Dynamic detection method and system of advertisement based on image recognition

Abstract

This invention discloses a dynamic detection method and system for advertisements based on image recognition. The invention relates to the field of image recognition technology, determining the feature change gradient matrix based on image recognition of advertisement feature maps; performing temporal tracking on the feature change gradient matrix to extract gradual trajectory features representing the transition of illegal content from concealment to manifestation, and stroboscopic trajectory features representing the periodic flashing of illegal elements, thus improving the accuracy of the gradual trajectory features and stroboscopic trajectory features. The advertisement feature vector is mapped to a graph query node and input into a violation knowledge graph to determine the topological distance between the graph query node and each illegal semantic entity node in the violation knowledge graph; the placement scenario of the advertisement to be detected is obtained, and the corresponding violation frame sequence is determined by combining the corresponding traffic warning content and the topological distance; based on the violation frame sequence, multiple corresponding sub-violation content are determined, improving the accuracy of the dynamic detection report.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more particularly to a method and system for dynamic detection of advertisements based on image recognition. Background Technology

[0002] With the rapid development of digital media technology and video editing algorithms, the content of internet advertising is becoming increasingly complex and dynamic. Especially in video advertising in the fields of beauty and skincare, in order to pursue exaggerated product efficacy displays, advertisers often use highly concealed digital special effects to illegally tamper with video footage. This kind of tampering is no longer a simple single-frame static patch or a crude scene jump cut, but has evolved into a dynamic cheating method that is extremely difficult to detect in the temporal dimension. For example, facial contours are slowly altered by finely adjusting the liquefaction effect frame by frame, or skin texture is periodically faked by using a high-frequency flashing mask effect.

[0003] Existing technologies typically extract frames from videos and treat them as independent static images for spatial texture and color distribution detection. This isolated single-frame detection mode disrupts the temporal continuity of the video and fails to capture the subtle pixel displacement fields and motion edge changes caused by hidden effects between adjacent frames. At the same time, existing technologies often ignore or simply splice audio spectral features. Due to the lack of advertising feature maps that construct multimodal spatiotemporal joint representations, existing technologies struggle to accurately track feature change gradients in a temporal sequence, affecting the accuracy of gradient trajectory features and flicker trajectory features, resulting in low accuracy of dynamic detection reports. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a dynamic detection method and system for advertisements based on image recognition.

[0005] This invention provides a method for dynamic detection of advertisements based on image recognition, comprising: The system collects advertisements to be detected, performs multi-scale segmentation on the advertisements to be detected, extracts visual appearance features and optical flow motion features respectively, extracts audio spectrum features from the advertisements to be detected simultaneously, and generates corresponding advertisement feature maps by combining visual appearance features and optical flow motion features. The feature change gradient matrix is ​​determined by image recognition based on the advertising feature map; the feature change gradient matrix is ​​tracked in time to extract the gradual trajectory features representing the illegal content from concealment to manifestation and the stroboscopic trajectory features representing the periodic flashing of illegal elements. The gradual trajectory features and the stroboscopic trajectory features are fused to generate the corresponding advertising feature vectors; the advertising feature vectors are mapped to graph query nodes and input into the violation knowledge graph to determine the topological distance between the graph query nodes and each violation semantic entity node in the violation knowledge graph; The system obtains the placement scenario of the ad to be detected, and determines the corresponding violation frame sequence by combining the corresponding traffic warning content and the topological distance. Based on the violation frame sequence, it determines multiple sub-violation contents and determines the corresponding dynamic detection report by combining the ad constraint factors of the ad to be detected.

[0006] This invention provides a dynamic advertising detection system based on image recognition, which is applied to the aforementioned dynamic advertising detection method based on image recognition; the dynamic advertising detection system based on image recognition includes: The advertising feature map module is used to collect advertisements to be detected, perform multi-scale segmentation on the advertisements to be detected, extract visual appearance features and optical flow motion features respectively, extract audio spectrum features in the advertisements to be detected simultaneously, and generate corresponding advertising feature maps by combining visual appearance features and optical flow motion features. The image recognition module is used to determine the feature change gradient matrix based on the image recognition of the advertising feature map; and to perform time-series tracking of the feature change gradient matrix to extract the gradual trajectory features representing the illegal content from concealment to manifestation and the strobe trajectory features representing the periodic flashing of illegal elements. The knowledge graph module is used to fuse the features of gradual trajectory and strobe trajectory to generate the corresponding advertising feature vector; the advertising feature vector is mapped to a graph query node and input into the violation knowledge graph to determine the topological distance between the graph query node and each violation semantic entity node in the violation knowledge graph; The dynamic detection module is used to obtain the placement scenario of the advertisement to be detected, and determine the corresponding violation frame sequence by combining the corresponding traffic warning content and the topological distance. Based on the violation frame sequence, it determines multiple sub-violation content, and determines the corresponding dynamic detection report by combining the advertising constraint factors of the advertisement to be detected.

[0007] Compared with the prior art, the beneficial effects of the present invention are: (1) Collect the advertisement to be detected, perform multi-scale segmentation on the advertisement to be detected, extract visual appearance features and optical flow motion features respectively, extract audio spectrum features in the advertisement to be detected simultaneously, and generate corresponding advertisement feature maps by combining visual appearance features and optical flow motion features; determine the feature change gradient matrix based on image recognition of the advertisement feature map; perform time-series tracking on the feature change gradient matrix to extract the gradual trajectory features representing the illegal content from concealment to manifestation and the stroboscopic trajectory features representing the periodic flashing of illegal elements. The advertisement feature map is introduced to realize the image recognition of the advertisement feature map and improve the accuracy of the gradual trajectory features and stroboscopic trajectory features.

[0008] (2) Feature fusion is performed on the gradual trajectory feature and the flashing trajectory feature to generate the corresponding advertising feature vector; the advertising feature vector is mapped to a graph query node and input into the violation knowledge graph to determine the topological distance between the graph query node and each violation semantic entity node in the violation knowledge graph; the placement scenario of the advertisement to be detected is obtained, and the corresponding violation frame sequence is determined by combining the corresponding traffic warning content and the topological distance; multiple sub-violation contents are determined according to the violation frame sequence, and the corresponding dynamic detection report is determined by combining the advertising constraint factors of the advertisement to be detected. The violation frame sequence is further controlled, and multiple sub-violation contents and advertising constraint factors of the advertisement to be detected are fully considered, which improves the accuracy of the dynamic detection report. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the dynamic detection method for advertisements based on image recognition in an embodiment of the present invention. Figure 2 This is a flowchart illustrating step S11 of the image recognition-based dynamic detection method for advertisements in this embodiment of the invention. Figure 3 This is a flowchart illustrating step S12 in the image recognition-based dynamic detection method for advertisements in this embodiment of the invention. Figure 4 This is a flowchart illustrating step S13 in the image recognition-based dynamic detection method for advertisements in this embodiment of the invention. Figure 5 This is a flowchart illustrating step S14 of the image recognition-based dynamic detection method for advertisements in this embodiment of the invention. Figure 6 This is a schematic diagram of the structural composition of the image recognition-based dynamic advertising detection system in an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0011] Please see Figures 1 to 6 A dynamic detection method for advertisements based on image recognition is applied to image recognition scenarios. The dynamic detection method for advertisements based on image recognition includes: Step S11: Collect the advertisement to be detected, perform multi-scale segmentation on the advertisement to be detected, extract visual appearance features and optical flow motion features respectively, extract audio spectrum features in the advertisement to be detected simultaneously, and generate corresponding advertisement feature maps by combining visual appearance features and optical flow motion features. Step S12: Determine the feature change gradient matrix based on image recognition of the advertising feature map; perform time-series tracking on the feature change gradient matrix to extract the gradual trajectory features representing the illegal content from concealment to manifestation and the strobe trajectory features representing the periodic flashing of illegal elements. Step S13: Perform feature fusion on the gradual trajectory features and the flashing trajectory features to generate the corresponding advertising feature vector; map the advertising feature vector to a graph query node and input it into the violation knowledge graph to determine the topological distance between the graph query node and each violation semantic entity node in the violation knowledge graph; Step S14: Obtain the placement scenario of the advertisement to be detected, and determine the corresponding violation frame sequence by combining the corresponding traffic warning content and the topological distance. Based on the violation frame sequence, determine the corresponding multiple sub-violation contents, and determine the corresponding dynamic detection report by combining the advertising constraint factors of the advertisement to be detected.

[0012] refer to Figure 2 In step S11, the specific steps are as follows: S111: Based on the traversal of the advertising database, the advertisements to be detected are collected, and the multi-scale segmentation of the advertisements to be detected is triggered by the convolutional network. In the process of multi-scale segmentation, visual appearance features are extracted, and optical flow motion features are extracted simultaneously by the convolutional group. At this time, the visual appearance features represent the spatial texture and color distribution; the optical flow motion features represent the displacement field and motion edge of adjacent frame pixels. S112: Mark the audio stream of the advertisement to be detected, determine the audio spectrum features based on the analysis of the audio stream, and combine the visual appearance features and optical flow motion features for feature alignment and bidirectional fusion to output a fused feature group. The fused feature group is spliced ​​along the channel dimension and compressed by a 1x1 convolution kernel to generate an advertisement feature map that represents the multimodal spatiotemporal joint representation.

[0013] In the embodiments of this application, advertisements to be detected are collected based on the traversal of the advertisement database, and multi-scale segmentation of the advertisements to be detected is triggered by a convolutional network. In this process, visual appearance features are extracted, and optical flow motion features are extracted simultaneously by a convolutional group. At this time, visual appearance features represent spatial texture and color distribution; optical flow motion features represent the displacement field and motion edge of adjacent frame pixels, thus introducing visual appearance features and optical flow motion features.

[0014] At this point, the system, based on a distributed stream processing architecture, traverses the advertising data to be delivered or already delivered stored in the advertising database. According to the preset scheduling strategy and priority queue, it extracts the encapsulated data stream of the advertisement to be detected, and then performs a demultiplexing operation on the encapsulated data stream to separate the continuous video image frame sequence and the synchronous audio stream, thus completing the acquisition and preprocessing of the underlying data.

[0015] The above video image frame sequence is input into a multi-scale feature extraction convolutional network. During the forward propagation of the feature map, the convolutional network triggers a multi-scale segmentation mechanism. At this time, the network uses global pooling or large receptive field convolutional kernels to generate global feature patches in the deep semantic layer to capture the macro composition and overall color tone of the advertisement image. At the same time, it uses sliding window cropping or small-scale convolutional kernels to generate local feature patches in the shallow spatial layer to lock high-frequency detail areas.

[0016] During the multi-scale segmentation process, the convolutional network extracts the visual appearance features of each scale segment through the spatial filtering effect of its internal weights. At this time, the visual appearance features are specifically mapped to the image's "spatial texture features, such as the direction of local edges and the statistical characteristics of the gray-level co-occurrence matrix" and "color distribution features, such as the moment features of the color histogram and the joint probability distribution of color channels" in the data representation, thereby completing the accurate high-dimensional encoding of the static physical attributes of the advertisement image.

[0017] Within the synchronous temporal window where the convolutional network extracts static visual appearance features, a parallel optical flow extraction convolutional group is introduced to process adjacent temporal video frame pairs. This convolutional group concatenates the temporally adjacent previous frame and the current frame in the channel dimension as input. Through the cascaded operation of multiple layers of depthwise separable convolution and transposed convolution, a two-dimensional motion vector field is regressed pixel by pixel. At this time, the optical flow motion features are specifically represented in data representation as "adjacent frame pixel displacement field, that is, the magnitude and angle of the motion vector of each pixel in the horizontal and vertical directions" and "motion edge, that is, the boundary of the region where the motion vector in the displacement field undergoes a step change, reflecting the dynamic deformation boundary of the object contour", thereby completing the temporal modeling of the dynamic physical evolution process in the advertisement image.

[0018] Specifically, the system traverses the beauty advertising database and collects the video frame sequence of the beauty advertisement to be detected. This sequence contains continuous images of the facial contours of the advertising model slowly changing before and after using beauty products due to "hidden digital effects, such as keyframe-driven liquefaction deformation".

[0019] Multi-scale convolutional networks segment and extract appearance features for each frame: global-scale segments capture the overall proportions of the model's face ("three courts and five eyes") and the distribution of skin tone, while local small-scale segments precisely cover micro-areas such as the jaw angle and cheekbone, which are easily tampered with by digital effects. Because modern liquefaction effects have extremely high image fidelity, the pixel transitions in the tampered areas remain smooth. Therefore, the visual appearance features extracted in this stage exhibit high coherence and legitimacy in a single frame static view, making it difficult to directly expose traces of tampering.

[0020] Optical flow extraction convolution is the core method for exposing this violation. When digital effects fine-tune the model's jawline frame by frame, such as liquefaction processing that shrinks inward by 0.5 pixels per frame, although the visual appearance of a single frame is the same, in the temporal dimension of adjacent frames, the pixels in the jaw angle region produce a forced displacement that violates the laws of rigid body motion in human physics. At this time, the synchronously extracted optical flow motion features accurately capture this anomaly: in the pixel displacement field, a convergent displacement distribution is shown where the pixel vectors in the jaw angle region all point to the inside of the face; on the motion edge, a sharp deformed contour line is outlined due to the effect of the edge mask of the special effects, which does not naturally decay with the traction force of muscles.

[0021] Furthermore, the audio stream of the advertisement to be detected is labeled, and the audio spectrum features are determined based on the analysis of the audio stream. The visual appearance features and optical flow motion features are combined for feature alignment and bidirectional fusion to output a fused feature set. The fused feature set is spliced ​​along the channel dimension and compressed by a 1x1 convolution kernel to generate an advertisement feature map that represents the spatiotemporal joint representation of multiple modalities, ensuring the accuracy of the advertisement feature map.

[0022] At this point, the audio stream obtained by demultiplexing the advertisement to be detected is timestamped and framed and windowed to convert the time-domain audio signal to the frequency domain. Specifically, the log-Mel spectrum is extracted by fast Fourier transform, and a multi-scale one-dimensional convolutional network is introduced to encode the spectrum along the temporal dimension to aggregate acoustic information under different receptive fields at different times, thereby outputting audio spectrum features that characterize the audio acoustic properties and semantic excitation distribution.

[0023] Based on the timestamp marking mechanism, temporally continuous audio spectral features, visual appearance features, and optical flow motion features are aligned along the time axis. To address the issue that audio features typically have low temporal resolution while visual and optical flow features have high temporal resolution, a temporal upsampling and linear interpolation strategy is adopted to extend low-frequency audio spectral features in the time dimension to be strictly in sync with high-frequency visual and optical flow features, ensuring that acoustic excitation, spatial texture, and dynamic displacement field at the same moment are accurately aligned on the feature temporal coordinates.

[0024] The aligned visual appearance features, optical flow motion features, and audio spectrum features are input into the cross-modal bidirectional attention fusion module. In the forward fusion path, the audio spectrum features are used as the query matrix, and the visual appearance features and optical flow motion features are used as the key and value matrices, respectively, to calculate the attention weight distribution, enabling acoustic stimulation to retrieve and highlight relevant regions in visual space and dynamic displacement. In the reverse fusion path, the visual and optical flow joint features are used as the query matrix, and the audio spectrum features are used as the key and value matrices, so that visually anomalous regions can also inversely enhance the response weight of audio semantics in the corresponding time period. Through the above bidirectional cross-attention operation, visual-audio fusion features and optical flow-audio fusion features containing deep intermodal interaction information are output, which together constitute the fusion feature group.

[0025] The modal fusion features in the above-mentioned fusion feature group are spliced ​​along the channel dimension to form a multi-channel joint feature tensor with high dimensionality redundancy. This joint feature tensor is then guided to perform cross-channel linear combination and dimensionality reduction compression through a 1x1 convolution kernel. While retaining key multimodal interaction information, redundant features are eliminated, thereby generating an advertising feature map that represents the spatiotemporal joint representation of multimodality. Specifically, the system marks and analyzes the audio stream of the beauty advertisement. The extracted audio spectrum features not only include the spectrum envelope of the background music, but more importantly, they include the high-frequency speech stimulation features of the narrator at specific moments, such as the spectrum response of highly persuasive speech segments like "instant lifting" and "V-face appearance" that appear between the 8th and 12th seconds.

[0026] The feature alignment mechanism played a precise spatiotemporal anchoring role. In this beauty advertisement, the audio spectrum feature moment when the narrator shouted "V face appears" was precisely aligned with the optical flow motion feature moment when the model's jaw angle area was undergoing a hidden contraction in the video stream through timestamp and upsampling interpolation. This ensured that the semantic excitation of the audio and the tampering action of the visual were strictly synchronized on the millisecond time axis, avoiding subsequent fusion failure caused by audio-visual misalignment.

[0027] The bidirectional fusion mechanism is key to revealing this hidden effect: In the forward fusion path, the audio spectrum features of "V-face appearance" serve as a query matrix, actively retrieving and significantly increasing the response weight of the optical flow displacement field in the mandibular angle region of the visual feature map at the same time, so that the originally weak and hidden local pixel displacement field is significantly amplified under the excitation of acoustic semantics; In the reverse fusion path, the abnormal optical flow movement edge in the mandibular angle region serves as a visual anomaly, which inversely enhances the attention weight of semantics related to "slimming face" and "contour" in the audio features of the corresponding time period; The resulting fusion feature group deeply binds and mutually confirms "inducing speech" and "hidden liquefied optical flow" at the feature level.

[0028] The highly correlated audio-visual fusion feature groups are spliced ​​along the channels. Due to the high feature dimension and the redundant mapping generated by cross-attention, a 1x1 convolutional kernel performs cross-channel weighted summation and compression dimensionality reduction. The resulting advertising feature map is no longer a simple appearance texture or isolated displacement vector, but a joint representation that deeply integrates acoustic intent and spatiotemporal deformation. In this feature map, the tiny optical flow anomalies caused by the concealment effect are amplified and solidified into high-response activation values ​​by audio semantic excitation, providing an input source with extremely high signal-to-noise ratio for the subsequent step S12 to extract the gradient trajectory and stroboscopic trajectory.

[0029] refer to Figure 3 In step S12, the specific steps are as follows: S121: Perform image recognition on the advertising feature map, and determine the gradient matrix of feature changes that characterize the intensity and direction of spatiotemporal features by measuring the absolute difference between the feature response values ​​of adjacent time frames during the recognition process; S122: Input the feature change gradient matrix into the long short-term memory network and perform time-series tracking based on time steps to extract the gradual trajectory feature of the feature response value continuously increasing over time. This gradual trajectory specifically presents the trajectory of the illegal content from being hidden to being revealed. S123: The time-series fluctuation signal of the gradient matrix of the marked feature changes is combined with the Fourier transform mechanism to trigger the frequency domain transformation of the time-series fluctuation signal, thereby extracting the stroboscopic trajectory feature with a set periodic high-frequency peak in the frequency domain. This stroboscopic trajectory feature presents the trajectory of the periodic flashing of the illegal element.

[0030] In the embodiments of this application, image recognition is performed on the advertising feature map, and the absolute difference between the feature response values ​​between adjacent time frames is used to determine the feature change gradient matrix that characterizes the intensity and direction of spatiotemporal features. The feature change gradient matrix is ​​introduced.

[0031] At this point, the advertising feature map representing the multimodal spatiotemporal joint representation output in step S112 is input into a temporal three-dimensional convolutional network or a spatiotemporal attention recognition network. The network performs feature response mapping and high-dimensional nonlinear transformation on the advertising feature map along the spatial and temporal dimensions to extract and activate the semantic concepts and dynamic patterns contained therein, thereby outputting a highly abstract feature response tensor containing the temporal dimension. Each element in the feature response tensor represents the confidence and activation intensity of a specific semantic pattern at the corresponding spatiotemporal location.

[0032] Based on the aforementioned feature response tensor, a difference operation is performed on its adjacent temporal frame slices in the time dimension. That is, the difference between the feature response tensor of the current t-th frame and the feature response tensor of the previous t-1 frame at the corresponding spatial position is calculated. In order to eliminate the interference of positive and negative fluctuations in the feature response value and to purely quantify the amplitude of feature changes, the absolute value of the difference is taken to form a difference absolute value tensor sequence. This difference absolute value tensor sequence accurately strips away the static background and constant semantics that persist in the video frame, retaining only the feature increments that change in the temporal stream.

[0033] The difference absolute value tensor sequence is recombined with the gradient direction information of the original feature response tensor in the time dimension. At this time, not only is the difference absolute value used to represent the drasticness of feature change, but also the sign information of the difference operation (positive sign represents the enhancement of feature response, negative sign represents the weakening of feature response) is introduced to represent the directional attribute of feature change. Then, the drasticness is used as the gradient magnitude, and the sign of change is associated with the spatial position as the gradient orientation. The two are jointly reconstructed into the feature change gradient matrix. In terms of mathematical form, the feature change gradient matrix not only maps the drasticness of feature response abrupt or gradual change in the spatiotemporal domain, but also indicates whether the feature evolution shows a convergent or divergent trend, thus providing an accurate spatiotemporal dynamic basis for the subsequent extraction of gradual trajectory features and stroboscopic trajectory features.

[0034] Specifically, the temporal 3D convolutional network identifies and maps the feature map of the beauty advertisement. Since the S112 has performed bidirectional audio-visual fusion in the early stage, when the advertisement plays to the area of ​​the model's jaw angle where there is a hidden liquefaction effect, accompanied by the voice "V face appears", the feature response tensor of the network at the corresponding spatiotemporal location will be strongly activated, generating a high-confidence response value. This response value is actually a quantitative expression of the high-order semantic of "suspicious facial deformation with audio-visual synchronization".

[0035] The system calculates the absolute difference of the feature response value between adjacent frames. Since the digital effects of this beauty advertisement use a hidden gradient liquefaction strategy of 0.5 pixels per frame, there is almost no difference in visual appearance in a single frame. However, in this high-dimensional feature space, the feature response values ​​of the corresponding jaw angle region in adjacent frames will show a small but constant difference. By taking the absolute value, the positive and negative oscillations of the feature response value caused by the nonlinear activation function of the network are removed, and the feature increment amplitude brought about by the "hidden deformation" at each moment is purely extracted. Thus, the weak tampering action that was originally submerged in the time axis is converted into a continuous non-zero difference absolute value sequence.

[0036] The system reconstructs the gradient matrix of feature changes. For this hidden liquefaction effect, the absolute value of the difference represents the "intensity" of the feature changes in the mandibular angle region, which is manifested here as a continuous low-amplitude but persistent gradient. The "direction" determined by combining the original symbol information accurately indicates the physical meaning of the feature changes. For example, if the difference sign is continuously negative, it indicates that the optical flow displacement field in this region continues to contract towards the inner side of the face, that is, the mandibular contour converges and the feature response shows a weakening trend. The feature change gradient matrix constructed in this way not only keenly captures the weak local tampering action, but also clearly outlines the spatiotemporal dynamics of the dynamic evolution process of "hidden digital special effects slowly slimming the mandibular angle" through the joint representation of gradient amplitude and gradient orientation.

[0037] Furthermore, the feature change gradient matrix is ​​input into the long short-term memory network, and time-series tracking is performed based on time steps to extract the gradual trajectory feature of the feature response value continuously increasing over time. This gradual trajectory specifically presents the trajectory of the illegal content from concealment to manifestation, thus introducing the gradual trajectory feature.

[0038] At this point, the feature change gradient matrix determined in step S121 is decomposed into a continuous gradient tensor sequence according to the time step size, and the sequence is sequentially input into the input gate channel of the Long Short-Term Memory (LSTM) network. The LSTM network uses its unique forget gate mechanism, input gate mechanism and output gate mechanism to perform nonlinear gating screening on the gradient features input at the current time to determine which instantaneous gradient fluctuations should be filtered out and which gradient features with continuous evolution potential should be retained and written into the cell state.

[0039] During the time-step iteration of the Long Short-Term Memory (LSTM) network, the network not only relies on the gradient matrix input at the current time step, but also transmits historical context information across time steps through hidden state vectors and cell state vectors. The gradient features of the current time step and the historically accumulated cell states are adaptively weighted and nonlinearly activated, thereby establishing a dynamic temporal tracking link in the time dimension. This enables the network to perceive and remember the evolution trend and pattern of feature changes over a long time span, rather than just focusing on isolated gradient mutations in a single frame.

[0040] Based on time-series tracking, trend discrimination and feature mapping are performed on the hidden state vectors output by the Long Short-Term Memory Network step by step. Since the transformation of illegal content from concealment to manifestation is often accompanied by the continuous activation and accumulation of feature responses in a specific spatial region, the network selectively retains gradient changes with the same sign direction and the magnitude of the gradient changes with a cumulative trend through its long-term memory cells. In this way, the gradual trajectory feature of the feature response value continuously increasing over time is extracted from the output state. This gradual trajectory feature constitutes a directional evolution curve with a clear start time and end time in the feature space. In terms of physical semantics, it precisely presents the complete dynamic trajectory of illegal content from the initially imperceptible concealed state to the final manifestation.

[0041] Specifically, the gradient matrix representing the slow inward contraction of the model's jawline is input frame by frame into the Long Short-Term Memory (LSTM) network. Since normal head movements, lighting changes, or background flickering are inevitable in the advertisement, these disturbances are represented as irregular random positive and negative gradient pulses in the gradient matrix. The forgetting gate of the LTM network plays a crucial role at this point, assigning extremely low retention weights to isolated, non-directional high-frequency random gradient pulses, thus filtering out most of the interference from normal physical motion at the input.

[0042] As the time step progresses, the Long Short-Term Memory (LSTM) network continuously tracks the dynamic features of the beauty advertisement. When the advertisement plays from the 5th to the 12th second, although the gradient magnitude of the inward displacement of the jaw angle in each frame is extremely small, for example, only a gradient difference of 0.01, the displacement direction remains highly consistent over dozens of consecutive frames. The cell states of the network accumulate this tiny and unidirectional gradient change into the historical memory frame by frame through the transmission of hidden state vectors, forming a continuous temporal tracking link spanning 7 seconds. This gives the weak signal that was originally submerged in the noise of a single frame a significant statistical accumulation advantage in the time dimension.

[0043] The network extracts the gradual trajectory feature of the feature response value continuously increasing over time. For this hidden liquefaction effect, the extracted gradual trajectory feature in the feature space is a smooth trajectory curve that monotonically decreases (due to the contraction of the mandibular angle, the corresponding feature response is negatively and continuously increasing) and the gradient accumulation value crosses the preset threshold boundary. This gradual trajectory feature clearly shows the complete illegal evolution trajectory of the model's facial contour from the original physiological state at the 5th second (hidden period), through dozens of frames of extremely small digital effects push and pull (quantitative change period), and finally completely presents the "V face" shape that violates the physiological structure at the 12th second (manifestation period). Thus, the originally highly deceptive single-frame visual disguise is completely revealed as a dynamic illegal fact with a clear intention to tamper under the perspective of long-term temporal dependence.

[0044] Therefore, the temporal fluctuation signal of the gradient matrix of the marked feature changes is used, and the frequency domain transformation of the temporal fluctuation signal is triggered by the Fourier transform mechanism. This allows the extraction of the stroboscopic trajectory feature with a set periodic high-frequency peak in the frequency domain. The stroboscopic trajectory feature presents the trajectory of the periodic flashing of the illegal element. The advertising feature map is introduced, and the image recognition of the advertising feature map is realized, which improves the accuracy of the gradual trajectory feature and the stroboscopic trajectory feature.

[0045] At this point, the feature change gradient matrix output in step S121 is processed and labeled in the time dimension. Then, the magnitude of the gradient matrix is ​​global pooled or activated by setting a spatial threshold along the time axis. The high-dimensional spatiotemporal gradient tensor is reduced and projected into a one-dimensional time series scalar signal, thereby obtaining a time-series fluctuation signal that represents the fluctuation of the feature change gradient over time. At the same time, the pulse surge points and drop points in the time-series fluctuation signal are timestamped to anchor the time coordinates of feature mutation.

[0046] The one-dimensional time-series fluctuation signal marked with time-series coordinates is input into the Fast Fourier Transform (FFT) engine, triggering a mapping transformation from the time domain to the frequency domain. By decomposing the continuous time-series fluctuation signal into a superposition of a series of sine and cosine functions with different frequencies, amplitudes, and phases, the transient pulses that were originally intertwined and overlapping on the time axis are orthogonally decoupled from the continuous background noise. Thus, the complex periodic patterns in the time domain that are difficult to identify with the naked eye or conventional time-series networks are explicitly converted into discrete spectral energy distributions in the frequency domain.

[0047] Based on the spectral energy distribution obtained from the above frequency domain transformation, spectral scanning and peak detection are performed according to the preset threshold range of illegal stroboscopic frequencies. When the energy amplitude at a specific frequency _bin_ is significantly higher than its adjacent background noise floor, and the frequency falls within the "set high-frequency band, which corresponds to the flicker frequency outside the human visual persistence limit, i.e., the frequency band that the human eye cannot perceive but the machine can perceive", the peak is determined to be a set periodic high-frequency peak. Then, the frequency value, energy amplitude, and timing trigger phase obtained by inverse transformation or phase spectrum analysis of the high-frequency peak are extracted and combined and encoded into a stroboscopic trajectory feature. In terms of data representation, the stroboscopic trajectory feature accurately presents the dynamic trajectory of the illegal element flashing periodically with an extremely short duty cycle in the original time sequence.

[0048] Specifically, the system performs dimensionality reduction and labeling on the feature change gradient matrix. In this beauty advertisement, in addition to the slow liquefaction effect, if the advertiser inserts local skin smoothing effect frames periodically with extremely short single-frame durations in the video stream to enhance the "wrinkle removal" effect, then in the feature change gradient matrix, the local skin smoothing area will generate a huge gradient amplitude jump at the moment of alternation between normal frames and effect frames. The system aggregates these jumps into a one-dimensional temporal fluctuation signal and marks these sharp pulse points representing the instantaneous smoothing and instantaneous recovery of texture in the signal.

[0049] Since the single-frame skin smoothing effect lasts only about 40 milliseconds each time, which is equivalent to one frame, it appears as extremely dense and weak spike-like fluctuations in the continuously played time-domain signal. Conventional time-series networks can easily treat it as Gaussian noise and filter it out. At this point, the Fourier transform mechanism intervenes to decompose this time-series fluctuation signal into the frequency domain. In the frequency domain, those transient spikes that appear chaotic in the time domain are transformed into high-energy resonance peaks concentrated on the "specific frequency _bin_, which is the frequency of the video frame rate / 10" on the spectrogram due to their strict periodic insertion, such as inserting one frame of skin smoothing effect every 10 frames.

[0050] The system scanned the spectrum and detected a set periodic high-frequency peak at a specific frequency _bin_ with an energy amplitude far exceeding the background noise. The system extracted the feature parameters of this high-frequency peak and generated a stroboscopic trajectory feature. In this beauty scene, the stroboscopic trajectory feature not only includes a 25Hz flicker frequency and high energy amplitude, but also reconstructs the precise rhythm of the illegal skin smoothing effect "flash-hide-flash" on the video timeline through phase information. Thus, in addition to the macroscopic slow liquefaction gradient trajectory, it further reveals the illegal operation trajectory of the advertisement injecting false skin texture effects through high-frequency stroboscopic at the microscopic time scale, forming a complete closed-loop evidence chain of multiple concealed tampering techniques of the advertisement.

[0051] refer to Figure 4 In step S13, the specific steps are as follows: S131: Perform bilinear pooling-based feature fusion on the gradient trajectory features and the flashing trajectory features to enhance the nonlinear interaction response between heterogeneous trajectories, thereby generating the corresponding advertising feature vector, and mapping the advertising feature vector to graph query nodes through a linear mapping layer; S132: Input the graph query node into the violation knowledge graph containing violation semantic entity nodes and entity association edges, and perform message passing and aggregation calculation based on the graph neural network to determine the topological distance between the graph query node and each violation semantic entity node in the violation knowledge graph in the low-dimensional embedding space. The topological distance is negatively correlated with the semantic association tightness between the graph query node and the violation semantic entity node, that is, the smaller the topological distance, the higher the semantic association tightness.

[0052] In the embodiments of this application, feature fusion based on bilinear pooling is performed on the gradual trajectory features and the stroboscopic trajectory features to enhance the nonlinear interaction response between heterogeneous trajectories, thereby generating the corresponding advertising feature vector. The advertising feature vector is then mapped to a graph query node through a linear mapping layer, thus introducing the advertising feature vector.

[0053] At this point, the gradient trajectory features extracted in step S122 and the flicker trajectory features extracted in step S123 are used as the input dual signals of the bilinear pooling module. In this module, instead of simple vector concatenation or element-wise addition, the two feature vectors are subjected to outer product operations to construct a second-order cross matrix of the feature vectors. The cross matrix is ​​flattened and subjected to sign square root processing, and combined with normalization operations to compress extreme values, thereby realizing the full pairwise combination of the two heterogeneous trajectory features in all dimensions. This bilinear pooling process forces each dimension in the gradient feature space to perform a multiplicative interaction with each dimension in the flicker feature space, thereby significantly strengthening the latent nonlinear interaction response between heterogeneous trajectories, so that composite violation patterns such as "specific flicker accompanied by specific gradient" are expressed intensified, and finally a high-order fused feature vector is output.

[0054] The high-order fusion feature vector obtained by bilinear pooling is input into the multilayer perceptron module. Through linear transformation of the fully connected layer and layer-by-layer extraction of nonlinear activation functions, the high-dimensional redundant second-order cross information is reduced in dimensionality and feature selection is performed. The multilayer perceptron module adaptively filters out invalid noise combinations generated during the cross process and retains the core interaction features with high discriminative power for identifying violations. The hidden layer output with compact representation capability after filtering is determined as the advertising feature vector of the advertisement to be detected. Mathematically, the advertising feature vector is a comprehensive semantic encoding of the multimodal and multi-timescale violation dynamics in the original advertisement image.

[0055] Based on the preset vector space dimension and metric space settings of the violation knowledge graph, the aforementioned advertising feature vectors are input into a linear mapping layer. This linear mapping layer consists of unbiased affine transformation matrices. Through parameterized spatial rotation and scaling operations, the advertising feature vectors are projected without distortion from the original visual temporal feature space to a graph embedding space consistent with the violation knowledge graph. The resulting vector is the graph query node. The coordinate position of the graph query node in the graph embedding space represents the retrieval anchor point of the current advertisement to be detected in the semantic topology of the knowledge graph, which is used to calculate its proximity relationship with existing violation semantic entity nodes in the graph.

[0056] Specifically, the system performs bilinear pooling fusion on the extracted gradient trajectory features of "slow inward convergence of the jaw angle" and the stroboscopic trajectory features of "local high-frequency single-frame skin smoothing flashes." In the complex fraudulent techniques of beauty advertising, these two effects often appear synergistically: while advertisers slim the contours, they use high-frequency flash frames to cover up the local texture stretching distortion caused by liquefaction. If only simple splicing is used, the model can hardly detect the causal relationship between the two. However, through bilinear pooling with outer product operation, the dimension representing "jaw angle displacement direction" in the gradient features and the dimension representing "texture smoothing frequency" in the flash features are multiplicatively crossed, so that the combination mode of "facial contour deformation in a specific direction" and "skin texture repair flash frames at a specific frequency" produces a very strong nonlinear interactive response, and the interaction result is significantly amplified in the feature space.

[0057] The high-dimensional second-order cross matrix output by bilinear pooling is fed into a multilayer perceptron. At this time, the perceptron network automatically suppresses cross-combination noise unrelated to facial tampering, such as "slow gradation of background light" and "high-frequency flashing of product logo", by adjusting the weight parameters. At the same time, it highly preserves the interactive features with strong violation indication significance, such as "continuous displacement of facial contour" and "high-frequency texture flashing in the same area". Finally, it compresses and generates an advertising feature vector with fixed dimensions and highly focused on the semantics of "concealed facial tampering".

[0058] To connect this visual temporal semantics with the pre-constructed violation rule network, the advertising feature vector is spatially projected through a linear mapping layer. The projected graph query node is precisely located in the high-dimensional topological space of the violation knowledge graph, in the intersection neighborhood of violation semantic entity clusters such as "excessive facial contouring" and "fake skin texture presentation". This seamlessly transforms the underlying pixel-level gradation and frame-level flashing physical observations into semantic-level query probes that can be used for logical reasoning and topological distance calculation in the knowledge graph.

[0059] Furthermore, the graph query node is input into the violation knowledge graph containing violation semantic entity nodes and entity association edges. Message passing and aggregation calculations are performed based on the graph neural network to determine the topological distance between the graph query node and each violation semantic entity node in the violation knowledge graph in the low-dimensional embedding space. The topological distance is negatively correlated with the semantic association tightness between the graph query node and the violation semantic entity node. That is, the smaller the topological distance, the higher the semantic association tightness, thus further controlling the topological distance.

[0060] At this point, the graph query node generated in step S131 is used as an external query vector and injected into the graph structure data of the pre-constructed violation knowledge graph. This violation knowledge graph consists of "violation semantic entity nodes" that represent specific violation concepts, such as specific violation actions, prohibited items, violation intentions, etc., and "entity association edges" that connect the nodes, such as causal association, temporal association, spatial dependence, and other relational predicates. In mathematical representation, the input operation is to place the initial embedding vectors of the graph query node and the existing nodes in the graph in the same low-dimensional manifold space to establish a cross-modal retrieval initialization state.

[0061] On the extended graph structure containing graph query nodes, an iterative computation mechanism of the graph neural network is triggered. In each iteration, for the target node in the graph, including the graph query node and each non-compliant semantic entity node, message passing is performed along the topological path of the entity association edge, that is, the current feature vector and edge attribute features of its neighboring nodes are extracted into message packets. At the target node, aggregation computation is performed, and a specific aggregation function is used to fuse and update the received multi-source message packets with the target node's own feature vector. Through multi-hop iterative message passing, the high-order dynamic semantic features of the graph query node can diffuse and permeate along the graph topology, while the prior rule features of each non-compliant semantic entity node are also backpropagated to the graph query node, thereby achieving adaptive smoothing of global graph features and contextual semantic enhancement.

[0062] After sufficient iterative updates by the graph neural network, the final embedding vectors of the graph query nodes and the final embedding vectors of each non-compliant semantic entity node are extracted. The metric distance between the two is calculated in the low-dimensional embedding space as the topological distance, such as using Euclidean distance, cosine distance or Mahalanobis distance. This topological distance essentially quantifies the proximity of the dynamic features of the advertisement to be detected represented by the graph query node and the prior violation patterns represented by each non-compliant semantic entity node in the manifold space.

[0063] The judgment logic is established based on the distribution characteristics of manifold space. That is, the topological distance is negatively correlated with the semantic association between the graph query node and the illegal semantic entity node. The physical meaning is that if the topological distance between the graph query node and a certain illegal semantic entity node in the low-dimensional embedding space is smaller, it means that the two are closer in feature distribution and topological context. After message passing, the feature overlap is higher, which indicates that the advertisement to be detected is more closely associated with the illegal semantics and the confidence of the violation is stronger. Conversely, the larger the topological distance, the more distant the semantics are and the weaker the suspicion of violation.

[0064] Specifically, the graph query node derived from the fusion and mapping of the beauty advertisement's "gradient liquefaction of the jawline" and "high-frequency flashing skin smoothing" is input into the violation knowledge graph. This knowledge graph has pre-set violation semantic entity nodes such as "excessive deformation of facial contours", "fake skin texture presentation", and "creating appearance anxiety". There is an entity association edge between "excessive deformation of facial contours" and "fake skin texture presentation", and its relation predicate is labeled as "temporal accompanying modification", which constitutes a priori knowledge network for cheating in the beauty field of special effects.

[0065] The graph neural network initiates message passing and aggregation computation; the composite features of "local high-frequency flashing frames" and "continuous inward convergence displacement" contained in the graph query node are passed along the entity association edges to the "excessive facial contour deformation" and "false skin texture presentation" nodes; at the same time, the prior features of the "false skin texture presentation" node are also passed back along the edges to the graph query node, and are weighted and aggregated through the attention mechanism; since the dynamic features of the advertisement to be detected happen to match the "temporal co-occurrence" relationship between the two, after the graph query node aggregates the messages from these two adjacent nodes, its feature vector is strongly attracted in the manifold space and shifted towards the center of the violation subgraph cluster.

[0066] The system calculates the topological distance between the updated graph query node and each entity in the graph. At this point, the low-dimensional embedding vectors of the graph query node and the two nodes "excessive facial contour deformation" and "false skin texture presentation" are extremely close in space, and the calculated Euclidean topological distance is very small. However, the embedding vectors of nodes that are completely unrelated to nodes in the graph, such as "display of prohibited drugs" and "vulgar pornographic innuendo", are far apart in space, and the topological distance is extremely large.

[0067] Semantic tightness is determined based on negative correlation mapping; extremely small topological distance directly maps to extremely high semantic association tightness. The system is thus convinced that the beauty advertisement represented by the graph query node is not a normal model display, but a violation pattern that highly matches "using hidden digital special effects to tamper with facial contours and fake skin texture". This high tightness obtained through graph structure reasoning completely avoids the defect of relying solely on visual classifiers and being easily deceived by high-fidelity special effects, and achieves a leapfrog locking from isolated dynamic features to the system's violation semantics.

[0068] refer to Figure 5 In step S14, the specific steps are as follows: S141: Obtain the placement scenario of the advertisement to be detected, determine the corresponding scenario tag based on the identification of the placement scenario, retrieve the traffic warning content corresponding to the scenario tag, and combine the danger level weight of the traffic warning content to apply a weighted penalty to the topological distance, thereby obtaining a dynamic violation confidence score. S142: Filter time-series frames whose dynamic violation confidence exceeds a preset threshold, construct a violation frame sequence from each time-series frame, segment the violation elements in the violation frame sequence according to their spatiotemporal continuity, determine multiple corresponding sub-violation content, and combine the advertising constraints of the advertisement to be detected under the current delivery scenario, such as regulatory requirements and audience limitation, to match the corresponding violation risk level and penalty avoidance strategy for each sub-violation content, thereby generating a dynamic detection report. This dynamic detection report includes violation timestamp, violation type semantics, and risk avoidance suggestions.

[0069] In the embodiments of this application, the placement scenario of the advertisement to be detected is obtained, the corresponding scenario tag is determined based on the identification of the placement scenario, the traffic warning content corresponding to the scenario tag is retrieved, and the topological distance is weighted and penalized in combination with the danger level weight of the traffic warning content, thereby obtaining a dynamic violation confidence level, which is compatible with the overall consideration of the placement scenario and ensures the accuracy of the corresponding scenario tag.

[0070] At this point, the system obtains the preset placement scenario configuration data of the advertisement to be detected through the business distribution interface, parses and semantically recognizes the configuration data, and extracts metadata that represents the target audience, media carrier and display environment of the advertisement. The metadata is then input into a pre-built scenario tag classification system, and through text matching or semantic mapping, it is accurately anchored to the corresponding scenario tag, such as a channel for minors, a medical and health information stream, or a pan-entertainment splash screen, thereby completing the discrete tag labeling of the business placement context.

[0071] Based on the aforementioned determined scenario tags, the scenario-warning mapping table in the risk control strategy engine is retrieved, and the traffic warning content bound to the scenario tag is dynamically called up. The traffic warning content includes a list of strictly regulated illegal semantic entities and their corresponding danger level weights in the specific delivery scenario. The danger level weights reflect the sensitivity and tolerance limit of specific illegal semantics in triggering regulatory penalties or user complaints in the current scenario.

[0072] Based on the retrieved traffic warning content, the danger level weights corresponding to each illegal semantic entity node associated with the graph query node are extracted, and the danger level weights are introduced as penalty coefficients into the topological distance calculated in step S132. At this time, for the same topological distance value, if the corresponding illegal semantic entity has a higher danger level weight in the traffic warning content of the current scenario, a multiplicative compression penalty with a scaling factor of less than 1 is applied to the topological distance, so that the distance representation of high-risk violations is forcibly narrowed under the same feature similarity.

[0073] Based on the corrected topological distance after being weighted and penalized by the risk level, the distance metric is converted into a probabilistic confidence index through a preset mapping function, such as a negative exponential function or a Gaussian radial basis function. Since the topological distance is negatively correlated with the semantic relevance, and the high-risk weight applies a compression penalty to the topological distance, the smaller the penalized topological distance, the higher the dynamic violation confidence of the mapping output. This dynamic violation confidence combines the objectivity of algorithm feature matching with the subjective rigidity of risk control in business scenarios, and serves as the quantitative basis for the final violation judgment.

[0074] Specifically, the system obtains the placement scenario configuration data of the beauty advertisement, identifying that it is planned to be placed on two channels: "Makeup Tutorial Zone for Adolescent Women" and "Open Screen for All Audiences". Through the scenario classification system, the former is labeled with the highly sensitive scenario tag "[Exclusive to Minors / Young Women]", while the latter is labeled with the regular scenario tag "[Open Screen for General Audiences]", thus completing the division of the business audience context. The system retrieves traffic warning content based on the scenario tags. For the scenario tag "[Exclusive to Minors / Young Women]", the traffic warning content issued by the risk control strategy engine clearly includes illegal semantic entities such as "creating appearance anxiety" and "false skin texture presentation", and its danger level weight is set to the highest level H (weight coefficient 0.1, the smaller the value, the heavier the penalty). For the scenario tag "[Open Screen for General Audiences]", the danger level weight of the same illegal semantics may only be set to the medium level M (weight coefficient 0.5).

[0075] Suppose that step S132 calculates the original topological distance between the graph query node and the "excessive facial contour deformation" entity node to be 0.6. When the target audience is "[Exclusively for minors / young women]", this topological distance of 0.6 is multiplied by a risk level weight of 0.1 for weighted penalty, and the corrected topological distance drops sharply to 0.06. However, when the target audience is "[Splash screen for general audiences]", the same topological distance of 0.6 is multiplied by a weight of 0.5, and the corrected topological distance is 0.3. This weighted penalty mechanism is essentially a mathematical manifestation of a zero-tolerance attitude towards potential violations in highly sensitive scenarios.

[0076] The corrected topological distance is input into the negative exponential mapping function. In the scenario of "specifically for minors / young women", the extremely small corrected topological distance of 0.06 is mapped to a dynamic violation confidence level as high as 0.95. However, in the scenario of "splash screen for a general audience", the corrected topological distance of 0.3 may only map to a dynamic violation confidence level of 0.65. It can be seen that even with the same technique of "hidden digital effects to tamper with facial contours" and the same feature extraction results, due to the different delivery scenarios, the system outputs differentiated dynamic violation confidence levels by integrating traffic warning risk levels for weighted penalties. This ensures stricter interception and crackdown on illegal beauty ads targeting vulnerable audiences, and achieves scenario-adaptive and dynamically accurate matching of risk control strategies.

[0077] Furthermore, time-series frames with dynamic violation confidence exceeding a preset threshold are selected and combined to form a violation frame sequence. Based on the spatiotemporal continuity of the violation elements in the violation frame sequence, multiple sub-violation contents are identified. Combining the advertising constraints of the advertisement under test in the current delivery scenario, such as regulatory requirements and audience limitations, corresponding violation risk levels and penalty avoidance strategies are matched for each sub-violation content, thereby generating a dynamic detection report. This dynamic detection report includes a violation timestamp, violation type semantics, and risk avoidance suggestions, further controlling the violation frame sequence and fully considering the advertising constraints of multiple sub-violation contents and the advertisement under test, thus improving the accuracy of the dynamic detection report.

[0078] At this point, through temporal positioning, spatiotemporal segmentation, and business constraint mapping, a structured detection report guiding the review and handling of advertisements is finally output. The following section provides a detailed explanation of each sub-step in this process, following the standard specifications for writing patents of large companies, and finally provides specific explanations in conjunction with typical violation scenarios of beauty advertisements.

[0079] The dynamic violation confidence score calculated in step S141 is compared frame by frame along the time axis with a preset violation judgment threshold. Continuous or discrete time-series frames with a dynamic violation confidence score greater than or equal to the threshold are extracted. The time-series frames that meet the threshold condition are rearranged in their original time order to construct a violation frame sequence. This operation realizes a dimensionality reduction mapping from the probability judgment of the global advertising dimension to the precise positioning of local time-series segments, stripping away compliant images and focusing on the spatiotemporal interval of violations.

[0080] Based on the feature change gradient matrix and optical flow motion features extracted in the previous steps, spatial tracking and temporal correlation are performed on the illegal elements in the illegal frame sequence. According to the connected component distribution of the illegal elements in the spatiotemporal dimension, spatiotemporal continuity segmentation is performed. At this time, when a certain illegal element has a smooth transition and physical coherence in its spatial position, motion trajectory and feature response in a continuous frame sequence, it is assigned to the same spatiotemporal connected component. Conversely, if there is a sudden change in spatial position, truncation of motion vector or long-term extinction of feature response, it is segmented according to this breakpoint. In this way, the illegal frame sequence is deconstructed into multiple sub-illegal contents that are independent of each other in the spatiotemporal dimension and highly continuous internally. Each sub-illegal content corresponds to an independent illegal dynamic evolution process.

[0081] The system acquires the advertising constraints of the advertisement under test in the current delivery scenario. These constraints include "platform regulatory requirements, such as prohibited words in advertising law and special permit requirements" and "audience restrictions, such as mechanisms for the protection of minors." It then performs rule matching between the core semantics of each sub-violation and the aforementioned advertising constraints, dynamically assigning each sub-violation a corresponding violation risk level based on the severity of the constraints, such as high-risk blocking, medium-risk traffic restriction, and low-risk rectification. Simultaneously, based on the causal relationship between risk level and violation semantics, it retrieves corresponding penalty avoidance strategies from the strategy library, such as removing specific frame segments, replacing the violating local area, and supplementing qualification certificates, forming a remediation guide for business operations.

[0082] By aggregating the analysis results of all the above sub-violations, a dynamic detection report is generated according to a preset structured template. The time interval of the violation frame sequence corresponding to the sub-violation is converted into a violation timestamp, the graph entities and constraint rules matched by the sub-violation are converted into violation type semantics, and the corresponding penalty avoidance strategies are converted into risk avoidance suggestions. This dynamic detection report not only provides a chain of evidence for the qualitative nature of the violation, but also provides quantitative spatiotemporal positioning and executable rectification solutions, completing the full-link output from signal detection to business closed loop.

[0083] Specifically, the system compares the dynamic violation confidence level with a preset threshold, such as 0.85, frame by frame. In the first 5 seconds of the beauty advertisement, when the model is displaying normally, the confidence level is below 0.2 and is filtered out. From the 6th second, the jawline slowly liquefies until the V-shaped face is fully formed at the 15th second, and the confidence level curve gradually rises and continues to exceed 0.85. The system then filters out the continuous video frames from the 6th to the 15th second, as well as the isolated video frame with the skin smoothing effect flashing in the single frame at the 18th second, which together constitute the violation frame sequence of the advertisement, thus achieving precise identification of the hidden tampering time window.

[0084] The system performs spatiotemporal continuity segmentation on the violation frame sequence. By tracing the optical flow motion features, it was found that between the 6th and 15th seconds, there is a continuously inward gradient pixel displacement field in the jaw angle region. This spatiotemporal connected region was segmented into "first sub-violation content: facial contour gradient liquefaction". In the isolated frame at the 18th second, its optical flow features show transient high-frequency pulses and have no displacement field continuity with the preceding and following frames. This isolated spatiotemporal block was segmented into "second sub-violation content: local skin texture single-frame flashing forgery". Through spatiotemporal continuity segmentation, the complex cheating method was successfully decoupled into two dynamic violation units with independent mechanisms.

[0085] The system incorporates advertising constraints specific to the current campaign scenario of "exclusively for minors / young women." Based on the regulatory requirement of "strictly prohibiting the creation of appearance anxiety" and the audience limitation of "minors being easily misled" in this scenario, the system determines that the "first sub-violation content" directly touches the red line of appearance anxiety, matching it as a "high-risk level," and generates a penalty avoidance strategy of "physically cutting out the footage from the 6th to the 15th second or replacing it with unaltered material." While the "second sub-violation content" is fraudulent, its impact is relatively small, matching it as a "medium-risk level," and generating an avoidance strategy of "recommending to reduce the frequency of skin smoothing flashes or adopting compliant uniform gradient."

[0086] The system integrates the above information to generate a dynamic detection report. The report clearly lists the violation timestamps as [00:06-00:15] and [00:18], and the violation types are semantically "using gradient effects to tamper with facial contours to create appearance anxiety" and "using strobe effects to fake skin texture," respectively. It also includes corresponding risk avoidance suggestions. This report not only allows advertising reviewers to intuitively understand the dynamic operation of hidden digital effects, but also provides a precise rectification path to avoid regulatory penalties without taking down the entire video. It achieves deep integration of technology and business and flexible risk control.

[0087] Please see Figure 6 The image recognition-based dynamic detection system for advertisements is applied to the aforementioned image recognition-based dynamic detection method for advertisements; the image recognition-based dynamic detection system for advertisements includes: The advertising feature map module 21 is used to collect the advertisement to be detected, perform multi-scale segmentation on the advertisement to be detected, extract visual appearance features and optical flow motion features respectively, extract audio spectrum features in the advertisement to be detected simultaneously, and generate the corresponding advertising feature map by combining visual appearance features and optical flow motion features. Image recognition module 22 is used to determine the feature change gradient matrix based on image recognition of advertising feature maps; and to perform time-series tracking of the feature change gradient matrix to extract the gradual trajectory features representing the illegal content from concealment to manifestation and the strobe trajectory features representing the periodic flashing of illegal elements. The knowledge graph module 23 is used to perform feature fusion on the gradual trajectory features and the flashing trajectory features to generate the corresponding advertising feature vector; the advertising feature vector is mapped to a graph query node and input into the violation knowledge graph to determine the topological distance between the graph query node and each violation semantic entity node in the violation knowledge graph; The dynamic detection module 24 is used to obtain the placement scenario of the advertisement to be detected, and determine the corresponding violation frame sequence by combining the corresponding traffic warning content and the topological distance. Based on the violation frame sequence, it determines multiple sub-violation contents and determines the corresponding dynamic detection report by combining the advertising constraint factors of the advertisement to be detected.

[0088] It should be noted that although multiple modules are mentioned in the detailed description above, this division is not mandatory; in fact, according to the embodiments of this disclosure, the features and functions of two or more modules or described above can be embodied in one module; conversely, the features and functions of one module described above can be further divided into multiple modules to be embodied.

[0089] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein; this application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein; the specification and embodiments are to be considered exemplary only.

[0090] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A dynamic detection method for advertisements based on image recognition, characterized in that, include: The system collects advertisements to be detected, performs multi-scale segmentation on the advertisements to be detected, extracts visual appearance features and optical flow motion features respectively, extracts audio spectrum features from the advertisements to be detected simultaneously, and generates corresponding advertisement feature maps by combining visual appearance features and optical flow motion features. Determining the feature change gradient matrix based on image recognition of advertising feature maps; The gradient matrix of feature changes is tracked over time to extract the gradual trajectory features that represent the violation content from concealment to manifestation, as well as the stroboscopic trajectory features that represent the periodic flashing of the violation element. The gradual trajectory features and the flashing trajectory features are fused to generate the corresponding advertising feature vector; The advertising feature vector is mapped to a graph query node and input into the violation knowledge graph to determine the topological distance between the graph query node and each violation semantic entity node in the violation knowledge graph. The system obtains the placement scenario of the ad to be detected, and determines the corresponding violation frame sequence by combining the corresponding traffic warning content and the topological distance. Based on the violation frame sequence, it determines multiple sub-violation contents and determines the corresponding dynamic detection report by combining the ad constraint factors of the ad to be detected.

2. The method for dynamic detection of advertisements based on image recognition according to claim 1, characterized in that, The process involves collecting the advertisement to be detected, segmenting the advertisement into multiple scales, extracting visual appearance features and optical flow motion features, simultaneously extracting audio spectral features from the advertisement, and combining the visual appearance features and optical flow motion features to generate a corresponding advertisement feature map, including: The system collects ads to be detected by traversing the ad database and combines convolutional networks to trigger multi-scale segmentation of the ads to be detected. In the process of multi-scale segmentation, visual appearance features are extracted and optical flow motion features are extracted simultaneously by convolutional groups. At this time, visual appearance features represent spatial texture and color distribution; optical flow motion features represent the displacement field and motion edge of adjacent frame pixels.

3. The method for dynamic detection of advertisements based on image recognition according to claim 2, characterized in that, The process of collecting the advertisement to be detected involves multi-scale segmentation of the advertisement, extracting visual appearance features and optical flow motion features separately, simultaneously extracting audio spectrum features from the advertisement, and combining the visual appearance features and optical flow motion features to generate a corresponding advertisement feature map. It also includes: The audio stream of the advertisement to be detected is labeled, the audio spectrum features are determined based on the analysis of the audio stream, and the visual appearance features and optical flow motion features are combined for feature alignment and bidirectional fusion to output a fused feature group. The fused feature group is spliced ​​along the channel dimension and compressed by a 1x1 convolution kernel to generate an advertisement feature map that represents the spatiotemporal joint representation of multiple modalities.

4. The method for dynamic detection of advertisements based on image recognition according to claim 1, characterized in that, The feature change gradient matrix is ​​determined by image recognition based on advertising feature maps; Temporal tracking of the feature change gradient matrix is ​​performed to extract the gradual trajectory features representing the violation content from concealment to manifestation, as well as the stroboscopic trajectory features representing the periodic flashing of the violation element, including: Image recognition is performed on the advertising feature map, and the absolute difference between the feature response values ​​of adjacent time frames is used to determine the feature change gradient matrix that represents the intensity and direction of spatiotemporal features.

5. The method for dynamic detection of advertisements based on image recognition according to claim 4, characterized in that, The feature change gradient matrix is ​​determined by image recognition based on advertising feature maps; The gradient matrix of feature changes is tracked over time to extract the gradual trajectory features representing the transition of illegal content from concealment to manifestation, as well as the stroboscopic trajectory features representing the periodic flashing of illegal elements. This also includes: The gradient matrix of feature changes is input into a long short-term memory network, and time-series tracking is performed based on time steps to extract the gradual trajectory feature of the feature response value continuously increasing over time. This gradual trajectory specifically presents the trajectory of the illegal content from being hidden to being revealed. The temporal fluctuation signal of the gradient matrix of the marked feature changes is combined with the Fourier transform mechanism to trigger the frequency domain transformation of the temporal fluctuation signal, thereby extracting the stroboscopic trajectory feature with a set periodic high-frequency peak in the frequency domain. This stroboscopic trajectory feature presents the trajectory of the periodic flashing of the illegal element.

6. The method for dynamic detection of advertisements based on image recognition according to claim 1, characterized in that, The feature fusion of the gradual trajectory feature and the strobe trajectory feature is performed to generate the corresponding advertising feature vector; The advertising feature vectors are mapped to graph query nodes and input into the violation knowledge graph to determine the topological distance between the graph query nodes and each violation semantic entity node in the violation knowledge graph, including: The feature fusion based on bilinear pooling is performed on the gradual trajectory features and the stroboscopic trajectory features to enhance the nonlinear interaction response between heterogeneous trajectories, thereby generating the corresponding advertising feature vector. The advertising feature vector is then mapped to a graph query node through a linear mapping layer.

7. The method for dynamic detection of advertisements based on image recognition according to claim 6, characterized in that, The feature fusion of the gradual trajectory feature and the strobe trajectory feature is performed to generate the corresponding advertising feature vector; The advertising feature vectors are mapped to graph query nodes and input into the violation knowledge graph to determine the topological distance between the graph query nodes and each violation semantic entity node in the violation knowledge graph. This also includes: The graph query node is input into the violation knowledge graph containing violation semantic entity nodes and entity association edges. Message passing and aggregation calculations are performed based on the graph neural network to determine the topological distance between the graph query node and each violation semantic entity node in the violation knowledge graph in the low-dimensional embedding space. The topological distance is negatively correlated with the semantic association tightness between the graph query node and the violation semantic entity node, that is, the smaller the topological distance, the higher the semantic association tightness.

8. The method for dynamic detection of advertisements based on image recognition according to claim 1, characterized in that, The process involves obtaining the placement scenario of the advertisement to be detected, determining the corresponding violation frame sequence based on the corresponding traffic warning content and the topological distance, identifying multiple sub-violation contents based on the violation frame sequence, and determining the corresponding dynamic detection report based on the advertising constraint factors of the advertisement to be detected, including: The system obtains the placement scenario of the advertisement to be detected, determines the corresponding scenario tag based on the identification of the placement scenario, retrieves the traffic warning content corresponding to the scenario tag, and applies a weighted penalty to the topological distance based on the danger level weight of the traffic warning content, thereby obtaining a dynamic violation confidence score.

9. The method for dynamic detection of advertisements based on image recognition according to claim 8, characterized in that, The process of obtaining the placement scenario of the advertisement to be detected, determining the corresponding violation frame sequence by combining the corresponding traffic warning content and the topological distance, determining multiple sub-violation contents based on the violation frame sequence, and determining the corresponding dynamic detection report by combining the advertising constraint factors of the advertisement to be detected, also includes: The system filters time-series frames whose dynamic violation confidence exceeds a preset threshold, constructs a violation frame sequence from these frames, segments the violation elements within the sequence based on their spatiotemporal continuity, identifies multiple corresponding sub-violations, and combines these with advertising constraints such as regulatory requirements and audience limitations in the current advertising scenario to match each sub-violation with a corresponding violation risk level and penalty avoidance strategy. This generates a dynamic detection report, which includes a violation timestamp, violation type semantics, and risk avoidance suggestions.

10. A dynamic detection system for advertisements based on image recognition, characterized in that, The image recognition-based dynamic detection system for advertisements is applied to the image recognition-based dynamic detection method for advertisements as described in any one of claims 1-9; the image recognition-based dynamic detection system for advertisements includes: The advertising feature map module is used to collect advertisements to be detected, perform multi-scale segmentation on the advertisements to be detected, extract visual appearance features and optical flow motion features respectively, extract audio spectrum features in the advertisements to be detected simultaneously, and generate corresponding advertising feature maps by combining visual appearance features and optical flow motion features. The image recognition module is used to determine the feature change gradient matrix based on the image recognition of the advertising feature map; and to perform time-series tracking of the feature change gradient matrix to extract the gradual trajectory features representing the illegal content from concealment to manifestation and the strobe trajectory features representing the periodic flashing of illegal elements. The knowledge graph module is used to fuse the features of gradual trajectory and strobe trajectory to generate the corresponding advertising feature vector; the advertising feature vector is mapped to a graph query node and input into the violation knowledge graph to determine the topological distance between the graph query node and each violation semantic entity node in the violation knowledge graph; The dynamic detection module is used to obtain the placement scenario of the advertisement to be detected, and determine the corresponding violation frame sequence by combining the corresponding traffic warning content and the topological distance. Based on the violation frame sequence, it determines multiple sub-violation content, and determines the corresponding dynamic detection report by combining the advertising constraint factors of the advertisement to be detected.

Images