Portable X-ray back scattering security check system based on AI image enhancement

Abstract

The invention discloses a portable X-ray back scattering security check system based on AI image enhancement. The system belongs to the technical field of radiation imaging, and comprises a sensing unit used for sensing physical situation information in an imaging process so as to generate time-sequenced physical prior data; the prediction unit is used for predicting a future scanning behavior based on the physical prior data to generate a feed-forward control instruction and prediction confidence; the regulation and control unit is used for adaptively regulating and controlling scanning parameters based on the feedforward control instruction and recording execution information including regulation and control residual errors; and the AI reconstruction processor is used for performing enhanced reconstruction on the original X-ray image based on the physical priori data, the prediction confidence and the execution information. By constructing an intelligent closed loop of perception, prediction, regulation and control and reconstruction, physical uncertainty information in the imaging process is quantized and used for guiding AI reconstruction, so that the problem of information fault between physical acquisition and digital reconstruction is solved, and the quality and reliability of a final image are improved.

Description

Technical Field

[0001] This invention relates to the field of radiation imaging technology, and in particular to a portable X-ray backscatter security inspection system based on AI image enhancement. Background Technology

[0002] Portable X-ray backscattering security inspection systems typically include an X-ray tube to generate X-rays, an X-ray detector to receive the scattered signals, and a data processor. Due to its single-sided imaging and flexible operation, this system is widely used in security inspections and non-destructive testing. To improve image quality, existing technologies typically employ artificial intelligence (AI) image enhancement algorithms for post-processing after acquiring the raw X-ray image. However, existing AI enhancement schemes have a technical flaw: the AI ​​model processes an already "damaged" image but is unaware of the physical causes of the damage; that is, there is a profound "information gap" between the physical acquisition process and the digital reconstruction process. During imaging, high-frequency random blurring caused by operator physiological micro-tremors and insufficient sampling in key areas due to uneven information density of the scanned target are two key physical uncertainties. Lacking prior information about these physical processes, existing AI models can only treat these complex, non-uniform distortions as ordinary noise and "blindly" repair them. Their performance has an inherent bottleneck, making it difficult to effectively recover image details obscured by the uncertainties of the physical processes, thus affecting the accuracy and reliability of the detection results. Summary of the Invention

[0003] This invention provides a portable X-ray backscatter security inspection system based on AI image enhancement, aiming to solve the technical problem in the prior art where the information gap between physical acquisition and digital reconstruction causes the AI ​​image enhancement model to be unable to effectively compensate for the image quality degradation caused by physical uncertainties during the imaging process.

[0004] In view of the above problems, the present invention provides a portable X-ray backscatter security inspection system based on AI image enhancement. The system includes a processor and a memory, wherein the memory stores a computer program, and the computer program executed by the processor includes: The sensing unit is used to sense physical context information during the imaging process in order to generate temporally sequenced physical prior data; The prediction unit is used to predict future scanning behavior based on the physical prior data, so as to generate feedforward control commands and prediction confidence. The control unit is used to adaptively control the scanning parameters of the imaging system based on the feedforward control command, and record the execution information of the control, the execution information including the control residual. An AI reconstruction processor is used to enhance and reconstruct the original X-ray image acquired by the X-ray detector based on the physical prior data, the prediction confidence level, and the execution information, so as to generate the final X-ray image.

[0005] The technical solution provided in this application has at least the following technical effects: This invention constructs an intelligent closed-loop system architecture of "perception-prediction-regulation-reconstruction," quantifying dynamically changing physical context information (such as operator micro-tremors and target information density) during the imaging process in real time. This high-dimensional physical prior information, containing the system's own cognitive uncertainty (prediction confidence) and execution bias (regulation residuals), is injected into the entire AI image reconstruction process. This method fundamentally bridges the information gap between physical acquisition and digital reconstruction, transforming AI enhancement from unguided "blind" post-processing to physically-based precise restoration, thereby simultaneously improving the clarity, signal-to-noise ratio, and overall fidelity of the final generated image. Attached Figure Description

[0006] Figure 1 This is a structural block diagram of a portable X-ray backscatter security inspection system based on AI image enhancement, provided as an embodiment of the present invention. Detailed Implementation

[0007] The above technical solutions will now be described in detail with reference to the accompanying drawings and specific embodiments to provide a better understanding of them. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments used only to explain the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Furthermore, it should be noted that, for ease of description, only the parts related to the present invention are shown in the drawings, not all of them.

[0008] Please see Figure 1 A portable X-ray backscatter security inspection system based on AI image enhancement, the system comprising a processor and a memory, the memory storing a computer program, the computer program running on the processor comprising: The sensing unit is used to sense physical context information during the imaging process in order to generate temporally sequenced physical prior data; The prediction unit is used to predict future scanning behavior based on the physical prior data, so as to generate feedforward control commands and prediction confidence. The control unit is used to adaptively control the scanning parameters of the imaging system based on the feedforward control command, and record the execution information of the control, the execution information including the control residual. An AI reconstruction processor is used to enhance and reconstruct the original X-ray image acquired by the X-ray detector based on the physical prior data, the prediction confidence level, and the execution information, so as to generate the final X-ray image.

[0009] This invention provides a portable X-ray backscatter security inspection system based on AI image enhancement, comprising a processor and a memory, the memory storing a computer program. When the processor runs the computer program, it implements functional units such as a sensing unit, a prediction unit, a control unit, and an AI reconstruction processor to execute an intelligent image enhancement process that includes physical process perception. The system's workflow begins with the sensing unit's perception and quantification of the imaging physical process.

[0010] The first step performed by the sensing unit is the perception and generation of physical prior data. This step is achieved by acquiring and processing multimodal physical signals in parallel. In a specific implementation path, the sensing unit acquires input from two physical sensors in parallel and continuously: an X-ray detector provides a raw X-ray photon signal stream; and an inertial measurement unit, rigidly connected to the system, provides a raw data stream of triaxial angular velocity and triaxial acceleration characterizing the system's motion state at a sampling frequency of not less than 1 kHz.

[0011] After acquiring the raw data stream from the inertial measurement unit (IMU), the signal processing module within the sensing unit acquires the real-time motion vector of the imaging system. This module calculates the time-varying motion vector representing the imaging system's translation and rotation in three-dimensional space by integrating the received three-axis angular velocity and three-axis acceleration data or using a preset attitude calculation algorithm. This motion vector is temporarily buffered for subsequent timestamp synchronization. For example, the attitude calculation algorithm can be a complementary filtering algorithm that fuses the accurate attitude measurements from the accelerometer in the low-frequency range with the accurate attitude change measurements from the gyroscope in the high-frequency range, thereby outputting a relatively accurate motion vector across the entire frequency band.

[0012] In parallel with the acquisition of real-time motion vectors, the signal processing module of the sensing unit also performs real-time quantization of local information density based on the X-ray photon signal stream received from the X-ray detector. One specific implementation of this quantization is the calculation of the short-term variance of the signal. The signal processing module sets a fixed-length sliding time window, for example, a window of 10 milliseconds. This sliding time window continuously moves across the received X-ray photon signal stream. For each time point, the signal processing module calculates and outputs the variance of all X-ray signal intensity values ​​within that time window. The calculated signal variance, as a quantized value, directly characterizes the image content complexity or information density of the currently scanned area. A higher signal variance value corresponds to a state where the image is being scanned across object edges, boundaries between different materials, or areas with complex internal structures; conversely, a lower signal variance value corresponds to a state where the image is being scanned across a region with uniform material. The calculated signal variance sequence is also temporarily cached.

[0013] After acquiring the real-time motion vector sequence and signal variance sequence, the core functional module of the sensing unit, namely a hardware synchronization module, performs timestamp synchronization and aggregation of multi-source physical information. In a specific implementation path, this hardware synchronization module is preferably a Field-Programmable Gate Array (FPGA) for nanosecond-level timing operations. This FPGA has a built-in high-frequency unified hardware clock. When the FPGA receives any X-ray data packet from the X-ray detector, it immediately appends a precise hardware timestamp based on this unified hardware clock to the X-ray data packet. Subsequently, using this hardware timestamp as an index, the FPGA searches the cached real-time motion vector sequence and signal variance sequence, respectively, and calculates the instantaneous motion vector value and instantaneous signal variance value corresponding to the precise moment of the hardware timestamp using linear interpolation or higher-order interpolation algorithms. Finally, the FPGA bundles and aggregates the original X-ray data packet, the calculated instantaneous motion vector, and the instantaneous signal variance into a new, unified data structure and continuously outputs it. This continuously outputting data stream containing synchronized multi-source physical information is the physical prior data in this embodiment of the invention.

[0014] After generating physical prior data, the system's workflow enters the prediction phase, where intelligent feedforward regulation is performed by the prediction unit. This prediction unit continuously receives physical prior data generated by the perception unit and predicts future scanning behavior based on this data. In one specific implementation, this prediction unit includes a pre-trained recurrent neural network (RNN) model. This RNN model takes a time series from the physical prior data (e.g., a sequence of motion vectors and signal variance over the past 100 milliseconds) as input. The RNN model is trained to identify the association between specific scanning patterns and image features. For example, the RNN model can be trained to recognize an "edge approach" pattern, where the model predicts that a high-contrast object edge is about to be scanned when the signal variance continuously increases and the motion vector direction is stable. Alternatively, the model can be trained to recognize a "repeating texture" pattern, where the model predicts that a surface with a regular texture is being scanned when the signal variance exhibits periodic fluctuations.

[0015] Based on the prediction results from the recurrent neural network model, the prediction unit generates feedforward control instructions. If the model's prediction outputs a specific future image feature, such as predicting an encounter with a high-density material point in the next 50 milliseconds, the prediction unit will generate one or more corresponding feedforward control instructions. These feedforward control instructions are structured data packets containing the target execution time, the scan parameters to be adjusted, and the target values. A specific feedforward control instruction could be: "At time T+50ms, increase the tube current of the X-ray tube by 15%."

[0016] Simultaneously with generating feedforward control instructions, the prediction unit also performs the concurrent generation of prediction confidence. This prediction confidence is a quantifiable value used to characterize the prediction model's self-assessment of the accuracy of its current prediction. In a specific implementation, the prediction confidence value can be directly taken from the probability value of the output layer of the recurrent neural network model (e.g., a Softmax layer). If the model predicts an edge with a 95% probability, the prediction confidence value is 0.95. This prediction confidence, along with the corresponding feedforward control instructions, is cached for subsequent system adjustments and image reconstruction.

[0017] After generating feedforward control commands and prediction confidence levels, the system's workflow is executed by the control unit. The control unit, in conjunction with the hardware synchronization module, performs adaptive dwell time control based on commands and information density. This control employs two parallel triggering mechanisms. The first is prediction-based feedforward control: when the prediction unit generates a feedforward control command for a future time, the command is sent to the control unit. When the system clock reaches the target execution time of the command, the control unit drives the execution unit (e.g., a haptic feedback motor connected to the operator's handle) to operate, providing a brief vibration feedback to the operator, prompting them to slow down the scanning speed in that area to increase dwell time. The second is real-time perception-based reactive control: the control unit independently and continuously monitors the real-time signal variance in the physical prior data. When this signal variance exceeds a preset dynamic threshold, it indicates that an unpredicted area with a sudden increase in information density is being scanned. At this point, the control unit immediately drives the execution unit, again providing vibration feedback to prompt the operator to immediately reduce the scanning speed. These two mechanisms together ensure that the dwell time during the scanning process dynamically matches the local information density of the scanned target.

[0018] While executing control commands, the control unit also performs closed-loop calculation and recording of control residuals. The goal of this step is to quantify the deviation between the system command and the actual execution result in the physical world. In a specific implementation path, this closed-loop calculation is performed by the control unit. Taking a feedforward control command as an example, the command includes a target scanning parameter value. The control unit obtains the actual change in the scanning parameter after the command is executed through feedback sensors within the system (e.g., by analyzing actual scanning speed changes or the operator's response time to feedback). The control unit then calculates the difference between the target value of the command and the actual value of the feedback, and records this difference as a quantified control residual. For example, if the command's goal is for the operator to reduce the scanning speed by 50%, but the actual feedback calculates a speed reduction of only 40%, then the 10% difference is recorded as the control residual. The calculated control residual, along with the corresponding timestamp and physical location information, is recorded to form complete execution information for final image reconstruction.

[0019] After completing the perception, prediction, and control of the physical process, the system's workflow enters the AI ​​enhancement and reconstruction stage executed by the AI ​​image enhancement processor. This AI image enhancement processor receives physical prior data, prediction confidence, and control residuals containing all physical process information, output from the perception, prediction, and control units. Based on this, it enhances the acquired raw X-ray images. The primary task of this enhancement and reconstruction process is to transform diverse process information into structured data that can be efficiently utilized by subsequent neural networks. The first step in this enhancement and reconstruction process is the construction of a multi-dimensional physical prior map. The goal of this step is to transform one-dimensional time-series data (e.g., motion vectors, confidence levels, etc.) into a two-dimensional image data structure that corresponds one-to-one with the pixels of the original X-ray image, facilitating subsequent neural network processing.

[0020] In a preferred embodiment, to improve processing efficiency and system integration, the aforementioned sensing unit, prediction unit, and control unit can be physically or logically integrated into a dedicated intelligent imaging engine module. This intelligent imaging engine module can be a dedicated board containing an FPGA and an embedded processor, specifically responsible for performing real-time physical information processing and control during the imaging process.

[0021] The construction of this multidimensional physical prior map first involves the dynamic construction of a local blur kernel based on real-time motion vectors. The AI ​​image enhancement processor traverses the received physical prior data, extracting the instantaneous motion vector corresponding to each pixel in the original X-ray image at the time of acquisition. For each pixel, the AI ​​image enhancement processor dynamically generates a two-dimensional blur kernel, i.e., a point spread function, to characterize the local blurring properties of that point based on its unique instantaneous motion vector. In a specific implementation, this point spread function can be modeled as a two-dimensional Gaussian function, whose covariance matrix is ​​determined by the instantaneous motion vector. For example, the principal axis of the probability density distribution of this two-dimensional Gaussian function is aligned with the direction of the instantaneous motion vector, and its standard deviation along the principal axis is proportional to the magnitude of the instantaneous motion vector. The local blur kernels generated for all pixels collectively constitute a "blur kernel map" of the same size as the original X-ray image.

[0022] While constructing the blurred kernel map, the AI ​​image enhancement processor also performs confidence weight calculations based on a fully coupled model. The goal of this calculation is to assign a quantified weight value representing the credibility of each pixel in the original X-ray image. This confidence weight calculation integrates information from multiple physical processes throughout the imaging process. The AI ​​image enhancement processor extracts the actual dwell time, instantaneous motion vector, prediction confidence, and control residual corresponding to each pixel, and uses these parameters as inputs into a pre-defined fully coupled mathematical model to calculate the final confidence weight for that pixel. The calculated confidence weights for all pixels together form a "confidence weight map" of the same size as the original X-ray image.

[0023] The core algorithm for confidence weighting is primarily reflected in its computational dependence on multi-source physical parameters. Specifically, to calculate the confidence weight of any pixel, the AI ​​image enhancement processor must simultaneously access three parameters from different dimensions of the imaging physical process: the first parameter is the instantaneous motion vector from the perception module, characterizing the quality of physical signal acquisition; the second parameter is the prediction confidence from the prediction module, characterizing the system's cognitive certainty; and the third parameter is the control residual from the action module, characterizing the precision of physical execution. This mandatory coupling of the mathematical calculations among these three parameters ensures that the final generated confidence weight comprehensively and holistically reflects the overall credibility of the pixel information.

[0024] In a preferred embodiment, the coupling dependency is weighted using a non-linear coupling mathematical formula. The AI ​​image enhancement processor calculates the confidence weights using the following formula ( ): Each component of this formula has its corresponding physical meaning and calculation path. Part One The confidence level of the basic data, among which It is the data saturation calculated based on the actual scanning dwell time of the pixel. One specific calculation method is... β is the preset sensitivity factor; It is based on the instantaneous motion vector corresponding to that pixel ( The calculated context distortion, one specific calculation method is... α is the preset distortion penalty factor. Part Two To predict the enhancement factor, where This is the prediction confidence level corresponding to that pixel, and γ is the preset enhancement index. Part Three To adjust the accuracy penalty factor, among which It is the adjustment residual corresponding to that pixel, and δ is the preset penalty coefficient.

[0025] To more clearly illustrate the numerical implementation of this weight calculation, a specific example is provided. Assume that when the AI ​​image enhancement processor processes a pixel, it extracts the following parameters from the received data stream: actual dwell time. Calculated data saturation The context distortion is 0.95, calculated from the instantaneous motion vector. The prediction confidence level is 0.4; the prediction confidence level corresponding to this pixel is 0.4. It is 0.8; the adjustment residual corresponding to this pixel. The value is 0.1. The system's preset weighting factors are: enhancement index γ = 1.5, penalty coefficient δ = 2.0. The AI ​​image enhancement processor first calculates the confidence level of the basic data as 0.1. The prediction enhancement factor was then calculated as follows: The result is approximately 2.415. Next, the penalty factor for control precision is calculated as follows: Finally, multiplying these three parts together gives the final confidence weight of the pixel. The calculated value of 0.765 will be used in subsequent neural network enhancement steps.

[0026] After constructing the corresponding blur kernel map and confidence weight map for the original X-ray image, the AI ​​image enhancement processor performs the final step: neural network enhancement based on multi-channel input. This step first prepares the original X-ray image, blur kernel map, and confidence weight map. The AI ​​image enhancement processor stacks these three identical, pixel-by-pixel two-dimensional data maps in the data dimension to form an input tensor with three channels. The first channel of this input tensor stores the grayscale values ​​of the original X-ray image, the second channel stores information about the local blur kernel (for example, a two-dimensional kernel can be flattened into a one-dimensional vector as the pixel value of that channel), and the third channel stores the calculated confidence weight values.

[0027] Subsequently, the prepared three-channel input tensor is fed into a pre-trained multi-channel AI neural network model. In one specific embodiment, the architecture of this multi-channel AI neural network model can be an improved U-Net network. The encoder portion of this U-Net network has its first convolutional layer's input channel count modified to 3 to receive the aforementioned three-channel input tensor. The rest of the network, including the encoder's downsampling path, the decoder's upsampling path, and the skip connections between them, can maintain the standard U-Net structure. This multi-channel AI neural network model is trained with a large number of training samples, with the goal of optimally reconstructing a clear, unblurred, and noise-free original image when receiving a three-channel input containing physical prior information.

[0028] Ultimately, the multi-channel AI neural network model performs image reconstruction and output based on multi-dimensional prior guidance. The input three-channel tensor undergoes forward propagation within the network. During feature extraction and fusion, information from the blur kernel map channel provides direct guidance for deconvolutional layers or similar operations in the network on how to specifically eliminate local blur. Simultaneously, information from the confidence weight map channel can be used by the network to weight features from different regions, making the network focus more on features extracted from pixel regions with high confidence and suppressing the influence of features from low-confidence regions on the final result. After a complete encoding-decoding process, the network's final output layer outputs a single-channel two-dimensional image data with the same size as the original image. This output image data is the X-ray image ultimately generated in this embodiment of the invention, which has undergone precise restoration and enhancement based on physical process information.

[0029] The embodiments described above are only some embodiments of the present invention, and not all embodiments. It should be noted that the present invention is not limited to the embodiments described above. Any modifications, equivalent substitutions, variations, or improvements made by those skilled in the art based on the above embodiments without creative effort, within the scope of the concept and principles of the present invention, should fall within the protection scope claimed by the claims of the present invention.

Claims

1. A portable X-ray backscatter security system based on AI image enhancement, characterized in that, The system comprises a processor and a memory, the memory storing a computer program, the computer program executed by the processor comprising: a perception unit configured to perceive physical context information in an imaging process to generate time-sequenced physical prior data; a prediction unit configured to predict future scanning behavior based on the physical prior data to generate feedforward control instructions and a prediction confidence; a regulation unit configured to adaptively regulate scanning parameters of an imaging system based on the feedforward control instructions, and record execution information of the regulation, the execution information comprising a regulation residual; an AI reconstruction processor configured to enhance reconstruction of raw X-ray images acquired by an X-ray detector based on the physical prior data, the prediction confidence, and the execution information to generate a final X-ray image.

2. The portable X-ray backscatter security system based on AI image enhancement according to claim 1, wherein, The perception unit is specifically configured to: acquire real-time motion vectors of the imaging system; real-time analyze X-ray signal streams acquired by the X-ray detector to acquire signal variances representing local information density of a scanning target; and synchronize the real-time motion vectors and the signal variances with corresponding X-ray signals in time stamps to generate the physical prior data. 3.The portable X-ray backscatter security inspection system based on AI image enhancement of claim 1, wherein, The regulation unit is specifically configured to: adjust dwell time in a scanning process based on the feedforward control instructions and the signal variances; and determine the regulation residual by recording differences between target values of the feedforward control instructions and actual values after execution of the imaging system. 4.The portable X-ray backscatter security inspection system based on AI image enhancement of claim 1, wherein, The AI reconstruction processor is specifically configured to calculate corresponding confidence weights for pixels in the raw X-ray images.

5. The portable X-ray backscatter security system based on AI image enhancement according to claim 4, characterized in that, The AI reconstruction processor takes the real-time motion vectors corresponding to the pixels, the prediction confidence corresponding to the pixels, and the regulation residual corresponding to the pixels as calculation inputs to determine the confidence weights when calculating the confidence weights.

6. The portable X-ray backscatter security system based on AI image enhancement according to claim 5, wherein, The AI reconstruction processor is configured to calculate the confidence weight (C) by a non-linear coupling model as follows: ): ； wherein, is a data saturation based on scan dwell time computation, is a context distortion based on the real-time motion vector (M) ) computation, is the prediction confidence, is the regulation residual, and γ and δ are preset weight factors.

7. The portable X-ray backscatter security system based on AI image enhancement according to claim 1, wherein, The AI reconstruction processor is further configured to: construct corresponding local blur kernels for pixels in the raw X-ray images according to the real-time motion vectors.

8. The portable X-ray backscatter security system based on AI image enhancement according to claim 4 and 7, characterized in that, The AI reconstruction processor comprises: a multi-channel AI neural network model; The AI reconstruction processor further takes the raw X-ray images, a blur kernel graph composed of the local blur kernels, and a confidence weight graph composed of the confidence weights as inputs, and inputs them into the multi-channel AI neural network model to generate the final X-ray image.

9. The portable X-ray backscatter security system based on AI image enhancement according to claim 2 or 3, characterized in that, The system further comprises an intelligent imaging engine module, and the perception unit, the prediction unit, and the regulation unit are integrated in the intelligent imaging engine module.

Images