Video object detection method based on deep neural network architecture

Abstract

The application relates to the technical field of computer system and deep learning model logic control, and discloses a video target detection method based on a deep neural network architecture, which comprises the following steps: acquiring an input tensor at a sampling moment and calling a reference feature tensor stored by a hidden state feature resident unit by a feature extraction branch; calculating a residual tensor and extracting a perception energy deviation therefrom; collecting periodic working condition data and extracting a dominant frequency of the perception energy deviation through a phase observer, thereby anchoring a phase and calculating a perception energy gating threshold fluctuating with a physical beat; comparing the perception energy deviation with the perception energy gating threshold, and controlling a calculation route to switch between a tensor increment evolution path and a logic mute path; and the application reconstructs full-amount reasoning into an increment evolution mode to reduce a calculation load, shields periodic environmental interference by using a time axis phase locking mechanism, and improves reasoning stability of target detection logic under complex industrial working conditions.

Description

Technical Field

[0001] This invention relates to a video target detection method based on a deep neural network architecture, belonging to the field of computer systems and deep learning model logic control technology. Background Technology

[0002] Current video object detection methods based on deep neural network architectures are used to achieve automated production quality inspection. From the perspective of the operation mechanism of the computing model, the mainstream computing architecture adopts a static independent inference method, which splits the continuous video stream into isolated image frames and performs full feedforward operations on each frame. This static computing logic ignores the inherent correlation of neural network computing models when processing time-series signals, causing the underlying computing graph of the computer system to repeatedly execute a large number of redundant convolution operators, resulting in ineffective scheduling and allocation of computing resources within the computing architecture, and affecting the execution performance of specific computing models in real-time systems.

[0003] In precision manufacturing scenarios such as continuous metal rolling or high-speed stamping, there is pixel redundancy and physical continuity between adjacent video frames. Neural network models assume that the semantic contribution of each frame is equal during computation, leading to computational resource consumption on repetitive feature extraction of known backgrounds or stable targets. Furthermore, frequent activation of deep convolutional branches generates latency jitter when processing high frame rate image sequences. While simply reducing the sampling frequency or using model quantization may decrease the total computation, it often neglects the logical consistency of the temporal dimension, resulting in flickering prediction boxes or decreased sensitivity to abrupt changes. Existing architectures lack underlying response mechanisms for temporal correlations and cannot dynamically adjust the computational path and logical flow of deep learning models based on the feature evolution of the input data, causing ineffective use of computing resources by the underlying computing architecture of the computer system. It is difficult to achieve efficient collaboration between the computational model and the hardware execution unit. For example, Chinese invention patent CN113065422B discloses a training method and device for video target detection, which introduces a matching network to support image and video pipelines and handles cold start and feature consistency in small sample environments. However, in complex production sites, this technology still belongs to the feature comparison method. The core logic is anchored on multi-frame feature fusion and matching, without touching the incremental transformation of the convolution operation itself. When dealing with high redundant data streams, the underlying primitive computation scheduling burden is heavy. Such solutions lack a phase recognition mechanism for periodic mechanical motion. When faced with temporal background interference caused by the rotation of rolls or regular vibration of strip, it is easy to induce logical mis-triggering due to static threshold or matching drift. It is difficult to achieve closed-loop adaptive optimization of computing power while ensuring sensitivity.

[0004] Therefore, how to construct a dedicated computer system architecture with dynamic routing and scheduling capabilities and incremental evolution of feature tensors, so that deep learning models can dynamically select computation paths based on the evolution characteristics of hidden states, is the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A video target detection method based on a deep neural network architecture, comprising the following steps: Step S101: Obtain the input tensor at the current sampling time, and call the reference feature tensor stored in the hidden feature resident unit, which was generated by the previous sampling time, by the feature extraction branch of the deep neural network; Step S102: Calculate the residual tensor between the input tensor and the reference feature tensor, and extract the perceptual energy bias that characterizes the distribution density of activation elements within the residual tensor. Step S103: Obtain periodic operating condition data that characterizes the operating cycle of the external operating condition source; extract the dominant frequency of the sensing energy deviation in the time domain by using an autocorrelation phase observer embedded in the latent feature resident unit; anchor the dominant frequency to the periodic operating condition data in phase; and calculate the sensing energy gating threshold that fluctuates nonlinearly with the phase based on the phase change law of the periodic operating condition data. Step S104: Compare the perceived energy deviation with the perceived energy gating threshold, and control the calculation route to switch between the tensor incremental evolution path and the logical silent path based on the comparison result: If the perceived energy deviation exceeds the perceived energy gating threshold, the detection head branch of the deep neural network is activated to execute the detection logic; if the perceived energy deviation does not exceed the perceived energy gating threshold, the feature dwelling state of the hidden feature dwelling unit is maintained.

[0006] Preferably, step S103 includes: inputting the sensing energy deviation in the continuous sampling sequence into a fast Fourier transform operator to extract the dominant frequency component in the time domain signal; identifying the cyclic interference frequency in the dominant frequency component that is associated with the external operating environment; and adjusting the amplitude of the dynamic envelope within the resonant phase interval between the dominant frequency component and the cyclic interference frequency so that the sensing energy gating threshold is incrementally compensated in real time to follow the physical fluctuation cycle of the external operating environment.

[0007] Preferably, step S102 includes: performing spatial dimensionality reduction on the residual tensor to obtain a feature residual map; statistically analyzing the pixel proportion and average activation intensity of non-zero elements in the feature residual map; and weighting the pixel proportion and average activation intensity to generate a perceptual energy bias.

[0008] Preferably, after step S103, the method further includes: analyzing the spatial connectivity and gradient consistency of the residual tensor; if it is determined that the non-zero elements of the residual tensor exhibit a globally diffuse distribution and the gradient direction lacks directionality, it is identified as an external sensing channel failure and a path breaker command is generated; in response to the path breaker command, the deep neural network closes the activation paths of the feature extraction branch and the detection head branch of the deep neural network, so that the deep neural network enters a low-power sleep mode.

[0009] Preferably, the tensor increment evolution path in step S104 includes: using the residual tensor to perform element-wise superposition of the reference feature tensor to generate an updated feature tensor; feeding the updated feature tensor back to the hidden feature resident unit to replace the reference feature tensor as a comparison benchmark for subsequent sampling times.

[0010] Preferably, it also includes an active perturbation detection process: injecting standardized perturbation tensors into the intermediate feature layer of the deep neural network at preset time intervals; monitoring the response echoes generated by the hidden feature residing units; if the intensity of the response echoes decays to below a preset sensitivity threshold, then applying gain compensation to the feature weights in the hidden feature residing units based on the amplitude deviation of the response echoes.

[0011] Preferably, the feature extraction branch of the deep neural network adopts a residual network structure, and the hidden feature resident unit is set in the bottleneck layer of the residual network structure to cut off the full inference pipeline of the deep neural network according to the switching state of the computation route.

[0012] Preferably, the quantitative determination logic for the sensing energy gating threshold and the sensing energy deviation follows the following discriminant: ,in, This is a route switching switch indicator. To detect energy deviation, The preset phase sensitivity factor, The phase determined by the dominant frequency at different sampling times The change in compensation amplitude, The preset base energy threshold, γ as well as All calculations are performed using dimensionless numerical values.

[0013] Preferably, the calculation cycle of the perceived energy deviation is synchronized with the inference cycle of the deep neural network; when a feature mutation signal or sensor reset signal is detected in the input data stream, the hidden feature resident unit is cleared.

[0014] Preferably, the external operating condition source is an industrial processing device with periodic rotation or reciprocating motion characteristics, and the periodic operating condition data is provided by the operating frequency of the industrial processing device; the autocorrelation phase observer identifies the mechanical beat of the industrial processing device and shields the time-domain background interference components generated by the rotation or reciprocating motion.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In video object detection, by establishing latent feature resident units in the underlying architecture of the neural network computing model, the full feedforward inference is transformed into a tensor incremental evolution mode. Under the condition that there is a high temporal correlation in the video data stream, the system only needs to process the incremental information generated by the perception energy deviation, reduce the activation frequency of deep convolution branches, and reduce the logical load in the operation of the computation graph. This realizes the transformation of the underlying computing mode of the deep neural network model in the computer system from data-driven to event-increment-driven, and avoids the waste of underlying computing resources caused by the deep neural network architecture performing equal-weighted repeated calculations on video frame sequences at the system architecture level.

[0016] 2. By introducing a time-axis phase-locking mechanism in the energy deviation judgment stage, the dominant phase in the time domain of the time-series tensor sequence is extracted, causing the judgment threshold to fluctuate nonlinearly with the physical motion beat. In application scenarios with mechanical cycle characteristics such as metal rolling or high-speed stamping, the system can identify and shield periodic interference caused by the rotation of rolls and the regular vibration of strip. This enables deep decoupling and logical reconstruction of the physical environment beat at the computing model architecture level, avoiding the mis-triggered routing of the underlying computing logic caused by regular background noise, and improving the computing power scheduling accuracy and inference stability of the computer system under complex industrial conditions.

[0017] 3. By utilizing the active detection mechanism of injecting standardized micro-perturbation tensors into the intermediate layer of the neural network, the system sensitivity state is inverted based on the response echo generated by the hidden feature resident unit. Under the condition that the monitoring screen is in a static or extremely low dynamic state for a long time, nonlinear gain compensation is applied to the feature weights to solve the feature passivation problem caused by the lack of effective signal stimulation in the incremental computing architecture for a long time, ensuring that the system maintains a high sensitivity to capture sudden moving targets. Attached Figure Description

[0018] Figure 1 This is a flowchart of the video target detection method based on tensor incremental evolution logic of the present invention; Figure 2 This is a schematic diagram of the overall system architecture for the integrated timing phase locking mechanism of this invention. Detailed Implementation

[0019] The present invention will be further described in detail below through specific embodiments. It should be understood that the following embodiments are intended to explain the present invention and not to limit the scope of protection of the present invention.

[0020] This invention provides a video target detection method based on a deep neural network architecture. It reconstructs traditional frame-by-frame inference into a temporally correlated tensor incremental evolution approach. By establishing latent feature residing units in the underlying architecture of the deep neural network, it achieves the storage and reuse of feature distributions from previous sampling moments. A perceptual energy gating mechanism based on temporal axis phase locking is introduced. An autocorrelation phase observer extracts the dominant frequency of the working environment and anchors the phase, thereby controlling the computational route to dynamically switch between the tensor incremental evolution path and the logically silent path. Finally, combined with active perturbation detection and computational validity arbitration procedures, it achieves closed-loop adaptive optimization of computational resource allocation and system sensitivity compensation. In precision manufacturing monitoring scenarios such as continuous metal rolling or high-speed stamping, video streams have high frame rates and semantic redundancy between adjacent frames. To address this challenge, this invention transforms the video stream into a continuous temporal tensor sequence and calls the reference feature tensors generated from previous sampling moments stored in the latent feature residing units in the feature extraction branch of the deep neural network. The hidden feature residing unit is located in the bottleneck layer of the residual network structure and is used to store the high-dimensional feature tensor of the previous reference frame as the current input tensor. By establishing a temporal correlation at the initial stage of computation, the system provides a reference field for subsequent incremental processing, reducing the frequency of redundant convolution operators being repeatedly executed in the underlying computation graph. The system utilizes the processor to calculate the residual tensor between the input tensor and the reference feature tensor, and extracts the perceptual energy bias from it. The processor performs spatial dimensionality reduction on the residual tensor to obtain the feature residual map, and then calculates the pixel percentage of non-zero elements in the feature residual map. With average activation intensity The two are then weighted to generate a perceived energy bias. The calculation formula is as follows: ,in, and The energy deviation is perceived using preset weighting coefficients. The distribution density of activation elements within the residual tensor is used to characterize the evolution of the current video frame relative to the dwelling features, thereby quantifying image differences into a logical indicator for controlling route switching.

[0021] For industrial processes such as metal rolling that involve periodic rotation or reciprocating motion, the system acquires periodic operating condition data that characterizes the operating cycle of external operating conditions. By using an autocorrelation phase observer embedded in the latent feature-resident unit, the sensing energy bias is extracted. In the time domain, the system inputs the sensed energy deviation within the continuously sampled sequence into the Fast Fourier Transform operator to extract the dominant frequency component from the time-domain signal. It identifies cyclic interference frequencies associated with the external operating environment and then controls the dominant frequency. With periodic operating condition data Phase anchoring is performed, and based on the phase variation pattern of periodic operating data, the sensing energy gating threshold, which fluctuates nonlinearly with the phase, is calculated. The judgment logic is as follows: ,in, This is a route switching switch identifier; To sense energy deviation; The preset phase sensitivity factor; The phase determined by the dominant frequency at different sampling times The magnitude of the compensation for the change; The preset base energy threshold is used; the energy gating threshold is sensed by adjusting the amplitude of the dynamic envelope within the phase interval. It can incrementally compensate for physical fluctuations in the external environment in real time, thereby shielding background interference components generated by roll rotation or mechanical vibration; and determine the sensing energy gating threshold. At the same time, the autocorrelation phase observer is used to collect the sensed energy deviation during the complete mechanical operation cycle. A baseline fluctuation set is established based on the sequence, and the dominant frequency components in the baseline fluctuation set are extracted using the Fast Fourier Transform operator. and initial phase, periodic operating condition data Real-time rotation angle normalized to to The radian interval determines the phase of the sampling time. Based on dominant frequency components Constructing a sinusoidal compensation function The amplitude is taken from the mean envelope of the reference fluctuation set at the corresponding phase, according to the formula. Generate sensing energy gating threshold This enables the gated boundary to follow the physical vibration trajectory of external operating conditions in real time.

[0022] The processor determines the energy deviation based on the sensor. With sensing energy gating threshold The comparison results of the magnitude values ​​control the computational route to switch between the tensor incremental evolution path and the logical silent path, if an energy deviation is detected. Exceeding the sensing energy gating threshold The system activates the detection head branch of the deep neural network to execute the detection logic and enters the tensor increment evolution path. In this path, the residual tensor is used to perform element-wise superposition on the reference feature tensor to generate an updated feature tensor, which is then fed back to the hidden feature resident unit for replacement, serving as the benchmark for subsequent time steps. If a perceptual energy deviation is detected... Not exceeding the sensing energy gating threshold The system maintains the latent feature residency state and reuses the residency features, while the deep convolutional branch remains silent. This procedure reconstructs full inference into an incremental evolution mode triggered on demand, reducing the instantaneous peak dependence of computing resources. To address invalid computations caused by lens occlusion or sensor physical failure, the system executes a computation validity arbitration procedure. The processor analyzes the spatial connectivity and gradient consistency indices of the residual tensor, uses a binarization operator to perform threshold segmentation on the feature residual map to extract the connected components of the active pixels, and calculates the gradient response vectors of the active pixels in the horizontal and vertical directions. and Furthermore, the gradient consistency index is derived through the cosine similarity operator. The specific calculation formula is as follows: ,in, As a gradient consistency index, For horizontal gradient components, The vertical gradient component; the percentage of the total area of ​​the activated pixel connected components. A value greater than 0.75 and gradient consistency index When the value is below 0.15, the system determines that the current feature increment has lost its semantic attributes due to interference from the globally diffuse water mist occlusion, determines that the external sensing channel has failed, and generates a path breaker command. The 0.75 area proportion threshold and 0.15 gradient consistency threshold involved in the above determination are determined by the occlusion calibration program in the system initialization phase: after the device is installed, 20 frames of reference residual maps are obtained by artificially simulating the lens being completely occluded. The average connected component distribution and directional gradient entropy in this state are statistically analyzed. The minimum area proportion in the occlusion state is reduced by 10% as the upper limit of determination, and the maximum gradient consistency is increased by 15% as the lower limit of determination. This establishes a dedicated breaker criterion for the current physical installation environment. The deep neural network responds to this command by forcibly closing the feature extraction branch and the detection head branch, causing the system to enter a low-power sleep mode. The system maintains the lowest frequency residual monitoring logic until the residual distribution returns to the preset natural feature range, thereby achieving deep coupling between the computational logic and the physical state of the hardware.

[0023] To address the potential feature passivation issue that may occur in incremental computing architectures under long-term silent operation, the system executes an active perturbation detection process. The system injects standardized perturbation tensors into the intermediate feature layers of the deep neural network at preset time intervals. And monitor the response echoes generated by the hidden characteristic resident units. If the response echo When the intensity decreases below a preset sensitivity threshold, the response echo... The amplitude deviation is compensated by applying nonlinear gain to the feature weights in the resident unit to maintain the neural network's sensitivity to small incremental signals; during the active perturbation detection process, a pseudo-random number generator is used to generate Gaussian distributed data with a mean of 0 and a variance of 0.01 to fill a tensor space of size 16×16×256 to generate a standardized micro-stage perturbation tensor. The processor injects the micro-stage perturbation tensor into the output of the fourth residual block of the deep neural network and captures the echo responses of the hidden feature resident units. By calculating the response echo The ratio of the real-time response intensity to the L2 norm of the injected tensor is obtained. When this ratio is lower than the preset sensitivity threshold At that time, the gain compensation coefficient is used. Element-wise multiplication is performed on the feature weight matrix of the latent feature residing unit. The compensation coefficient is taken as the ratio of the sensitivity threshold to the real-time intensity, with the upper limit set at 2.25 to maintain the sensitivity of the feature weights to small incremental signals. In addition, the system calculates the feature entropy based on the confidence vector output by the target detection head. And combined with the inter-frame offset of the position coordinates Constructing a semantic uncertainty function When the semantic uncertainty function When the value exceeds the preset boundary, the system sends a reset command to the neural network and updates the feature tensor in the hidden feature resident unit through a full convolution operation. This feedback adjustment link realizes closed-loop adaptive optimization of computing resource allocation and ensures the consistency of the system output during long-term continuous operation.

[0024] Example 1: In the video monitoring environment of a continuous rolling mill production line, the camera equipment captures video of the strip running at a frequency of 60 frames per second. The large rolls in the production area generate recurring mechanical vibrations and periodic cooling water mist spray reflections during rotation. These physical disturbances have fixed frequency characteristics on the time axis. Due to the periodic abrupt changes in background pixels, the traditional frame-by-frame inference method triggers the scheduling of deep convolution branches, causing the processor load at the bottom layer of the computer system to remain above 90%. The system executes a tensor incremental evolution process, where the hidden feature resident unit stores the reference feature tensor generated from the background image. When the current input tensor contains periodic reflection interference Upon entering the processing link, the processor calculates its sensed energy deviation. The autocorrelation phase observer performs a fast Fourier transform on the historical sequence of perceived energy deviations to extract the dominant frequency component whose frequency value corresponds to the roll speed frequency. Acquire synchronously collected periodic operating condition data Phase information Based on this, the sensing energy gating threshold, which fluctuates nonlinearly with phase, is calculated. Specifically, the energy-sensing threshold The calculation formula is ,in, To sense the energy gating threshold. The preset base energy threshold, The preset phase sensitivity factor, The phase determined by the dominant frequency at different sampling times The changing compensation amplitude.

[0025] When the phase corresponding to the reflective point When at the peak of amplitude, the sensing energy gating threshold The corresponding upward movement reduces the perceived energy deviation affected by physical cycle noise. Below the gate envelope, the computation route remains within a logically silent path and reuses the feature distribution in the latent feature-residing units until non-periodic crack defects appear on the strip surface, reducing the proportion of non-zero pixels in the residual tensor. The resulting increment penetrates the sensing energy gating threshold. The system triggers the detection head branch to perform feature updates and target recognition, utilizing periodic operating condition data. Phase alignment is performed on the sensing energy gate, enabling the neural network to identify and ignore predictable background evolution processes, and allocate computing resources to sudden incremental events that deviate from physical laws. Under the condition of high-frequency fluctuations in the rolling mill, the instruction scheduling latency is reduced from 50ms to less than 5ms.

[0026] Example 2: In the performance verification test of the simulated industrial video surveillance platform, the system used a strip production line monitoring sequence with a frame rate of 60fps and a resolution of 1920x1080 as the raw input data. Gaussian white noise with a signal-to-noise ratio of 20.0dB and periodic flickering disturbances with a frequency of 15.0Hz were actively superimposed at the input to simulate electromagnetic interference and physical cyclic reflection noise in a precision manufacturing environment. In the parameter setting procedure, the sensing energy deviation weighting coefficient... The weighting coefficient is set to 0.60. A value of 0.40 is chosen to balance the contributions of spatial activation density and average activation intensity in the residual tensor to semantic evolution determination; this is the basic energy threshold. The value of is determined based on the engineering trade-off between detection sensitivity and false trigger rate. When the value is 0.12, the system can effectively shield the background thermal noise while ensuring the capture rate of small defects with a diameter of not less than 2 mm. See Table 1, which records the performance comparison results of the sample group of the present invention under the above parameter configuration with the control group using the frame-by-frame full inference method, the partially missing control group with missing phase locking mechanism, and the out-of-range control group with the threshold setting exceeding the optimal window.

[0027] Table 1: Comparison Test Data of Video Target Detection Performance

[0028] Analysis of the experimental data shown in Table 1 reveals that when the environmental interference frequency is 15.0 Hz and the interference phase reaches the peak of the reflection wave, the control group experiences a processor utilization rate of 92.4% due to the full convolution operation, and a single-frame inference latency of 48.65 ms. In contrast, the sample group of this invention utilizes an autocorrelation phase observer to calculate the sensing energy gating threshold that varies with the phase. This causes a deviation in perceived energy, which includes periodic reflective components. Below the dynamic threshold envelope, the computational route remains on the logically silent path and reuses the feature distribution in the latent feature residing unit. The single-frame processing latency is reduced to 3.12ms and the processor utilization rate is reduced to 11.6%. In addition, as the amplitude of the simulated interference source increases, the sample of this invention maintains the fluctuation range of the single-frame inference latency within 0.5ms through the phase-locking mechanism, while the out-of-range control group has a detection accuracy rate of 65.2% due to the excessively high threshold setting. In the metal hot rolling monitoring condition including high-pressure water vapor spray, the instantaneous overflow of steam on site causes global fogging and occlusion of the camera lens. The non-zero pixels in the residual tensor exhibit disordered diffuse distribution characteristics. The processor detects the total area ratio of the active pixel connected regions. The value suddenly increased from the normal 0.05 to 0.82, and the resulting gradient consistency index was calculated. The value dropped from 0.88 to 0.09. After the aforementioned dimensional index penetrated the preset arbitration boundary value, the processor determined that the sensing channel had failed and instantly executed a path breaker operation, causing the computation graph scheduling of the deep learning model to enter a logical silent state. In this state, the system's single-frame power consumption was reduced from 15.0W to 2.5W, and it stopped outputting false defect detection commands until the water vapor dissipated. Restored to the preset safe range above 0.60.

[0029] Example 3: This example combines Figures 1 to 2 The following describes a video object detection method based on a deep neural network architecture, such as... Figure 1As shown, step S101 obtains the input tensor at the current sampling time, and the feature extraction branch of the deep neural network calls the reference feature tensor stored in the latent feature residence unit, which was generated by the previous sampling time. Step S102 calculates the residual tensor between the input tensor and the reference feature tensor, and extracts the sensing energy deviation that represents the distribution density of activation elements in the residual tensor. Then, step S103 obtains periodic operating data, extracts the dominant frequency of the sensing energy deviation through the autocorrelation phase observer, anchors the dominant frequency with the operating data in phase, and calculates the sensing energy gating threshold that fluctuates nonlinearly with the phase according to the phase change law. Finally, in step S104, the sensing energy deviation is compared with the sensing energy gating threshold, and the calculation route is controlled to switch between tensor incremental evolution and logical silent path. Specifically, if the threshold is exceeded, the detection head branch is activated; if it is not exceeded, the latent feature residence state is maintained.

[0030] like Figure 2 As shown, the overall system architecture includes a physical sensing domain, a temporal phase-locked controller, a tensor incremental evolution computation core, a latent feature residence unit, and a decision execution domain. Within the physical sensing domain, the operating condition signal acquisition unit is responsible for acquiring the machine's operating cycle time and transmitting the phase data to the temporal phase-locked controller. The high-speed camera module is responsible for acquiring video stream tensors and transmitting the input tensors to the tensor incremental evolution computation core. The temporal phase-locked controller generates a follow-up gating threshold based on the phase data and inputs it into the computation core. The tensor incremental evolution computation core, deployed on the edge computing host, integrates a residual computation engine and dynamic computation routing. This core interacts bidirectionally with the latent feature residence unit, which stores the reference feature benchmark, for feature residence or invocation. Furthermore, it outputs signals to the decision execution domain based on the processing results, enabling the monitoring and visualization terminal in the decision execution domain to receive the target detection results, while the safety circuit breaker actuator receives the path circuit breaker command.

[0031] Example 4: In extremely low-dynamic monitoring environments such as warehouses at night, the latent feature residing units are at risk of sensitivity passivation due to the quasi-static feature distribution. The system executes an active perturbation detection process to maintain the activity of stored features. The processor uses a pseudo-random sequence generator to generate a Gaussian noise field with a mean of 0 and a variance of 0.01, and maps the noise field to a tensor space of size 16×16×256 to construct a normalized perturbation tensor. The normalized perturbation tensor is combined using the tensor addition operator. Injected into the output of the fourth residual block in the deep neural network, the response echo of the hidden feature resident unit to the spatial topological perturbation is monitored. And calculate the response echo. L2 norm and standardized perturbation tensor The ratio of the L2 norms is used to obtain the response intensity ratio. Response intensity ratio The calculation formula is as follows: ,in, The ratio of response intensity, For L2 norm operators, For the observed response echo tensor, This refers to the injected standardized micro-perturbation tensor.

[0032] The processor selects a moving target video sequence containing a 1.0% brightness increment under controlled conditions and executes the detection process, records the response intensity ratio sequence, and selects the lower quartile of the response intensity ratio sequence as the sensitivity threshold. In this embodiment, the sensitivity threshold The calibration value is 0.85, and it is used for real-time detection. When the value is below 0.85, the resident cell is determined to be in a passivated state, using the formula... Calculate the gain compensation coefficient And use element-wise multiplication to adjust the gain compensation coefficient The feature weight matrix applied to the latent feature residing units enhances sensitivity; where, This is the gain compensation coefficient. The preset sensitivity threshold, This represents the real-time response intensity ratio; the processor acquires the confidence vector probability distribution output by the target detection head. And calculate the feature entropy Simultaneously calculate the position offset of the current detection box center point relative to the reference time. Constructing a semantic uncertainty function using a linear weighted model To control numerical drift; semantic uncertainty function The calculation formula is as follows: ,in, This represents the semantic uncertainty value. The preset first weighting coefficient, The second weighting coefficient is preset. For feature entropy, This is the pixel offset of the center point. The length of the video frame diagonal in pixels; when the semantic uncertainty function When the value exceeds three times the standard deviation of the mean calculated based on 300 historical sampling periods, the processor sends a reset instruction to the deep neural network, triggering a full inference path to update the hidden feature resident units and eliminate the cumulative effect of numerical truncation error in the tensor increment evolution process.

[0033] Example 5: In the initialization state of the newly deployed video target detection system, the processor executes a benchmark calibration procedure to determine the boundary values ​​of the sensing energy gating threshold, and acquires data against a static background excluding the target to be detected. The video sequence is processed frame by frame, and the perceptual energy deviation sequence between each frame is calculated. The mean and standard deviation of the perceptual energy deviation sequence are calculated, and then the result is obtained according to the formula. Determine the basic energy threshold ,in, Based on the basic energy threshold, To perceive the mean of the energy deviation sequence, To perceive the standard deviation of the energy bias sequence, The preset confidence level coefficient is used. In this embodiment, the total number of frames in the video sequence is... The value is set to 3.0, resulting in the basic energy threshold. Establish a discrimination criterion that is correlated with sensor sensitivity and background noise.

[0034] When the system faces the task of aligning the phase of an external rotating operating condition source with the sampling time of the video stream, the processor obtains the initial phase deviation through an autocorrelation phase observer and receives periodic operating condition data characterizing the physical rotation. and periodic operating condition data Normalize to 0 to Within the arc range, the energy deviation is monitored simultaneously within the time window. The peak distribution pattern can be determined by calculating the cross-correlation function between the two. Identify the time delay corresponding to the maximum correlation coefficient and the delay amount Converted into initial phase compensation values ​​in a deep neural network inference link ,in, It is a cross-correlation function. For delay quantity, As the initial value for phase compensation, this process makes the phase sensitivity factor... Compensation amplitude under action Synchronized with the reflective pulses generated by the external rotating operating condition source in the time axis direction, the route switching switch is indicated. The path switching decision is made based on the synchronized sensing energy gating threshold.

[0035] Example 6: In the deployment process of the strip defect detection system, the processor executes the offline calibration procedure of the hidden feature resident unit. The system selects a monitoring video sequence containing 5000 sampling periods as the calibration dataset. The processor uses the feature extraction branch of the deep neural network to calculate the reference feature tensor of the calibration dataset. The distribution of the processor statistical reference feature tensor The activation density in the convolution channel direction is adjusted, and the cluster center in the feature space is locked using the mean shift algorithm. In this process, the system completes the benchmark construction of the background semantic field.

[0036] When the system is in debugging mode to adapt to differences in the sensitivity of camera equipment, the processor performs stability tests to determine the gain compensation coefficient. The system injects a normalized micro-perturbation tensor of gradient changes into the deep neural network at the operating boundary. And collect the response intensity ratio The change curve, processor monitors response echo The numerical stability, when the gain compensation coefficient When the value exceeds 2.50, the feature weights within the hidden feature residing cells exhibit convergent oscillations, and the system adjusts the gain compensation coefficient accordingly. The saturation extremum is set to 2.25, and the processor uses a linear regression method to determine the semantic uncertainty function. First weighting coefficient With the second weighting coefficient and will Set to 0.70 and Set to 0.30 to adjust the threshold for determining system reset commands in environments with fluctuating background complexity.

[0037] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0038] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A video target detection method based on a deep neural network architecture, characterized in that, Includes the following steps: Step S101: Obtain the input tensor at the current sampling time, and call the reference feature tensor stored in the hidden feature resident unit, which was generated by the previous sampling time, by the feature extraction branch of the deep neural network; Step S102: Calculate the residual tensor between the input tensor and the reference feature tensor, and extract the perceptual energy bias that characterizes the distribution density of activation elements within the residual tensor. Step S103: Obtain periodic operating condition data that characterizes the operating cycle of the external operating condition source; extract the dominant frequency of the sensing energy deviation in the time domain by using an autocorrelation phase observer embedded in the latent feature resident unit; anchor the dominant frequency to the periodic operating condition data in phase; and calculate the sensing energy gating threshold that fluctuates nonlinearly with the phase based on the phase change law of the periodic operating condition data. Step S104: Compare the perceived energy deviation with the perceived energy gating threshold, and control the calculation route to switch between the tensor incremental evolution path and the logical silent path based on the comparison result: If the perceived energy deviation exceeds the perceived energy gating threshold, the detection head branch of the deep neural network is activated to execute the detection logic; if the perceived energy deviation does not exceed the perceived energy gating threshold, the feature dwelling state of the hidden feature dwelling unit is maintained.

2. The video target detection method based on a deep neural network architecture according to claim 1, characterized in that, Step S103 includes: inputting the sensing energy deviation in the continuous sampling sequence into the fast Fourier transform operator to extract the dominant frequency component in the time domain signal; identifying the cyclic interference frequency in the dominant frequency component that is associated with the external operating environment; and adjusting the amplitude of the dynamic envelope within the resonant phase interval between the dominant frequency component and the cyclic interference frequency so that the sensing energy gating threshold is incrementally compensated in real time according to the physical fluctuation cycle of the external operating environment.

3. The video target detection method based on a deep neural network architecture according to claim 1, characterized in that, Step S102 includes: performing spatial dimensionality reduction on the residual tensor to obtain a feature residual map; statistically analyzing the pixel proportion and average activation intensity of non-zero elements in the feature residual map; and weighting the pixel proportion and average activation intensity to generate a perceptual energy bias.

4. The video target detection method based on a deep neural network architecture according to claim 1, characterized in that, After step S103, the method further includes: analyzing the spatial connectivity and gradient consistency of the residual tensor; if it is determined that the non-zero elements of the residual tensor exhibit a globally diffuse distribution and the gradient direction lacks directionality, it is identified as an external sensing channel failure and a path breaker command is generated; in response to the path breaker command, the deep neural network closes the activation paths of the feature extraction branch and the detection head branch of the deep neural network, causing the deep neural network to enter a low-power sleep mode.

5. The video target detection method based on a deep neural network architecture according to claim 1, characterized in that, The tensor increment evolution path in step S104 includes: using the residual tensor to perform element-wise superposition of the reference feature tensor to generate the updated feature tensor; feeding the updated feature tensor back to the hidden feature resident unit to replace the reference feature tensor as a comparison benchmark for subsequent sampling times.

6. The video target detection method based on a deep neural network architecture according to claim 1, characterized in that, It also includes an active perturbation detection process: injecting standardized perturbation tensors into the intermediate feature layer of the deep neural network at preset time intervals; monitoring the response echoes generated by the hidden feature residing units; if the intensity of the response echoes decays to below a preset sensitivity threshold, then applying gain compensation to the feature weights in the hidden feature residing units based on the amplitude deviation of the response echoes.

7. The video target detection method based on a deep neural network architecture according to claim 1, characterized in that, The feature extraction branch of the deep neural network adopts a residual network structure, and the hidden feature resident unit is set in the bottleneck layer of the residual network structure to truncate the full inference pipeline of the deep neural network according to the switching state of the computation route.

8. The video target detection method based on a deep neural network architecture according to claim 1, characterized in that, The quantitative determination logic for the sensing energy gating threshold and sensing energy deviation follows the following discriminant: ,in, This is a route switching switch indicator. To detect energy deviation, The preset phase sensitivity factor, The phase determined by the dominant frequency as it varies with the sampling time. The changing compensation amplitude, The preset base energy threshold, , , as well as All calculations are performed using dimensionless numerical values.

9. The video target detection method based on a deep neural network architecture according to claim 1, characterized in that, The calculation cycle of the perceived energy deviation is synchronized with the inference cycle of the deep neural network; when a feature mutation signal or sensor reset signal is detected in the input data stream, the hidden feature resident units are cleared.

10. The video target detection method based on a deep neural network architecture according to claim 1, characterized in that, The external operating condition source is an industrial processing device with periodic rotation or reciprocating motion characteristics. The periodic operating condition data is provided by the operating frequency of the industrial processing device. The autocorrelation phase observer identifies the mechanical beat of the industrial processing device and shields the time-domain background interference components generated by the rotation or reciprocating motion.

Images