Evidence obtaining method and system based on image processing

Abstract

The invention provides an evidence obtaining method and system based on image processing, and the method comprises the following steps: S1, generating a pixel-level depth-of-field distribution diagram of an input image through a multi-scale encoder-decoder network, employing an edge perception optimization layer in a decoding stage, and improving the depth-of-field boundary precision through minimizing a local gradient consistency loss function; s2, performing depth-of-field rationality verification based on an optical imaging physical model, and triggering a first-level tampering alarm by calculating a defocusing fuzzy radius and a gradient direction of a selected region when a difference between the defocusing gradient directions of a target region and a background region exceeds a preset threshold value; and S3, dynamically positioning a key pixel region, identifying a depth-of-field mutation boundary by using an edge detector, calculating by combining local texture complexity, screening a pixel set of which the entropy value is higher than a threshold value and which is located at the mutation boundary, correlating metadata to verify the rationality of the physical size and the spatial position of an object, and eliminating false detection caused by perspective transformation.

Description

Technical Field

[0001] The present invention belongs to the field of image evidence collection, and in particular relates to an image processing-based evidence collection method and system. Background Art

[0002] At present, digital forensics methods based on image processing mainly use the inherent patterns, statistical inconsistencies or operation traces left in the process of image generation, editing, storage and transmission to identify authenticity, identify tampering or trace the source.

[0003] Its core technical background covers multiple directions: First, it uses camera imaging characteristics, such as sensor pattern noise, lens optical distortion, color filter array interpolation traces, and original image processing pipeline characteristics as device fingerprints; second, it analyzes the image file format structure, especially the quantization table, blocking effect, and specific statistical anomalies introduced by double compression in the JPEG compression process; third, it detects traces left by image editing operations, such as similarity features introduced by the copied area in copy-move forgery, edge discontinuities or lighting / color inconsistencies caused by splicing operations, and periodic interpolation traces generated by resampling operations; fourth, it uses the statistical laws of the image content itself, such as the specific noise distribution, lighting direction consistency, and physical reflection characteristics of natural images. When these laws are destroyed by tampering, they can be detected.

[0004] In addition, with the rise of deep learning, methods based on models such as convolutional neural networks that automatically learn the intrinsic features of images to distinguish between real and tampered areas have also developed rapidly, identifying more subtle tampering patterns by training a large number of samples. However, despite the continuous evolution of technology, current forensic methods based on image processing still face a series of significant shortcomings and fundamental challenges. Existing technologies are highly dependent on the presence of specific detectable traces in images, while many advanced tampering techniques, especially synthetic content produced using generative adversarial networks or deep fakes, can effectively simulate the statistical characteristics of real images, greatly reducing the abnormal traces that traditional forensic methods can rely on, resulting in detection failure. Most methods are extremely sensitive to conventional post-processing operations such as image compression, scaling, and format conversion. These common operations can seriously damage or cover up the subtle traces required for forensics, resulting in a large number of false negative results. In particular, when images are transmitted multiple times through social media platforms, the forensic features contained in them are often irreversibly attenuated or erased.

[0005] At the same time, existing algorithms generally suffer from insufficient generalization. Models trained for specific datasets or manipulation types often experience significant performance degradation when faced with images from unknown sources, new manipulation techniques, or images captured by different camera models, making them difficult to adapt to rapidly evolving forgery techniques. Many methods based on pattern noise or device fingerprints require comparison with reference images captured by the same camera, which is unavailable in most practical forensic scenarios, significantly limiting their practical application value.

[0006] In addition, complex forensic algorithms are usually computationally expensive and difficult to meet the needs of real-time detection or rapid screening of large-scale content platforms. In terms of detection accuracy, existing methods have limited ability to distinguish between malicious tampering and reasonable editing (such as simple brightness adjustment and cropping), are prone to confusion, and generally face the dilemma of balancing robustness and sensitivity: increasing the detection rate of slight tampering often leads to a sharp increase in false positives, while reducing false positives makes the system insensitive to fine tampering.

[0007] Classical methods such as copy-move detection have a significant decline in matching accuracy when faced with tampered regions that have undergone geometric transformations (rotation, scaling), noise addition, blur processing, or complex backgrounds. More importantly, forensic technology itself is facing serious challenges from anti-forensic means, attackers can actively use tools to eliminate or fake forensic features, misleading the detection system to produce false results. Finally, the field lacks a unified, large-scale, and realistic benchmark dataset and objective and fair evaluation criteria, making it difficult to compare the performance of different methods, hindering the effective iteration and reliable deployment of technology.

[0008] In summary, current image forensic technology has systematic deficiencies in handling high-quality forgeries, resisting routine post-processing, generalization ability, practicality, robustness against anti-forensics, and standardized evaluation, and breakthroughs are needed in basic theory and engineering implementation to address the increasingly severe digital content credibility crisis. SUMMARY

[0009] The present application proposes a forensic method and system based on image processing, which solves the three major problems of high false positive rate, failure to detect high-quality forgeries (such as generative adversarial network synthetic content), and inaccurate positioning of key tampered regions in complex scenarios caused by existing image forensic technology ignoring the physical laws of optical imaging, and realizes the deep integration of physical interpretability and artificial intelligence detection capability.

[0010] The technical solution of the present application is as follows: a forensic method based on image processing, the method comprising the following steps: S1: generating a pixel-level depth-of-field distribution map of the input image through a multi-scale encoder-decoder network, using an edge-aware optimization layer in the decoding stage to improve the accuracy of the depth-of-field boundary by minimizing the local gradient consistency loss function; S2: verifying the depth-of-field reasonableness based on the optical imaging physical model, calculating the defocus blur radius and gradient direction of the selected region, and triggering a first-level tampering alarm when the defocus gradient direction difference between the target region and the background region exceeds the preset threshold; S3: Dynamically locates key pixel areas, uses edge detectors to identify depth-of-field mutation boundaries, and combines local texture complexity calculations to screen pixel sets with entropy values ​​above a threshold and located at the mutation boundary. It then associates metadata to verify the rationality of the object's physical size and spatial position, eliminating false detections caused by perspective transformations. S4: Performs feature fusion analysis through two channels. Channel 1 uses guided filtering to separate high-frequency residual components and extract sensor pattern noise features. Channel 2 uses a graph convolutional network to model object occlusion relationships and shadow direction consistency. It uses an attention mechanism to weightedly fuse the dual-channel features and output a region tampering suspicion score. S5: When the first-level tampering alarm is activated, the tampering confirmation is directly output; otherwise, the second-level verification is started: the tampering type is determined through multi-directional sub-band energy distribution anomaly detection combined with the SVM classifier, and a tampering probability heat map and verification report are generated.

[0011] Existing technologies suffer from four fundamental flaws: First, traditional forensic methods rely on statistical features or single-mode noise analysis, and are unable to model the optical physical constraints of lenses (such as the spatial continuity relationship between defocus blur gradients and depth of field). This leads to a surge in the missed detection rate of tampering that conforms to statistical distribution but violates physical laws (such as the abnormal out-of-focus effect of artificially inserted objects); Second, although mainstream deep learning forensics solutions can learn hidden traces, they are essentially black-box models and lack an explainable physical rule verification mechanism. Their generalization ability drops sharply when faced with anti-forensic operations (such as noise injection and heavy compression); Third, key area positioning relies on simple edge detection or uniform blocking strategies. In scenes with complex textures, interwoven shadows, or perspective transformations, it is impossible to accurately separate tampered areas from normal content, resulting in feature extraction contamination; Fourth, physical rules and data-driven features fail to effectively coordinate in the cascade decision-making process. Existing solutions either only use post-fusion (such as feature splicing), resulting in computational redundancy, or ignore the priority of physical verification, making it difficult to balance detection sensitivity and false alarm rate.

[0012] The scheme breaks through the following technical difficulties: for the first time, a depth of field-defocus gradient direction coupling verification model is established (step S2), the defocus gradient direction difference between the target and the background is quantified (instead of relying only on the absolute value of the blur radius), the consistency contradiction of the focal plane in optical imaging is captured, and the difficulty of insufficient physical law modeling is overcome; a dynamic key area positioning-metadata collaborative filtering mechanism is designed (step S3), candidate areas are screened in combination with edge mutations and local texture entropy, and the rationality of the spatial position of the object is deduced by using the lens parameters in EXIF, thereby breaking through the bottleneck of positioning blur in complex scenes; an innovative dual-channel feature physical-semantic separation extraction architecture is designed (step S4), while channel one retains the underlying physical features such as device fingerprints, channel two analyzes high-level semantic conflicts (such as shadow direction contradictions) through a graph convolution network, thereby solving the problem that a single feature is easily attacked; a physical rule priority cascade decision flow is constructed (step S5), when the physical verification triggers an alarm, the calculation-intensive feature analysis is skipped, thereby significantly improving the real-time performance and robustness of the system, and realizing the joint optimization of detection efficiency and accuracy.

[0013] As a preferred embodiment, the edge-aware optimization layer in step S1 performs a multi-scale feature pyramid fusion operation, specifically including deploying a feature calibration unit at the output end of each stage of the encoder to capture depth context information of different receptive fields through a hollow spatial pyramid pooling; injecting an edge guide signal in the upsampling process on the decoder, which is derived from a Canny edge intensity map extracted by a high-frequency component separation module; using an adaptive gating mechanism to fuse shallow texture features and deep semantic features to generate a boundary sharpened depth distribution map; the local gradient consistency loss function includes a direction constraint term and an amplitude constraint term, wherein the direction constraint term forces the gradient vector field of adjacent pixels to satisfy the vorticity conservation condition, and the amplitude constraint term suppresses gradient mutations caused by noise through an anisotropic diffusion equation.

[0014] As a preferred embodiment, the optical imaging physical model verification of step S2 constructs a mapping relationship between object distance and defocus blur based on the lens imaging equation, and verifies the perspective consistency by analyzing the blur radius variation rate of parallel plane objects in the image; a local consistency measure is used to quantify the defocus gradient direction difference, including calculating the gradient vector field cosine similarity matrix of the target area and the background area, and triggering an alarm when the average of the elements outside the main diagonal of the matrix exceeds the preset tolerance; ray tracing simulation is performed on objects outside the focal plane to detect artificial inserted objects by comparing the deviation degree of the actual imaging spherical aberration and the simulation result; the physical model verification is performed on the luminance component of the YUV color space to avoid interference of chroma sampling.

[0015] As a preferred embodiment, the key pixel area positioning in step S3 is achieved through a multimodal feature collaborative analysis framework, and the depth of field mutation boundary detection adopts a phase consistency edge detector instead of a traditional operator, which enhances noise resistance through local phase spectrum analysis; the texture complexity analysis introduces a direction-controllable Gabor filter group, and extracts the multi-directional frequency response energy entropy as a screening basis; the metadata verification includes the reverse derivation of the lens optical distortion correction, and verifies the rationality of the object's spatial position through the reprojection error; a perspective geometry constraint engine is established, and the vanishing point detection and horizon calibration technology are used to exclude legitimate perspective transformation areas; the collaborative analysis framework outputs a mask area with a confidence score, and automatically merges suspected areas that are spatially topologically connected.

[0016] As a preferred embodiment, in the dual-channel feature fusion analysis of step S4: the sensor pattern noise extraction of channel one adopts the non-local mean denoising residual method to separate the device-related noise components under the guidance of the depth of field distribution map; the scene semantic relationship modeling of channel two generates an object-level semantic map based on the instance segmentation network, constructs a three-dimensional spatial relationship reasoning engine, and identifies physical contradictions through shadow direction consistency detection and light reflection analysis; the attention weighting mechanism implements cross-channel feature significance evaluation, specifically dynamically allocates channel weights through a differentiable feature selector; the multi-directional sub-band energy distribution anomaly detection includes: constructing a mutual information entropy measure between directional sub-bands to quantify the frequency domain correlation attenuation anomaly; and using a hierarchical anomaly response feature extractor to capture the frequency domain representation of forgery traces at multiple decomposition scales.

[0017] An image processing-based forensics system includes a depth estimation module comprising a multi-scale encoder-decoder network, wherein the decoder is embedded with an edge-aware optimization layer for outputting a high-precision depth distribution map. The optimization layer is trained using a local gradient consistency loss function. Physical rule verification module: configured to perform focal length-object distance-depth of field mapping calculations, compare the defocus blur gradient direction differences between the target area and the background area, and generate a level 1 tampering alarm signal when the difference exceeds a threshold; Key area positioning engine: including: edge detection unit to identify depth of field mutation boundaries, texture analysis unit to calculate local LBP entropy value, EXIF ​​verification unit to associate lens parameters with physical space logic and screen key pixel areas; The dual-channel feature analysis module consists of a first channel and a second channel. The first channel includes a guided filter and a pattern noise extractor, while the second channel uses a graph convolutional network-driven scene semantic relationship analyzer and an attention weighted fusion to output a tampering suspicion score. The first-level alarm signal of the physical rule verification module directly outputs the tampering result; otherwise, the second-level verification unit is activated, which includes: a Steerable Pyramid multi-directional sub-band energy analyzer, a pre-trained SVM classifier and a thermal map generator; the preprocessing interface is physically connected to the image sensor to obtain RAW format data and transmit it to the depth of field estimation module; the output priority of the physical rule verification module is higher than that of the dual-channel feature analysis module.

[0018] After adopting the above technical solution, the present invention has the following beneficial effects: at the level of detection accuracy, through the collaboration of depth of field physical constraint verification (step S2) and key area dynamic positioning (step S3), tampering that violates the laws of optical imaging (such as the mismatch between the defocus gradient direction of the forged object and the background) is accurately identified with near-zero false detection. Compared with traditional methods, experiments on the COVERAGE dataset show that the false alarm rate is reduced; at the level of anti-interference capability, the dual-channel feature fusion mechanism (step S4) separates and extracts physical fingerprints and semantic features, maintaining a high recall rate even in the face of forged content generated by a generative adversarial network or heavy JPEG compression attacks. This is because channel one retains strong robustness features such as sensor noise, and channel two captures semantic contradictions through occlusion relationship analysis. The dual protection increases the probability of failure of anti-forensic attacks. In terms of computational efficiency, the cascaded decision-making process (step S5) prioritizes lightweight physical rule verification, initiating deep feature analysis only for difficult cases that haven't triggered alarms. This reduces the system's average response time to one-fifth that of a pure end-to-end deep learning solution, meeting the demands of real-time forensics. Furthermore, key technological breakthroughs offer derivative advantages: dynamic key area positioning reduces ineffective feature calculations, significantly optimizing resource consumption; the edge-aware optimization layer (step S1) improves depth-of-field boundary accuracy to sub-pixel levels, making even the smallest tampering (such as hair-thin stitching seams) impossible to hide; and EXIF ​​metadata verification (step S3) establishes a closed-loop digital-physical space mapping, providing a traceable physical and logical chain of evidence for forensic evidence collection. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0022] Example: like Figure 1 As shown, a forensics method based on image processing comprises the following steps: S1: Generates a pixel-level depth map of the input image through a multi-scale encoder-decoder network. In the decoding stage, an edge-aware optimization layer is used to improve the accuracy of the depth boundary by minimizing the local gradient consistency loss function. S2: Depth of field rationality verification is performed based on the optical imaging physics model. By calculating the defocus blur radius and gradient direction of the selected area, a level 1 tampering alarm is triggered when the difference in the defocus gradient direction between the target area and the background area exceeds a preset threshold. S3: Dynamically locates key pixel areas, uses edge detectors to identify depth-of-field mutation boundaries, and combines local texture complexity calculations to screen pixel sets with entropy values ​​above a threshold and located at the mutation boundary. It then associates metadata to verify the rationality of the object's physical size and spatial position, eliminating false detections caused by perspective transformations. S4: Performs feature fusion analysis through two channels. Channel 1 uses guided filtering to separate high-frequency residual components and extract sensor pattern noise features. Channel 2 uses a graph convolutional network to model object occlusion relationships and shadow direction consistency. It uses an attention mechanism to weightedly fuse the dual-channel features and output a region tampering suspicion score. S5: When the first-level tampering alarm is activated, the tampering confirmation is directly output; otherwise, the second-level verification is started: the tampering type is determined through multi-directional sub-band energy distribution anomaly detection combined with the SVM classifier, and a tampering probability heat map and verification report are generated.

[0023] In this application document, the input image is passed through a multi-scale encoder-decoder network to generate an initial depth distribution map. The encoder extracts multi-level features through convolutional layers, and the decoder uses upsampling and skip connections to restore spatial details. The edge-aware optimization layer introduces high-frequency edge guidance signals and uses a local gradient consistency loss function to constrain the gradient direction continuity of adjacent pixels, eliminating the boundary blurring effect in traditional depth estimation. This process simulates the binocular fusion mechanism of the human visual system and achieves sub-pixel depth reconstruction by strengthening the coupling relationship between object contours and depth mutations, overcoming the geometric ambiguity problem of monocular depth prediction.

[0024] Based on the principle of lens imaging, a defocus blur gradient analysis model is constructed, and the defocus blur gradient vector field is calculated in the selected target area and adjacent background area. By quantifying the cosine similarity of the gradient direction of the two areas, the focal plane consistency violation phenomenon is detected. When the direction difference exceeds the allowed range of optical aberration, the first level of tampering alarm is triggered. This step first converts the spherical and coma characteristics of the optical system into quantifiable forensic indicators, breaking the limitations of traditional methods relying on absolute blur values, and has specific recognition ability for abnormal defocus effects of forged objects.

[0025] The phase consistency edge detection technology is used to identify the depth of field mutation boundary, and the local texture entropy value analysis is used to screen high complexity candidate areas. The optical parameters and spatial information in the image metadata are associated, and the physical size rationality of the object is deduced by perspective projection geometry, and the legal perspective transformation area is automatically excluded. This mechanism establishes a mapping verification closed loop from pixel space to physical space, solves the confusion problem of tampering areas and natural depth of field changes in complex scenes, and significantly reduces the false detection rate caused by building perspective or natural landscape.

[0026] Channel one uses a depth-guided directional filter to separate high-frequency residual components, extracts device-related pattern noise features, and preserves the physical intrinsic properties of sensor fingerprints. Channel two constructs a scene semantic graph model based on graph convolution network, analyzes the three-dimensional spatial consistency of occlusion relationship and shadow projection direction between objects. The dual-channel features are fused by attention weighting mechanism to generate regional tampering suspiciousness score. This architecture complements the orthogonal features of the physical layer and the semantic layer, and resists the attack of generated adversarial network fake content on a single feature dimension.

[0027] The system preferentially performs optical physics rule verification, and if the first level alarm is triggered, it directly outputs the tampering confirmation. The difficult cases that do not trigger the alarm enter the second verification: the image frequency domain features are decomposed by multi-directional sub-band filter bank to detect the frequency domain correlation attenuation caused by resampling or copy-move operation; combined with support vector machine classifier to determine the tampering type. Finally, a spatial positioning heat map and verification report are generated. This cascading mechanism realizes the dialectical unity of physically interpretable verification and data-driven analysis, while ensuring the reliability of forensic evidence and optimizing the allocation of computing resources.

[0028] In routine image forensics, after technicians import the image to be examined, the system automatically performs multi-scale feature pyramid fusion. Feature calibration units at each encoder stage use atrous spatial pyramid pooling (ASPP) to capture the depth context of the scene's near, mid, and far views, effectively resolving depth prediction ambiguities caused by a single receptive field. The decoder upsampling process injects the Canny edge intensity map extracted by the high-frequency component separation module as a guidance signal in real time, ensuring precise alignment of object contours with depth transition boundaries. An adaptive gating mechanism dynamically adjusts the fusion weights of shallow texture features (such as surface details) and deep semantic features (such as overall scene structure) to generate a sharpened depth distribution map. The direction constraint of the local gradient consistency loss function enforces curl conservation for adjacent pixel gradient vectors (a physical law governing continuous depth variation in natural scenes). The amplitude constraint, based on the anisotropic diffusion equation, suppresses noise interference, ensuring sub-pixel depth map accuracy even in low-light or highly compressed images. The system automatically outputs the optimized depth distribution map for subsequent modules to use without requiring manual operator intervention.

[0029] In criminal evidence collection or news image review scenarios, the physical model verification module automatically selects parallel planar objects in the image (such as floor tiles and building facades) and calculates the rate of change of the theoretical blur radius based on the lens imaging equation. Any nonlinear changes detected indicate perspective distortion. A local consistency measure identifies focal plane conflicts (such as the deviation in the out-of-focus direction between a forged object and the background) by calculating the cosine similarity matrix of the gradient vector fields of the target and background regions. When the mean of the off-diagonal elements of the matrix exceeds the tolerance allowed by the optical system, an alarm is automatically triggered and a report of the violating coordinates is generated. For suspicious objects outside the focal plane (such as suspended evidence), the system performs ray tracing simulations: reconstructing the ideal optical path based on lens parameters, comparing the deviation between the actual image's spherical aberration (such as comet-like tails) and the simulated results, identifying artificially inserted objects. This entire process is performed on the YUV luminance component, avoiding color aliasing caused by chroma subsampling. This alarm allows reviewers to quickly locate violating areas without requiring specialized optical expertise.

[0030] To detect manipulated images on social media, a phase congruency edge detector replaces the traditional Canny operator. This algorithm uses local phase spectrum analysis to identify true depth boundaries (e.g., hair-thin seams), maintaining a detection rate exceeding 90% despite JPEG compression artifacts. Texture complexity analysis uses a Gabor filter bank to scan the frequency response in eight directions, calculating energy entropy to filter out high-complexity areas (e.g., unnatural textures in forged areas) and avoid false positives caused by uniform textures. A metadata verification engine analyzes focal length and sensor parameters in EXIF ​​data and calculates the theoretical projected position of objects through inverse optical distortion correction. If the reprojection error exceeds the sensor pixel tolerance, a spatial inconsistency is identified. A perspective geometry constraint engine automatically detects vanishing points and horizon lines, eliminating areas of legitimate perspective distortion in architectural photography. The system outputs a binary mask with a confidence score and merges suspected areas that are spatially topologically connected (e.g., continuously manipulated objects). Operators can directly view these highlighted areas in a visual interface.

[0031] During forensic electronic evidence evaluation, channel one uses a non-local means denoising residual method, guided by a depth map, to separate device-related noise (such as the fixed pattern noise of the Sony IMX586 sensor) to prevent natural textures from being misidentified as forgeries. Channel two generates an object-level semantic map based on an instance segmentation network (such as Mask R-CNN). A 3D spatial relationship inference engine automatically verifies the consistency of shadow direction with light source position (e.g., illegal back projection) and the matching of specular reflection angle with object material (e.g., unnatural metallic reflections). A differentiable feature selector dynamically assigns weights to both channels based on image attributes: channel one is given a higher weight for images with obvious device fingerprints (e.g., RAW images); channel two is given a higher weight for scenes with prominent semantic contradictions (e.g., suspended objects that violate the laws of gravity). Multi-directional subband analysis uses a Steerable Pyramid decomposition into 16 directional subbands to calculate anomalies in the mutual information entropy decay between subbands (e.g., frequency domain correlation breaks caused by copy-move operations). A hierarchical feature extractor captures frequency domain representations of forgery artifacts at four scales (e.g., frequency holes synthesized by generative adversarial networks). Appraisers can combine thermal maps and physical contradiction reports to generate legally binding evidence conclusions.

[0032] An image processing-based forensics system includes a depth estimation module comprising a multi-scale encoder-decoder network, wherein the decoder is embedded with an edge-aware optimization layer for outputting a high-precision depth distribution map. The optimization layer is trained using a local gradient consistency loss function. Physical rule verification module: This module is configured to perform focal length-object distance-depth of field mapping calculations, compare the defocus gradient direction differences between the target area and the background area, and generate a level 1 tampering warning signal when the difference exceeds a threshold; Key area positioning engine: including: edge detection unit to identify depth of field mutation boundaries, texture analysis unit to calculate local LBP entropy value, EXIF ​​verification unit to associate lens parameters with physical space logic and screen key pixel areas; The dual-channel feature analysis module consists of a first channel and a second channel. The first channel includes a guided filter and a pattern noise extractor, while the second channel uses a graph convolutional network-driven scene semantic relationship analyzer and an attention weighted fusion to output a tampering suspicion score. The first-level alarm signal of the physical rule verification module directly outputs the tampering result; otherwise, the second-level verification unit is activated, which includes: a Steerable Pyramid multi-directional sub-band energy analyzer, a pre-trained SVM classifier and a thermal map generator; the preprocessing interface is physically connected to the image sensor to obtain RAW format data and transmit it to the depth of field estimation module; the output priority of the physical rule verification module is higher than that of the dual-channel feature analysis module.

[0033] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A forensics method based on image processing, characterized in that: The method comprises the following steps: S1: Generates a pixel-level depth map of the input image through a multi-scale encoder-decoder network. In the decoding stage, an edge-aware optimization layer is used to improve the accuracy of the depth boundary by minimizing the local gradient consistency loss function. S2: Depth of field rationality verification is performed based on the optical imaging physics model. By calculating the defocus blur radius and gradient direction of the selected area, a level 1 tampering alarm is triggered when the difference in the defocus gradient direction between the target area and the background area exceeds a preset threshold. S3: Dynamically locates key pixel areas, uses edge detectors to identify depth-of-field mutation boundaries, and combines local texture complexity calculations to screen pixel sets with entropy values ​​above a threshold and located at the mutation boundary. It then associates metadata to verify the rationality of the object's physical size and spatial position, eliminating false detections caused by perspective transformations. S4: Performs feature fusion analysis through two channels. Channel 1 uses guided filtering to separate high-frequency residual components and extract sensor pattern noise features. Channel 2 uses a graph convolutional network to model object occlusion relationships and shadow direction consistency. It uses an attention mechanism to weightedly fuse the dual-channel features and output a region tampering suspicion score. S5: When the first-level tampering alarm is activated, the tampering confirmation is directly output; otherwise, the second-level verification is started: the tampering type is determined through multi-directional sub-band energy distribution anomaly detection combined with the SVM classifier, and a tampering probability heat map and verification report are generated.

2. The image processing-based evidence collection method according to claim 1, characterized in that: The edge-aware optimization layer in step S1 performs a multi-scale feature pyramid fusion operation, specifically including deploying a feature calibration unit at the output end of each stage of the encoder, capturing depth context information of different receptive fields through void space pyramid pooling; injecting an edge guidance signal during the decoder upsampling process, which is derived from the Canny edge intensity map extracted by the high-frequency component separation module; using an adaptive gating mechanism to fuse shallow texture features and deep semantic features to generate a depth distribution map with sharpened boundaries; the local gradient consistency loss function includes a direction constraint term and an amplitude constraint term, wherein the direction constraint term forces the gradient vector field of adjacent pixels to satisfy the curl conservation condition, and the amplitude constraint term suppresses gradient mutations caused by noise through the anisotropic diffusion equation.

3. The image processing-based evidence collection method according to claim 1, wherein: The optical imaging physical model verification in step S2 constructs a mapping relationship between object distance and defocus blur based on the lens imaging equation, verifies perspective consistency by analyzing the rate of change of blur radius of parallel plane objects in the image; quantifies the difference in defocus gradient direction using a local consistency measure, including calculating the cosine similarity matrix of the gradient vector field of the target area and the background area, and triggers an alarm when the mean value of the elements outside the main diagonal of the matrix exceeds a preset tolerance; Focus Perform ray tracing simulation on out-of-plane objects and detect artificially inserted objects by comparing the deviation between the spherical aberration of the actual imaging and the simulation results; The physical model verification is performed on the luma component of the YUV color space to avoid chroma sampling interference.

4. The image processing-based evidence collection method according to claim 1, wherein: The key pixel area positioning in step S3 is achieved through a multimodal feature collaborative analysis framework. The depth of field mutation boundary detection adopts a phase consistency edge detector instead of a traditional operator, which enhances noise resistance through local phase spectrum analysis. The texture complexity analysis introduces a direction-controllable Gabor filter group and extracts the multi-directional frequency response energy entropy as a screening basis. The metadata verification includes the reverse derivation of the lens optical distortion correction and verifies the rationality of the object's spatial position through the reprojection error. A perspective geometry constraint engine is established, and the vanishing point detection and horizon calibration technology are used to exclude legitimate perspective transformation areas. The collaborative analysis framework outputs a mask area with a confidence score and automatically merges suspected areas that are spatially topologically connected.

5. The image processing-based evidence collection method according to claim 1, characterized in that: In the dual-channel feature fusion analysis of step S4: the sensor pattern noise of channel 1 is extracted using the non-local mean denoising residual method, and the device-related noise component is separated under the guidance of the depth of field distribution map; Channel 2's scene semantic relationship modeling generates object-level semantic graphs based on instance segmentation networks, builds a 3D spatial relationship reasoning engine, and identifies physical contradictions through shadow direction consistency detection and light reflection analysis. The attention weighting mechanism implements cross-channel feature saliency evaluation, specifically by dynamically assigning channel weights through a differentiable feature selector. The multi-directional sub-band energy distribution anomaly detection includes: constructing a mutual information entropy measure between directional sub-bands to quantify frequency domain correlation attenuation anomalies; A hierarchical abnormal response feature extractor is adopted to capture the frequency domain representation of forgery traces at multiple decomposition scales.

6. A forensics system based on image processing, characterized in that: Includes a depth estimation module: It contains a multi-scale encoder-decoder network whose decoder is embedded with an edge-aware optimization layer to output a high-precision depth distribution map. The optimization layer is trained with a local gradient consistency loss function. Physical rule verification module: configured to perform focal length-object distance-depth of field mapping calculations, compare the defocus blur gradient direction differences between the target area and the background area, and generate a level 1 tampering alarm signal when the difference exceeds a threshold; The key area positioning engine uses the edge detection unit to identify the depth of field mutation boundary, the texture analysis unit to calculate the local LBP entropy value, and the EXIF ​​verification unit to associate lens parameters with physical space logic to screen key pixel areas; The dual-channel feature analysis module consists of a first channel and a second channel. The first channel includes a guided filter and a pattern noise extractor, while the second channel uses a graph convolutional network-driven scene semantic relationship analyzer and an attention weighted fusion to output a tampering suspicion score. The first-level alarm signal of the physical rule verification module directly outputs the tampering result; Otherwise, a secondary verification unit is activated, wherein the secondary verification unit includes: a Steerable Pyramid multi-directional sub-band energy analyzer, a pre-trained SVM classifier, and a heat map generator; The preprocessing interface is physically connected to the image sensor and is used to obtain RAW format data and transmit it to the depth of field estimation module; the output priority of the physical rule verification module is higher than that of the dual-channel feature analysis module.

Images