the distributed computing framework TensorFlow's tf.distribute or PyTorch's Distributed Data Parallel to synchronize the model parameters of each node. Design a control center to coordinate the task allocation and result integration of each node. Splice the interference images generated by each node to form the final interference image. Check whether the interference effects generated by each node are consistent to ensure the effectiveness and stability of the overall interference effect.
[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.
Claims
1. A method for spatial imaging interference based on neural network, characterized in that: The specific steps include: S1. Environmental perception and feature extraction The environmental information of the target area is collected in real time through environmental sensors, and key features are extracted from the environmental information and multimodal images using feature extraction methods; S2. Multimodal feature fusion and adaptive interference strategy generation The extracted environmental features and multimodal image features are fused to generate a comprehensive feature vector, and an adaptive interference strategy is generated through a reinforcement learning model; S3. Dynamic interference generation According to the generated adaptive interference strategy, the interference information is embedded in the frequency domain using a multi-scale generation model and a time series generation model, and superimposed with the original image; S4. Environmental dependence and anti-interference resistance The environmental features are integrated with the interference information to ensure that the interference effect is highly consistent with the environment, and anti-interference training is performed by building a generative adversarial network; S5. Interpretability and reversibility design The attention mechanism is introduced to generate interpretable interference effects, and the interference model is designed using a reversible convolutional network to ensure that the original image can be restored under certain conditions; S6. Distributed Interference The image data is divided into multiple subtasks and assigned to different computing nodes for parallel processing, and the interference results of each node are integrated in the control center.
2. The method of spatial imaging interference based on a neural network according to claim 1, characterized in that: The step S1 of environmental perception and feature extraction further includes: using environmental sensors to collect temperature, humidity, and light intensity environmental information of the target area, using convolutional neural networks to extract key features from the environmental information, and image feature extraction using convolutional neural networks to extract key features from multimodal images.
3. The method of spatial imaging interference based on neural network according to claim 1, characterized in that: The step S2 of multimodal feature fusion and adaptive interference strategy generation further includes: Use attention mechanism or feature splicing method to fuse environmental features and multimodal image features, adaptive strategy generation defines states, actions and rewards, and use deep reinforcement learning model DQN or PPO to train the decision model.
4. The method of spatial imaging interference based on neural network according to claim 1, characterized in that: The step S3 of dynamic interference generation further includes: generating interference using a multi-scale generation model and a time series generation model, embedding interference information through frequency domain processing, generating corresponding interference according to the material classification result, and finally superimposing the interference information with the original image.
5. The method of spatial imaging interference based on neural network according to claim 1, characterized in that: The step S4 of environmental dependence and anti-interference confrontation further includes: fusing environmental features with interference information to ensure that the interference effect is highly consistent with the specific environment, constructing a generative adversarial network, including an interference model and an anti-interference model, and optimizing the interference model through adversarial training.
6. The method of spatial imaging interference based on neural network according to claim 1, characterized in that: The interpretability and reversibility design of step S5 further includes: introducing an attention mechanism to generate an attention map, and using a natural language processing model to generate text or graphic explanations, generating a key, and using a reversible convolutional network to design an interference generation model to ensure that it is difficult to restore the original image without the key.
7. The method of spatial imaging interference based on neural network according to claim 1, characterized in that: The S6 distributed interference further includes: dividing the image data into multiple subtasks and allocating them to different computing nodes, deploying the same interference generation model on each node, performing parallel processing, synchronizing the model parameters of each node, and finally integrating the results.