A security check image remote judgment task adaptive allocation method and system

Abstract

The present application belongs to the technical field of security check image distributed processing, and relates to a security check image remote image judgment task adaptive allocation method and system, comprising: extracting pixel gradient distribution and edge overlap degree features of to-be-processed security check image data; using static orthogonal mask matrix convolution dimension reduction to generate a path prediction vector representing logical branch trigger probability; receiving hardware running stream activity data reported by a distributed computing node, updating a cache hotspot distribution vector reflecting processor instruction cache hot state distribution; calculating a matching score of the path prediction vector and the hotspot distribution vector, determining a target node in combination with node real-time load parameters and distributing image data, the present application establishes an association mechanism of image features and processor instruction cache prefetch hit rate, avoids instruction path oscillation and cache page exchange loss in the process of heterogeneous data processing, guarantees hardware running continuity, and improves the limit throughput efficiency of the distributed system.

Description

Technical Field

[0001] This invention belongs to the field of distributed processing technology for security inspection images, and relates to an adaptive allocation method and system for remote image judgment tasks of security inspection images. It is used for remote image distribution and processing in high-traffic security inspection scenarios such as airports or transportation hubs, and can improve the operational continuity and throughput of distributed systems when processing heterogeneous image data. Background Technology

[0002] Currently, in the field of distributed systems, achieving balanced task allocation and efficient scheduling is key to improving overall system performance. However, existing technologies still have limitations in handling inter-node information consistency and real-time hardware resource feedback. First, decentralized scheduling technology based on synchronization mechanisms, such as the Chinese invention patent CN114257586B which discloses a distributed task scheduling method, achieves node self-scheduling by integrating scheduling algorithms into task modules and ensures decision consistency by using a unified scheduling execution time and basis time. The drawback of this scheme is that it mainly relies on preset time intervals for state synchronization. In the case of severe fluctuations in network links or sudden bursts of high-frequency tasks, fixed-period information interaction is difficult to align the task queue status of all nodes in real time, which can easily lead to scheduling decision conflicts or response delays.

[0003] Second, there is the data feature analysis-based traffic splitting and scheduling technology. For example, Chinese invention patent CN114359023B discloses a method for splitting and scheduling images to the center based on complexity. This method obtains the complexity of the image to be identified by analyzing it and then makes a traffic splitting decision accordingly. The drawback of this scheme is that it focuses on the static feature extraction of the data content itself and fails to establish a closed-loop relationship between data complexity and the real-time computing resource carrying capacity of the nodes. In scenarios with multiple concurrent nodes and dynamic changes in computing load, due to the lack of awareness of hardware layer performance indicators, it is difficult to achieve a physical match between computing pressure and processing capacity, resulting in uneven resource allocation.

[0004] Third, there is the task migration technology based on hardware performance monitoring. For example, Chinese invention patent application CN121722575A discloses a performance monitoring-driven heterogeneous computing task hardware scheduling system, which uses hardware performance index data formatted into feature vectors to drive task migration. The drawback of this solution is that it is a low-level pure hardware scheduling mechanism. Although it improves the scheduling accuracy at the surface level, it lacks the overall perception of the global task topology and business logic in a large-scale distributed network. At the same time, since this solution is highly dependent on a specific instruction set architecture and hardware monitoring registers, it has limited cross-platform compatibility in complex heterogeneous distributed environments.

[0005] Therefore, the technical problem to be solved by this invention is how to construct a predictive scheduling architecture with real-time global task status awareness and closed-loop feedback of underlying hardware load, so that the task distribution strategy can dynamically converge to the real resource security boundary while ensuring system information consistency. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art, solve the problem of computing power loss caused by software and hardware mismatch in distributed image analysis tasks, and provide an adaptive allocation method for remote image analysis tasks of security inspection images.

[0007] To achieve the above-mentioned objectives, the present invention provides an adaptive allocation method for remote image analysis tasks in security inspection, comprising the following steps: Step 101, Obtain security inspection image data to be processed: Obtain security inspection image data of features to be processed; Step 102, Dimensionality reduction of image features to generate path prediction vector: Extract the pixel gradient distribution and edge overlap features of the security inspection image data, and use a static orthogonal mask matrix to perform convolution dimensionality reduction on the pixel gradient distribution and edge overlap features to generate a computational path prediction vector composed of multiple components, where each component represents the probability of the security inspection image data triggering each preset logical branch in the remote image judgment core algorithm. Step 103, Update the cache hotspot vector based on activity data: Receive the heartbeat data packet sent by the distributed computing node, extract the running flow activity data collected based on the hardware performance counter in the heartbeat data packet, and update the cache hotspot distribution vector that represents the hot distribution of the processor instruction cache in the distributed computing node. Step 104, calculate the matching score between the prediction vector and the hotspot vector: perform a dot product operation on the prediction vector of the computing path and the cache hotspot distribution vector to calculate the cache matching score between the security inspection image data and the distributed computing node, so as to quantify the prefetch hit rate of the processor instruction cache on the security inspection image data processing path; Step 105, Selecting nodes and distributing images based on comprehensive cost: Calculate the allocation cost based on the cache matching score, the real-time load rate of the node, and the depth of the suspended task queue, select the computing node with the lowest allocation cost as the target node, and generate routing instructions to distribute the security inspection image data to the target node. Step 106, if resource rejection is encountered, cancel the distribution and reduce the weight: In response to the resource rejection signal returned by the target node, cancel the routing instruction within 10ms to 50ms, and reduce the allocation weight of the target node in the next scheduling cycle by a preset step size.

[0008] Step 102 of the present invention further includes the following sub-steps: Step 1021, calculate the local information entropy of each pixel block in the security inspection image data, and construct a spatial information gradient map representing the gradient distribution of the pixel; Step 1022, input the spatial information gradient map into the path prediction model, and use the path prediction model to predict the triggering frequency of each jump instruction when the remote image judgment core algorithm processes the security inspection image data, thereby generating the calculated path prediction vector.

[0009] Step 103 of the present invention further includes the following sub-steps: Step 1031, obtaining the instruction branch jump trajectory of the processor within the current sliding window; Step 1032, counting the number of hits of each instruction branch and filtering out active instruction branches with a hit frequency exceeding a preset threshold; Step 1033, updating the corresponding dimension component in the cache hotspot distribution vector according to the position of the active instruction branch in the logical topology of the remote graph judgment core algorithm, so as to track the instruction resident status of the processor's first-level instruction cache and second-level instruction cache in real time.

[0010] Step 105 of the present invention further includes the following sub-steps: Step 1051, obtaining the real-time load rate, memory usage rate, and suspended task queue depth of each computing node in the candidate node set; Step 1052, constructing a multi-dimensional cost function including resource cost item, latency cost item, and instruction cache mismatch cost item; Step 1053, determining that the instruction cache mismatch cost item is negatively correlated with the cache matching score, calculating the comprehensive cost value of each computing node, and selecting the node with the smallest comprehensive cost value as the target node.

[0011] Before distributing the security check image data as described in this invention, step 101 further includes: obtaining the passenger boarding time slack parameter captured by the front-end gate system; converting the boarding time slack parameter into a task priority weight and binding it to the header of the security check image data, so that the scheduling order in step 105 is dynamically arranged according to the task priority weight.

[0012] The process of extracting the spatial information gradient map in step 1021 of the present invention is as follows: using an orthogonal mask matrix to perform convolution operation on the brightness component of the security inspection image data to obtain edge saliency features, and determining the spatial feature complexity of the security inspection image data based on the histogram distribution of the edge saliency features.

[0013] In step 1032 of this invention, a decay factor is introduced when counting the number of hits; the value of the decay factor decreases as the interval between the recording time and the current time increases, so that the cache hotspot distribution vector accurately reflects the instruction cache hotspot distribution of the processor at the current time.

[0014] The distributed computing nodes described in this invention are distributed in a heterogeneous resource pool, and different data acquisition frequencies for step 103 are set for computing nodes with different processor architectures; the processor architecture includes reduced instruction set architecture and complex instruction set architecture.

[0015] The present invention further includes a self-correction step after distributing the security inspection image data: obtaining the actual instruction branch record triggered when the target node processes the security inspection image data; calculating the comparison residual between the actual triggered instruction branch record and the calculated path prediction vector; and correcting the neuron weights of the path prediction model based on the comparison residual to optimize the accuracy of subsequent task allocation.

[0016] This invention also provides a system for an adaptive allocation method of remote image analysis tasks for security inspection images, which is used to implement an adaptive allocation method of remote image analysis tasks for security inspection images, including: The feature extraction unit is used to acquire security inspection image data of the features to be processed, extract the pixel gradient distribution and edge overlap features of the security inspection image data, and use a static orthogonal mask matrix to perform convolution dimensionality reduction on the pixel gradient distribution and edge overlap features to generate a computational path prediction vector composed of multiple components; wherein, each component is used to characterize the probability of the security inspection image data triggering each preset logical branch in the remote image judgment core algorithm; The state awareness unit is used to receive heartbeat data packets sent by distributed computing nodes, extract the running flow activity data collected based on hardware performance counters in the heartbeat data packets, and update the cache hot spot distribution vector that represents the hot state distribution of processor instruction cache within the distributed computing node. The matching calculation unit is used to perform a dot product operation between the calculation path prediction vector and the cache hotspot distribution vector to calculate the cache matching score between the security inspection image data and the distributed computing node, so as to quantify the prefetch hit degree of the processor instruction cache for the security inspection image data processing path; The task distribution unit is used to calculate the allocation cost based on the cache matching score, the real-time load rate of the distributed computing node, and the depth of the suspended task queue of the distributed computing node, select the computing node with the lowest allocation cost as the target node, and generate routing instructions to distribute the security inspection image data to the target node. The feedback adjustment unit is used to respond to the resource rejection signal returned by the target node, cancel the routing instruction within 10ms to 50ms, and reduce the allocation weight of the target node in the next scheduling cycle by a preset step size.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: 1. In the adaptive allocation of remote image analysis tasks for security inspection images, the pixel gradient distribution of security inspection image data is converted into a logical branch probability vector and matched with the instruction dwell vector of the computing node. This enables data packets with similar underlying processing logic to be continuously and directionally distributed to nodes in the corresponding instruction hot cache state, reducing the paging operations of the CPU's first-level instruction cache and second-level instruction cache, and maintaining the processing continuity of the computing node at the surface level.

[0018] 2. By using feature transformation operators to predict the mathematical probability of triggering each logical branch after the data packet enters the parsing algorithm, a monotonically positive correlation mapping between data flow features and processor branch prediction hit rate is established. This avoids branch prediction failure caused by random allocation of heterogeneous data, thereby eliminating instruction idle time in the processor pipeline and increasing the number of effective instructions processed in a single clock cycle.

[0019] 3. By combining the real-time load rate of computing nodes, the depth of suspended task queues, and the affinity of instruction paths, the routing allocation strategy is adjusted in multiple dimensions through a cost estimation function. This transforms data scheduling from simply matching overall computing power to collaborative scheduling that fits the surface characteristics of the computing architecture. Without changing the physical hardware configuration, the overall throughput of the distributed processing system approaches the sum of the rated computing power of each computing node. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the adaptive allocation process for security inspection image tasks in this invention; Figure 2 This is a schematic diagram of the system interaction architecture between the remote map-based scheduling center and the computing nodes of this invention. Detailed Implementation

[0021] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings.

[0022] Example 1: This embodiment discloses an adaptive allocation method for remote image analysis tasks in security inspections, including the following steps: Step 101, Obtain security inspection image data to be processed: Obtain security inspection image data of features to be processed; Step 102, Dimensionality reduction of image features to generate path prediction vector: Extract the pixel gradient distribution and edge overlap features of the security inspection image data, and use a static orthogonal mask matrix to perform convolution dimensionality reduction on the pixel gradient distribution and edge overlap features to generate a computational path prediction vector composed of multiple components, where each component represents the probability of the security inspection image data triggering each preset logical branch in the remote image judgment core algorithm. Step 103, Update the cache hotspot vector based on activity data: Receive the heartbeat data packet sent by the distributed computing node, extract the running flow activity data collected based on the hardware performance counter in the heartbeat data packet, and update the cache hotspot distribution vector that represents the hot distribution of the processor instruction cache in the distributed computing node. Step 104, calculate the matching score between the prediction vector and the hotspot vector: perform a dot product operation on the prediction vector of the computing path and the cache hotspot distribution vector to calculate the cache matching score between the security inspection image data and the distributed computing node, so as to quantify the prefetch hit rate of the processor instruction cache on the security inspection image data processing path; Step 105, Selecting nodes and distributing images based on comprehensive cost: Calculate the allocation cost based on the cache matching score, the real-time load rate of the node, and the depth of the suspended task queue, select the computing node with the lowest allocation cost as the target node, and generate routing instructions to distribute the security inspection image data to the target node. Step 106, if resource rejection is encountered, cancel the distribution and reduce the weight: In response to the resource rejection signal returned by the target node, cancel the routing instruction within 10ms to 50ms, and reduce the allocation weight of the target node in the next scheduling cycle by a preset step size.

[0023] Step 102 in this embodiment further includes the following sub-steps: Step 1021, calculate the local information entropy of each pixel block in the security inspection image data, and construct a spatial information gradient map representing the gradient distribution of the pixel; Step 1022, input the spatial information gradient map into the path prediction model, and use the path prediction model to predict the trigger frequency of each jump instruction when the remote image judgment core algorithm processes the security inspection image data, thereby generating the calculated path prediction vector.

[0024] Step 103 in this embodiment further includes the following sub-steps: Step 1031, obtaining the instruction branch jump trajectory of the processor within the current sliding window; Step 1032, counting the number of hits of each instruction branch and filtering out active instruction branches with a hit frequency exceeding a preset threshold; Step 1033, updating the corresponding dimension component in the cache hotspot distribution vector according to the position of the active instruction branch in the logical topology of the remote graph judgment core algorithm, so as to track the instruction resident status of the processor's first-level instruction cache and second-level instruction cache in real time.

[0025] Step 105 in this embodiment further includes the following sub-steps: Step 1051, obtaining the real-time load rate, memory usage rate, and suspended task queue depth of each computing node in the candidate node set; Step 1052, constructing a multi-dimensional cost function that includes resource cost items, latency cost items, and instruction cache mismatch cost items; Step 1053, determining that the instruction cache mismatch cost item is negatively correlated with the cache matching score, calculating the comprehensive cost value of each computing node, and selecting the node with the smallest comprehensive cost value as the target node.

[0026] Before distributing the security check image data as described in this embodiment, step 101 further includes: obtaining the passenger boarding time slack parameter captured by the front-end gate system; converting the boarding time slack parameter into a task priority weight and binding it to the header of the security check image data, so that the scheduling order in step 105 is dynamically arranged according to the task priority weight.

[0027] The process of extracting the spatial information gradient map in step 1021 of this embodiment is as follows: using an orthogonal mask matrix to perform convolution operation on the brightness component of the security inspection image data to obtain edge saliency features, and determining the spatial feature complexity of the security inspection image data based on the histogram distribution of the edge saliency features.

[0028] In step 1032 of this embodiment, a decay factor is introduced when counting the number of hits. The value of the decay factor decreases as the interval between the recording time and the current time increases, so that the cache hotspot distribution vector accurately reflects the instruction cache hotspot distribution of the processor at the current time.

[0029] The distributed computing nodes described in this embodiment are distributed in a heterogeneous resource pool, and different data acquisition frequencies for step 103 are set for computing nodes with different processor architectures; the processor architecture includes reduced instruction set architecture and complex instruction set architecture.

[0030] This embodiment further includes a self-correction step after distributing the security inspection image data: obtaining the actual instruction branch record triggered when the target node processes the security inspection image data; calculating the comparison residual between the actual triggered instruction branch record and the calculated path prediction vector; and correcting the neuron weights of the path prediction model based on the comparison residual to optimize the accuracy of subsequent task allocation.

[0031] This embodiment also provides a system for an adaptive allocation method of remote image analysis tasks for security inspection, including: The feature extraction unit is used to acquire security inspection image data of the features to be processed, extract the pixel gradient distribution and edge overlap features of the security inspection image data, and use a static orthogonal mask matrix to perform convolution dimensionality reduction on the pixel gradient distribution and edge overlap features to generate a computational path prediction vector composed of multiple components; wherein, each component is used to characterize the probability of the security inspection image data triggering each preset logical branch in the remote image judgment core algorithm; The state awareness unit is used to receive heartbeat data packets sent by distributed computing nodes, extract the running flow activity data collected based on hardware performance counters in the heartbeat data packets, and update the cache hot spot distribution vector that represents the hot state distribution of processor instruction cache within the distributed computing node. The matching calculation unit is used to perform a dot product operation between the calculation path prediction vector and the cache hotspot distribution vector to calculate the cache matching score between the security inspection image data and the distributed computing node, so as to quantify the prefetch hit degree of the processor instruction cache for the security inspection image data processing path; The task distribution unit is used to calculate the allocation cost based on the cache matching score, the real-time load rate of the distributed computing node, and the depth of the suspended task queue of the distributed computing node, select the computing node with the lowest allocation cost as the target node, and generate routing instructions to distribute the security inspection image data to the target node. The feedback adjustment unit is used to respond to the resource rejection signal returned by the target node, cancel the routing instruction within 10ms to 50ms, and reduce the allocation weight of the target node in the next scheduling cycle by a preset step size.

[0032] Example 2: This embodiment relates to a security image distribution scenario handling high-concurrency traffic at airport security checkpoints. The system receives security image data collected from multiple security checkpoints. This data includes complex images composed of stacked metal objects and simple images composed of clothing. This data enters the data stream to be processed, causing fluctuations in the branch prediction accuracy of the central processing unit (CPU). This fluctuation stems from differences in the logical paths triggered by images of varying complexity in the remote image judgment core algorithm, leading to CPU instruction cache paging overhead. Before feature acquisition, the system uses the Cannibal edge detection operator to process the security images, generating binary edge images. These binary images are then traversed through a set grid window. Within each grid window, the number of intersecting edge pixels and the intersection density of multi-directional gradient vectors are calculated. The normalized absolute value of this intersection density is used as the edge overlap feature characterizing the complexity of the stacked items within the package. The system extracts the pixel gradient distribution and edge overlap features from the security image data. A static orthogonal mask matrix is ​​used to perform convolutional dimensionality reduction on the pixel gradient distribution and edge overlap features, generating a computational path prediction vector composed of multiple components. Each component represents the probability of security inspection image data triggering each preset logical branch in the remote image judgment core algorithm. Simultaneously, it receives heartbeat data packets sent by distributed computing nodes, extracts runtime activity data collected based on hardware performance counters, and updates the cache hotspot distribution vector, which represents the hot state distribution of processor instruction caches within the distributed computing nodes. The cache hotspot distribution vector It reflects the residency status of instructions corresponding to each preset logical branch within the processor in the instruction cache.

[0033] The scheduling center calculates the path prediction vector. With cache hotspot distribution vector The dot product is used to determine the cache fit score, which reflects the degree of matching between data characteristics and the processor instruction cache. The scheduling center obtains the real-time load rate and suspended task queue depth of each computing node in the candidate node set, and matches the results based on the cache score. The real-time load rate and the depth of the suspended task queue are used to calculate the comprehensive cost of each computing node according to the following formula: ,in, This is the comprehensive cost value; Cache matching scores; Real-time load rate; To determine the depth of the suspended task queue; , , These are the preset weighting coefficients.

[0034] The system selects the computation node with the minimum overall cost as the target node. Under this condition, it has a similar computation path prediction vector. Security inspection image data is continuously and directionally distributed to the current cache hotspot distribution vector. The corresponding distributed computing nodes maintain a hot-state distribution of processor instruction caches, reducing clock cycle wastage caused by instruction paging. When the target node returns a resource rejection signal, the scheduling center cancels the routing instruction within 10ms to 50ms. The process of canceling the routing instruction includes obtaining the physical flow location of security inspection image data within the distributed scheduling system. When the security inspection image data remains in the application layer sending queue of the scheduling center and has not yet been written to the underlying buffer of the transmission control protocol, the scheduling center releases the data packet by directly erasing the memory pointer of the corresponding queue. When it is determined that the security inspection image data has been sent to the underlying software-defined network switch, the scheduling center issues the highest priority directed discard flow table entry to the switch controller, intercepting and discarding the security inspection image data packet at the physical link end through hardware-level actions. After completing the interception, the scheduling center lowers the allocation weight of the target node in the next scheduling cycle to prevent backpressure effects caused by local blocking. This scheme achieves a match between data feature complexity and node real-time throughput capabilities, improving the overall throughput performance of the distributed system.

[0035] Example 3: This embodiment uses a cluster of 16 distributed computing nodes in the current experimental environment. Each distributed computing node uses a processor with a main frequency of 2.5 GHz and is equipped with 32 KB of L1 instruction cache and 256 KB of L2 instruction cache. The experimental data comes from a pre-calibrated security inspection image sample set, which contains 50,000 X-ray security inspection images. By calculating the pixel gradient distribution and edge overlap characteristics of the security inspection image data, the sample set is divided into three gradients: simple, normal, and complex. To simulate non-ideal network conditions, transmission latency fluctuations following a Gaussian distribution are injected into the network interface of the distributed computing nodes, and each distributed computing node is pre-loaded with 15% to 25% random background computing load. The experiment sets up a comparison sample group 1, a comparison sample group 2, and the sample group of the present invention. The comparison sample group 1 uses a round-robin allocation method, the comparison sample group 2 uses a dynamic allocation method based on the depth of the suspended task queue, and the sample group of the present invention uses a method based on the computation path prediction vector. With cache hotspot distribution vector The affinity matching allocation method, where the weight coefficients The value of is determined by balancing instruction cache affinity gain and node load balancing, when the weight coefficient When the value is in the range of 0.5 to 0.8, the processor instruction cache hit rate and the variance of task distribution among nodes tend to reach a steady state. When processing a regular gradient image stream, the L2 instruction cache miss rate of the processor in comparison sample group 1 is 8.45%, while that of comparison sample group 2 is 7.21%. The present invention's sample group uses cache matching scores... Guiding data flow to enable prediction vectors with similar computational paths The security inspection image data is continuously fed into the distributed computing node with corresponding instructions residing in a hot state. Its secondary instruction cache miss rate is 1.42%. The decrease in the secondary instruction cache miss rate results in an increase in system throughput. The number of image processing frames per unit time of the sample group of this invention is 22.5% higher than that of the comparative sample group 2.

[0036] When verifying the parameter boundary conditions, adjust the real-time load rate in the comprehensive cost value formula. Influence weight When weight When the value is less than 0.2, the system distributes tasks to nodes with high cache matching more frequently, leading to an increase in the depth of the suspended task queue. Exceeding a preset threshold triggers a resource rejection signal from a local node, when the weight... When the threshold exceeds 1.5, the load balancing logic's intervention in scheduling instructions increases, and the system's secondary instruction cache miss rate rises to over 6.5%. Experimental data shows that as the information entropy gradient of security inspection image data increases, the performance gain generated by the sample group of this invention exhibits non-linear growth. When processing complex gradient images containing layered metal features, the average single-frame parsing time is reduced from 1526ms in the second comparative sample group to 118.4ms. The above data confirms that vectorizing and aligning the data logic branch probability with the hardware cache hot state distribution avoids instruction path oscillations during heterogeneous data processing. Under network fluctuations and load disturbances, this solution establishes a deterministic mapping from pixel gradient features to surface execution state, eliminating computational power loss caused by software and hardware state mismatch and improving the processing efficiency of the distributed system.

[0037] Example 4: In this embodiment, within a civil aviation security image interpretation cluster operating environment capable of accessing multi-dimensional business data, the dispatch center allocates remote image interpretation tasks for security images in the following manner by acquiring passenger flow distribution across different security checkpoints; during the offline calibration phase, the system constructs a static orthogonal mask matrix for feature extraction, which is composed of... The pixel gradient template is constructed from a set of mutually orthogonal basis vectors. The gradient is then transformed using the Schmitt orthogonal transformation method to determine the gradient that satisfies... The distribution of mask coefficients, where, For the mask matrix, The identity matrix is ​​used. During real-time processing, the system extracts the pixel gradient distribution of the security inspection image data, divides it into 16×16 feature blocks, calculates the multiplication and addition results for each feature block using a static orthogonal mask matrix, projects the high-dimensional features onto a low-dimensional manifold space, and generates a path prediction vector. ,in, To compute the path prediction vector, the hardware performance counters within the distributed computing nodes sample the instruction cache hit status every 100ms. The scheduling center receives runtime activity data sent by the distributed computing nodes and updates the cache hotspot distribution vector accordingly. Then, an exponential decay model based on timestamps is introduced, and the calculation formula is as follows: ,in, Cache the th element in the hotspot distribution vector at the current time. The values ​​of each component; The first digit of the cached hotspot distribution vector from the previous time step. The values ​​of each component; The preset attenuation factor has a value range of 0.05 to 0.12; To record time intervals; For the current sampling period, the first The system converts discrete hardware sampling pulses into a continuous instruction resident state space using an exponential decay model, eliminating scheduling deviations caused by interference from historical outdated states. The decay factor and temperature coefficient ranges are determined through a balance calculation between system throughput and cache replacement cycle. If the decay factor is below 0.05, state convergence is sluggish, leading to the misjudgment of invalid instruction caches as hotspots. If it is above 0.12, historical states are forgotten too quickly, easily causing severe oscillations in the state vector due to instantaneous load fluctuations. Similarly, the temperature coefficient in subsequent normalization operations, if below 1.5, leads to severe polarization of trigger probability, causing nodes to degenerate into single-branch processors; if above 3.0, it excessively smooths feature differences, completely losing its guiding and distribution function for heterogeneous computing power. When processing related business data, the system obtains passenger boarding time slack parameters captured by the front-end gate system, calculates a one-way hash value for the passenger's unique identifier at the data entry point, generates anonymized association tags, and extracts the desensitized boarding time slack parameters. And according to the preset time gradient mapping table, the boarding time slack parameter is... Convert to dimensionless task priority weights The scheduling center will assign task priority weights. The header encapsulation bound to security inspection image data participates in the overall cost value. Dynamic sorting, when the boarding time slack parameter When the time is less than 20 minutes, the corresponding security inspection image task is prioritized for allocation to the cache matching score. Computational nodes with a value greater than 0.85.

[0038] The above scheme reduces fluctuations in the data processing path by orthogonalizing the feature reduction operator and quantifying the decay law of the cache vector. When processing anonymized passenger business data, the system achieves physical alignment between business priority logic and hardware instruction execution status, reducing system latency under high concurrency and ensuring the stability of distributed image analysis task distribution. In an engineering example of performing pixel-level complexity analysis on security inspection image data, the system performs convolution operations on the brightness components of the image using a static orthogonal mask matrix to obtain edge saliency features representing contour information, calculates the histogram distribution of these edge saliency features in grayscale space, extracts the information entropy of the histogram distribution and the second moment of the pixel gradient as complexity quantification indicators, and establishes a complexity quantification indicator and a calculation path prediction vector. The mapping relationship between the components is determined by a complex metric calculated using a predefined logical branch probability function. Images with larger pixel gradients correspond to different computational path prediction vectors. The higher the probability of triggering high-overhead jump logic in a component, the higher the probability of triggering the logic branch. The underlying layer of the preset logic branch probability function uses a multilayer perceptron as the path prediction model. This path prediction model includes an input layer for receiving feature coefficient sequences, two fully connected hidden layers with 64 and 32 nodes respectively, and an output layer with the same dimension as the preset total number of logic branches. During the offline training phase, the system extracts a large number of historical security inspection image features with real physical execution trajectory labels and inputs them into the path prediction model. The error between the predicted probability output and the real branch label is calculated using the cross-entropy loss function, and the weights of each layer of neurons in the path prediction model are updated through the backpropagation algorithm until the loss function converges. This enables the model to have the deterministic calculation capability from image dimensionality reduction features to code branch trigger probability. The remote image judgment core algorithm adopts a multi-level cascaded architecture, including a pre-judgment conditional branch based on image complexity and a post-convolutional network that performs parallel matrix operations. The logic branch probability function is based on the principle of normalized exponential function, which converts the real number system feature coefficients output by the static orthogonal mask matrix into a conditional branch trigger probability distribution. By introducing a temperature coefficient, exponential normalization is performed on the dimensionality reduction feature coefficients. The calculation formula is as follows: ,in, Characterizing the first The probability of triggering a logical branch is limited to a value greater than 0 and less than 1, and the sum of the probabilities of each component is 0. , and Characterizes the output eigenvalues ​​of the static orthogonal mask matrix. The temperature coefficient characterizing the smoothness of the adjusted probability distribution is set to a value between 1.5 and 3.0. The total number of preset conditional branches in the remote image judgment core algorithm is represented by the mapping from visual spatial features to program jump probability. The reason why the above mapping holds true at the physical level is that the remote image judgment core algorithm itself is designed with conditional branches based on the physical properties of the image. When the static orthogonal mask matrix extracts spatial feature coefficients with dense high-frequency edge components, it indicates that there are a large number of overlapping or high-density dangerous materials in the current security inspection image. This physical complexity directly corresponds to the deep iterative recognition logic branch in the core algorithm used to peel off interference layer by layer. Therefore, taking the high-amplitude feature coefficients obtained by linear visual transformation as input can objectively reflect the computational depth required for data parsing, and thus uniquely and definitively map it to the high-overhead logic branch trigger probability for processing such complex physical structures, realizing cross-domain equivalence between image physical spatial complexity and code execution trajectory.

[0039] Example 5: In this embodiment, during the parameter calibration scenario, the distributed computing cluster determines the correlation between the L2 instruction cache hit rate and the dispatch response latency based on the physical output collected by the processor hardware performance counter. Under controlled load, the system distributes standard security inspection image data with a fixed pixel gradient distribution and adjusts the attenuation factor in steps of 0.01 within a range of 0.01 to 0.20. Record the distribution vector of cache hotspots under each gradient. The convergence time and corresponding instruction cache prefetch hit rate are considered. When the variance of the prediction hit rate fluctuates less than 0.001 over 500 consecutive processing cycles, the system will adjust the current decay factor. As the operating baseline, the system collects images at different suspended task queue depths by distributing high-information-entropy image streams containing metal overlap features. The corresponding instruction bus conflict frequency is used to determine the overall cost value based on the rate of change of this conflict frequency. The response boundary of each weight coefficient in the equation.

[0040] When compute nodes are deployed to an environment with heterogeneous processor architectures, the scheduling center retrieves the preceding baseline alignment process, which distributes a pre-defined compute path prediction vector to the target node. The system obtains the number of instruction execution cycles under different kernel frequencies from the probe data packets. It then calculates the ratio of the actual parsing time of the probe data packets to the theoretical peak computing power and uses this ratio as a normalization factor in the real-time load rate. The calculation process aims to reduce the asymmetric impact of different hardware specifications on the allocation strategy. The system retrieves weight coefficients within the parameter space. , and The combination of parameters selects a numerical array that keeps the standard deviation of the system's processed frame rate within a preset steady-state range as the operating parameters, and records the comprehensive cost value at the inflection point where the distributed computing node returns a resource rejection signal. The comprehensive cost value The trigger criterion for dynamically adjusting the weight allocation is set as follows: , , As a weighting factor, in real-time sensing scenarios with integrated processor performance monitoring units, distributed computing nodes periodically collect physical counts of Level 1 instruction cache miss events and branch prediction failure events by calling hardware performance counters. The scheduling center receives the count increments from the heartbeat data packets. To separate background resident processes from instruction paging data caused by operating system kernel scheduling in the heterogeneous resource pool, a precise event filtering mechanism based on process unique identifiers is adopted. This mechanism reads hardware performance counter data bound to the remote graph judgment core algorithm process and removes global interference counts that do not match the process identifier. In practice, this filtering mechanism utilizes the extended Berkeley packet filter provided by the operating system kernel or a dedicated performance monitoring underlying system interface to filter specific micro-incident data in kernel mode. The processor monitoring register set is strongly bound to the context structure of the remote graph analysis core algorithm thread. The corresponding hardware performance counters are only activated and incremented when the CPU schedules the target process into the running state. Once a process context switch is triggered and the process is switched out, the system's underlying layer freezes the corresponding counting action. This eliminates interference data from all non-core processes at the physical hardware isolation level. A sliding time window of 5ms to 15ms is selected to perform low-pass filtering on the hit frequency of active instruction branches after interference removal. This filtering eliminates the spatiotemporal scale difference between the microsecond-level processor execution cycle and the millisecond-level network routing delay, extracting instruction cache hotspots with durations greater than the network packet transfer delay. These hotspots are then mapped to a cache hotspot distribution vector. The corresponding dimension, where the runtime activity data is inversely proportional to the branch prediction failure frequency, eliminates random interference caused by operating system kernel paging, making the cache hotspot distribution vector... It can reflect the instruction resident status of the processor's execution pipeline. In order to achieve the accurate reverse mapping of the above-mentioned physical memory instruction jump to the software logic topology, the scheduling center preloads the debug information table file generated when the remote graph judgment core algorithm is compiled. This file statically records the mapping association between the source code line number, logic branch mark and the memory address range of the final generated machine code. When the physical instruction memory address reported by the hardware performance counter is received, the system calculates its absolute address offset in the process virtual address space in kernel mode through the memory page table offset, and performs a binary search comparison in the debug information table to accurately locate the high-level language logic branch node to which the machine code belongs. In this way, the semantically detached low-level hardware physical pulse is effectively transformed into the activity state of a specific branch on the application layer algorithm logic tree.

[0041] Example 6: In this embodiment, during the pre-calibration of a security image analysis terminal, the system acquires 1000 standard security image samples containing metals and liquids. It calculates the discrete cosine transform coefficients of the standard security image samples in the horizontal and vertical directions, selects the top 16 frequency components with the highest energy percentage as the original basis vectors, and processes these original basis vectors using the Schmitt orthogonalization method to generate vectors that satisfy orthogonal constraints. static orthogonal mask matrix ,in, It is a static orthogonal mask matrix. As the identity matrix, this procedure determines the image texture to compute the path prediction vector. Discretization operator for projection; the system establishes cache matching scores based on the total capacity of the secondary instruction cache of the distributed computing nodes. The judgment logic, where To cache matching scores, Affinity distribution threshold is matched to the preset cache, and the distribution threshold is adjusted in steps of 0.05 within the range of 0.1 to 0.9. Monitor the task backlog rate and average processing latency of each distributed computing node under different thresholds, and calculate the average processing latency as a function of the distribution threshold. The distribution threshold is set by the fraction at which the second derivative first shows a positive abrupt change. The upper limit of operation, the scheduling center presets task status self-check steps, when the comprehensive cost value When the difference exceeds three times the preset operating variance, the system stops distributing new tasks to the corresponding distributed computing nodes and updates the cached hotspot distribution vector by calling the heartbeat data packet collection process. ,in This is the comprehensive cost value. To cache hotspot distribution vectors, this calibration procedure establishes the steady-state range of the allocation strategy in a heterogeneous hardware environment.

Claims

1. An adaptive allocation method for remote image analysis tasks in security inspections, characterized in that, Includes the following steps: Step 101: Obtain security inspection image data of the features to be processed; Step 102: Extract the pixel gradient distribution and edge overlap features from the security inspection image data. Use a static orthogonal mask matrix to perform convolutional dimensionality reduction on the pixel gradient distribution and edge overlap features, generating a computational path prediction vector composed of multiple components. The core remote image judgment algorithm employs a multi-level cascaded architecture, including a pre-judgment conditional branch based on image complexity and a post-convolutional network performing parallel matrix operations. The logical branch probability function is based on the principle of normalized exponential function, converting the real-number characteristic coefficients output from the static orthogonal mask matrix into a conditional branch trigger probability distribution. A temperature coefficient is introduced to perform exponential normalization on the dimensionality-reduced characteristic coefficients. The calculation formula is as follows: ,in, Characterizing the first The probability of triggering a logical branch is limited to a value greater than 0 and less than 1, and the sum of the probabilities of each component is 0. , and Characterizes the output eigenvalues ​​of the static orthogonal mask matrix. The temperature coefficient characterizing the smoothness of the adjusted probability distribution is set to a value between 1.5 and 3.

0. The total number of preset prerequisite branches in the core algorithm for remote image analysis; Step 103: Receive the heartbeat data packet sent by the distributed computing node, extract the running flow activity data collected based on the hardware performance counter in the heartbeat data packet, and update the cache hot spot distribution vector that represents the hot state distribution of the processor instruction cache within the distributed computing node. Step 104: Perform a dot product operation on the computation path prediction vector and the cache hotspot distribution vector to calculate the cache matching score between the security inspection image data and the distributed computing nodes, so as to quantify the prefetch hit rate of the processor instruction cache on the security inspection image data processing path. Step 105: Obtain the real-time load rate, memory usage rate, and suspended task queue depth of each computing node in the candidate node set; construct a multi-dimensional cost function that includes resource cost, latency cost, and instruction cache mismatch cost; determine that the instruction cache mismatch cost is negatively correlated with the cache matching score; calculate the comprehensive cost value of each computing node; select the node with the smallest comprehensive cost value as the target node; and generate routing instructions to distribute the security inspection image data to the target node. Step 106: In response to the resource rejection signal returned by the target node, the routing instruction is revoked within 10ms to 50ms, and the allocation weight of the target node in the next scheduling cycle is reduced by a preset step size.

2. The adaptive allocation method for remote image analysis tasks in security inspection according to claim 1, characterized in that, Step 102 further includes the following sub-steps: Step 1021, calculate the local information entropy of each pixel block in the security inspection image data, and construct a spatial information gradient map representing the pixel gradient distribution; Step 1022, input the spatial information gradient map into the path prediction model, and use the path prediction model to predict the triggering frequency of each jump instruction when the remote image judgment core algorithm processes the security inspection image data, thereby generating a calculated path prediction vector.

3. The adaptive allocation method for remote image analysis tasks in security inspection according to claim 1, characterized in that, Step 103 further includes the following sub-steps: Step 1031, obtaining the instruction branch jump trajectory of the processor within the current sliding window; Step 1032, counting the number of hits of each instruction branch, and filtering out active instruction branches whose hit frequency exceeds a preset threshold. Step 1033: Based on the position of the active instruction branch in the logical topology of the remote graph judgment core algorithm, update the corresponding dimension component in the cache hotspot distribution vector to track the instruction resident status of the processor's L1 and L2 instruction caches in real time.

4. The adaptive allocation method for remote image analysis tasks in security inspection according to claim 1, characterized in that, Before distributing the security check image data, step 101 also includes: obtaining the passenger boarding time slack parameter captured by the front-end gate system; converting the boarding time slack parameter into a task priority weight and binding it to the header of the security check image data, so that the scheduling order in step 105 is dynamically arranged according to the task priority weight.

5. The adaptive allocation method for remote image analysis tasks in security inspection according to claim 1, characterized in that, The process of extracting the spatial information gradient map in step 1021 is as follows: an orthogonal mask matrix is ​​used to perform convolution operation on the brightness component of the security inspection image data to obtain edge saliency features, and the spatial feature complexity of the security inspection image data is determined based on the histogram distribution of the edge saliency features.

6. The adaptive allocation method for remote image analysis tasks in security inspection according to claim 3, characterized in that, In step 1032, a decay factor is introduced when counting the number of hits; The value of the decay factor decreases as the interval between the recording time and the current time increases, so that the cache hotspot distribution vector accurately reflects the instruction cache hotspot distribution of the processor at the current time.

7. The adaptive allocation method for remote image analysis tasks in security inspection according to claim 1, characterized in that, The distributed computing nodes are distributed in a heterogeneous resource pool, and different data acquisition frequencies for step 103 are set for computing nodes with different processor architectures; the processor architectures include reduced instruction set architecture and complex instruction set architecture.

8. The adaptive allocation method for remote image analysis tasks in security inspection according to claim 2, characterized in that, After distributing the security inspection image data, a self-correction step is also included: obtaining the actual instruction branch records triggered when the target node processes the security inspection image data; calculating the comparison residual between the actual triggered instruction branch records and the calculated path prediction vector; and correcting the neuron weights of the path prediction model based on the comparison residual to optimize the accuracy of subsequent task allocation.

9. A system for an adaptive allocation method of remote image analysis tasks for security inspection images, used to implement the adaptive allocation method of remote image analysis tasks for security inspection images as described in claim 1, characterized in that, include: The feature extraction unit acquires security inspection image data to be processed, extracts pixel gradient distribution and edge overlap features from the security inspection image data, and performs convolutional dimensionality reduction on the pixel gradient distribution and edge overlap features using a static orthogonal mask matrix to generate a computational path prediction vector composed of multiple components. The remote image judgment core algorithm adopts a multi-level cascaded architecture, including a pre-judgment conditional branch based on image complexity and a post-convolutional network performing parallel matrix operations. The logical branch probability function is based on the principle of normalized exponential function, converting the real-number feature coefficients output by the static orthogonal mask matrix into a conditional branch trigger probability distribution. A temperature coefficient is introduced to perform exponential normalization on the dimensionality-reduced feature coefficients; the calculation formula is as follows: ,in, Characterizing the first The probability of triggering a logical branch is limited to a value greater than 0 and less than 1, and the sum of the probabilities of each component is 0. , and Characterizes the output eigenvalues ​​of the static orthogonal mask matrix. The temperature coefficient characterizing the smoothness of the adjusted probability distribution is set to a value between 1.5 and 3.

0. The total number of preset prerequisite branches in the core algorithm for remote image analysis; The state awareness unit is used to receive heartbeat data packets sent by distributed computing nodes, extract the running flow activity data collected based on hardware performance counters in the heartbeat data packets, and update the cache hot spot distribution vector that represents the hot state distribution of processor instruction cache within the distributed computing nodes. The matching calculation unit is used to perform a dot product operation on the calculation path prediction vector and the cache hotspot distribution vector to calculate the cache matching score between the security inspection image data and the distributed computing nodes, so as to quantify the prefetch hit degree of the processor instruction cache on the security inspection image data processing path; The task distribution unit is used to obtain the real-time load rate, memory usage rate and suspended task queue depth of each computing node in the candidate node set, construct a multi-dimensional cost function including resource cost item, latency cost item and instruction cache mismatch cost item, determine that the instruction cache mismatch cost item is negatively correlated with the cache matching score, calculate the comprehensive cost value of each computing node, select the node with the smallest comprehensive cost value as the target node, and generate routing instructions to distribute the security inspection image data to the target node. The feedback adjustment unit is used to revoke the routing instruction within 10ms to 50ms in response to the resource rejection signal returned by the target node, and to reduce the allocation weight of the target node in the next scheduling cycle by a preset step size.

Images