Method and system for intelligent inference of postmortem lividity image death time

Abstract

This invention belongs to the field of forensic image analysis technology and discloses an intelligent method and system for inferring the time of death from forensic livor mortis images. The method includes: acquiring livor mortis images under standardized light sources and performing color space conversion; separating livor mortis regions and extracting color feature vectors through a semantic segmentation network; inferring the preliminary time of death using an association modeling network that includes color analysis branches, distribution feature branches, and finger pressure test branches; obtaining the final estimate through environmental temperature correction and multi-site weighted fusion; and outputting the time of death interval and an identification auxiliary report. This invention achieves objective and quantitative analysis of the evolution of livor mortis color, providing objective, quantitative, and standardized image analysis technology support for forensic time of death identification.

Description

Technical Field

[0001] This invention belongs to the field of forensic image analysis technology, specifically relating to a method and system for intelligent estimation of forensic time of death based on color evolution analysis of livor mortis images. Background Technology

[0002] In forensic medical practice, accurately estimating the time of death is crucial for case investigation and judicial trials. Determining the time elapsed after death provides key temporal clues for case reconstruction, helping to define the suspect's behavioral trajectory, verify witness testimonies, and exclude or identify criminal suspects. However, the estimation of the time of death has long relied on the subjective experience and judgment of forensic personnel, lacking objective and quantitative technical support. The estimation results among different forensic personnel often differ significantly, and this subjectivity, to some extent, affects the scientific validity and credibility of forensic medical conclusions.

[0003] Currently, traditional methods used in forensic medicine to estimate time of death mainly include post-mortem temperature measurement, rigor mortis assessment, corneal opacity observation, and livor mortis observation. Post-mortem temperature measurement estimates time of death by measuring rectal temperature and combining it with a Henssge nomogram; however, this method is significantly affected by factors such as ambient temperature, weight, and clothing, and is only applicable to the early post-mortem stage. While rigor mortis assessment can provide a reference for time of death within a certain timeframe, its accuracy is limited due to multiple factors such as temperature and muscle condition. Corneal opacity observation has attracted attention in recent years, with some literature reporting the use of smartphones combined with image analysis for quantitative assessment of corneal opacity to estimate post-mortem intervals (Zheng et al., Journal of Forensic Sciences, 2020; Cantürk et al., Expert Systems, 2024). However, this method only focuses on local eye features and fails to fully utilize the broader information from the body's surface.

[0004] Lividity, as one of the earliest and longest-lasting post-mortem phenomena, contains rich information about the passage of time after death. Lividity is a characteristic color change that occurs when blood, under the influence of gravity, settles into the blood vessels in the lower parts of the body after blood circulation ceases, and shines through the skin. The formation and evolution of livor mortis typically involves a gradual color change from light red to dark red, purplish-red, and finally dark purple, accompanied by a shift from scattered distribution to gradual merging, and from a variable to a fixed pattern. These evolutionary characteristics are closely related to the passage of time after death, providing an objective material basis for inferring the time of death based on the characteristics of livor mortis.

[0005] Currently, Chinese patent CN120298301A discloses a method for determining the cause of death based on virtual anatomical images. This method acquires CT image data of a corpse, preprocesses it, and then inputs it into a 3D-DenseNet121 three-dimensional convolutional neural network model for feature extraction. A Softmax classifier is then used to map the feature vectors to categories of cause of death such as drowning, sudden death, fall from a height, and mechanical asphyxiation. The LIME algorithm is used to analyze the interpretability of the determination results. Although this method achieves high classification accuracy in determining the cause of death, its technical approach has the following shortcomings: First, the input to this method is a 3D CT image, requiring large medical imaging equipment, making rapid acquisition and analysis at the crime scene impossible. Second, the goal of this method is to determine the category of cause of death rather than to infer the time of death; its classification output cannot provide a quantitative estimate of the time elapsed since death. Third, this method does not involve time-related modeling of the evolution of livor mortis color, and cannot utilize livor mortis, the most time-sensitive physical characteristic in forensic medicine, for time inference. Furthermore, existing technologies for quantitative analysis of livor mortis characteristics are still mainly at the level of qualitative description. A quantitative mapping model from the color characteristics of livor mortis images to the passage of time after death has not yet been established, and there is a lack of systematic solutions that comprehensively consider the response to finger pressure tests, the influence of ambient temperature, and the fusion of livor mortis information from multiple locations.

[0006] Therefore, there is an urgent need for an intelligent analysis method and system that can objectively and quantitatively infer the time of death based on the color evolution characteristics of livor mortis images. This would compensate for the shortcomings of existing technologies in forensic time of death identification, improve the scientificity, consistency and repeatability of identification conclusions, reduce subjective differences among different experts, and provide more reliable technical support for criminal case investigation and judicial trials. Summary of the Invention

[0007] The technical problem to be solved by the present invention is that existing forensic methods for estimating the time of death rely on the subjective experience of the forensic personnel, the evolution characteristics of livor mortis are difficult to quantify objectively, there is a lack of a quantitative mapping model from the color characteristics of livor mortis images to the time elapsed after death, and the methods fail to comprehensively consider the reaction of finger pressure test, the influence of environmental temperature, and the fusion of information from multiple livor mortis sites.

[0008] To address the aforementioned technical problems, this invention provides an intelligent method for estimating the time of death from forensic livor mortis images, comprising the following steps:

[0009] Step S1, Standardized acquisition and preprocessing of livor mortis images: Under the control of a standardized light source and color calibration device, high-resolution images of the livor mortis area of ​​the lower part of the corpse are acquired, and white balance correction and color space conversion are performed on the acquired high-resolution images of the livor mortis area to convert the image from RGB color space to HSV color space to obtain a multi-channel livor mortis feature image containing hue channel, saturation channel and lightness channel.

[0010] Step S2, Lividity Region Segmentation and Color Feature Extraction: Automatic skin region segmentation is performed on the multi-channel livor mortis feature image. The livor mortis region is separated from the non-livor mortis region by pixel-level classification based on a semantic segmentation network. The hue distribution features, saturation distribution features, and brightness distribution features of the segmented livor mortis region are extracted using a color analysis network to construct the livor mortis color feature vector.

[0011] Step S3, Modeling the Correlation Between Lividity Color Evolution and Postmortem Time: The livor mortis color feature vector is input into the correlation modeling network for time inference. The correlation modeling network includes a color analysis branch, a distribution feature branch, and a finger pressure test branch. The color analysis branch quantifies the gradual color change process of livor mortis and establishes a time-color mapping curve. The distribution feature branch evaluates the degree of fusion and diffusion range of livor mortis. The finger pressure test branch determines the fixation of livor mortis by analyzing the differences between images before and after finger pressure. The outputs of the three branches are fused through a fusion layer to obtain a preliminary estimate of the postmortem time.

[0012] Step S4, Environmental Factor Correction and Multi-site Fusion: The environmental temperature correction module is used to perform nonlinear temperature correction on the preliminary estimate of postmortem time based on the ambient temperature of the body. The inference results of multiple livor mortis features on the neck, back and lower limbs are combined through the multi-site fusion branch, and an adaptive weighted fusion strategy is used to obtain the corrected estimate of postmortem time.

[0013] Step S5, Death Time Interval Estimation and Report Generation: Based on the corrected postmortem estimator and its confidence level, calculate the death time interval estimate and generate an inference result report containing livor mortis-annotated images, death time intervals, and auxiliary information of expert opinions.

[0014] This invention also provides an intelligent system for estimating the time of death from forensic livor mortis images, including a standardized acquisition and preprocessing module for livor mortis images, a livor mortis region segmentation and color feature extraction module, a livor mortis color evolution and post-mortem time correlation modeling module, an environmental factor correction and multi-site fusion module, and a death time interval estimation and report generation module. Each of these modules corresponds one-to-one with steps S1 to S5 in the method. The correlation modeling module includes built-in color analysis, distribution feature, and finger pressure test branches. The environmental factor correction and multi-site fusion module integrates a temperature correction sub-module and a multi-site adaptive weighted fusion sub-module.

[0015] The beneficial effects of this invention are as follows: First, by using standardized light sources and color calibration devices, the consistency of image acquisition is ensured, eliminating the interference of different acquisition environments on the analysis of livor mortis color. Second, a quantitative mapping model of livor mortis color evolution and post-mortem time is constructed, realizing objective quantitative analysis of livor mortis color changes and overcoming the subjectivity of traditional visual observation methods. Third, the introduction of a finger pressure test branch, through comparative analysis of images before and after finger pressure, determines the fixation of livor mortis, significantly improving the accuracy of post-mortem time stage division. Fourth, the environmental temperature correction module corrects the inference results based on the nonlinear temperature-time response relationship, effectively reducing the system bias caused by temperature factors. Fifth, the multi-site fusion strategy integrates livor mortis information from the neck, back, and lower limbs, and improves the robustness of inference through an adaptive weighting mechanism. Sixth, the system output includes visualized annotated images and an auxiliary report of the identification opinion, providing objective and traceable technical support for forensic time of death identification. Attached Figure Description

[0016] Figure 1 This is a flowchart of the intelligent inference method for time of death from forensic livor images according to the present invention;

[0017] Figure 2 This is an architecture diagram of the intelligent inference system for time of death from forensic livor images, as described in this invention. Detailed Implementation

[0018] The technical solution of the present invention will be further described in detail below through specific embodiments and in conjunction with the accompanying drawings. It should be noted that the following embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0019] like Figure 1 As shown, the present invention provides an intelligent method for estimating the time of death from forensic livor mortis images, comprising steps S1 to S5. The specific implementation methods for each step are described in detail below.

[0020] Step S1: Standardized Acquisition and Preprocessing of Lividity Images. During forensic scene investigation or autopsy, standardized image acquisition of the livor mortis area on the lower part of the body is required. In one embodiment of this invention, a portable acquisition device integrating a standardized light source module is used for image acquisition. Preferably, the standardized light source adopts a ring-shaped LED array layout, with a color temperature set in the range of 5000K to 6500K, preferably 5500K, to simulate natural sunlight conditions. The illumination uniformity is not less than 90%, ensuring consistent lighting conditions across all areas of the livor mortis surface. The light source power is preferably 15W, and the illumination distance is controlled between 20cm and 35cm, thereby forming a uniform illumination area on the livor mortis surface.

[0021] Before image acquisition, the acquisition system needs to be calibrated using a color calibration device. In one embodiment of the present invention, the color calibration device includes a standard 24-color chart. Before each acquisition task begins, the standard color chart is first photographed, and the system automatically calculates a color correction matrix based on the deviation between the known standard color values ​​in the color chart and the actual photographed color values. Preferably, the color correction matrix The 3×3 affine transformation matrix was obtained by fitting using the least squares method:

[0022] ,

[0023] in: The first one in the color chart The actual RGB values ​​of each color block are collected as a 3×1 vector. For the first Each color block has a standard reference RGB value, which is a 3×1 vector. The total number of color blocks in the color chart, in this embodiment ; This represents the Euclidean norm. This correction matrix effectively eliminates the impact of different camera sensor characteristics and ambient light variations on color reproduction accuracy.

[0024] After calibration, use a digital camera with a resolution of at least 12 megapixels or a dedicated acquisition device to photograph the livor mortis areas on the neck, back, and lower limbs of the cadaver. Preferably, at least two images are acquired for each acquisition site, one of which is a conventionally acquired image, and the other is an image acquired after finger pressure, i.e., after applying a standard finger pressure of approximately 4.9N to the livor mortis surface for 5 seconds and then immediately taking a picture. The preferred image acquisition format is RAW or lossless PNG, and the preferred size of a single image is 4000×3000 pixels.

[0025] After acquisition, white balance correction and color space conversion are performed on all images sequentially. White balance correction utilizes the aforementioned color correction matrix. A linear transformation is performed on the RGB values ​​of each pixel in the image. Then, the white-balance corrected RGB image is converted to the HSV color space. The RGB to HSV conversion follows a standard conversion formula, resulting in a multi-channel livor mortis feature image containing hue (H), saturation (S), and lightness (V) channels. The reason for choosing the HSV color space in this invention is that it can effectively separate color information (hue) from lightness information (lightness), allowing subsequent analysis to focus more on the variation patterns of the livor mortis color itself without being affected by fluctuations in ambient light intensity. Preferably, the hue (H) channel values ​​are normalized to the range of 0 to 360, and the saturation (S) and lightness (V) channel values ​​are normalized to the range of 0 to 1.

[0026] In a preferred embodiment of the present invention, the system further performs adaptive noise reduction processing on the preprocessed image. Considering that the lighting conditions at the forensic scene may not be ideal, and that the acquired image may contain a certain degree of sensor noise, the present invention uses a bilateral filter to reduce noise in the brightness V channel of the HSV image. The spatial domain standard deviation of the filter is set to 5 pixels, and the value domain standard deviation is set to 0.1 (normalized value domain). The advantage of bilateral filtering is that it can smooth noise while maintaining the clarity of the edges of the livor lesions, which is of great significance for accurate segmentation in subsequent steps. In addition, for any locally overexposed or underexposed areas in the image, the CLAHE (Contrast-Limited Adaptive Histogram Equalization) method is used to locally enhance the contrast of the brightness V channel, with a cropping limit set to 3.0 and a block size of 8×8 pixels. These preprocessing operations together ensure the consistency of the input image in terms of color accuracy, noise level, and contrast, laying the foundation for reliable analysis in subsequent steps.

[0027] II. Step S2: Libido Region Segmentation and Color Feature Extraction. After obtaining the multi-channel livor mortis feature image, automatic segmentation of the livor mortis region needs to be performed to accurately separate the livor mortis region from normal skin and background. In one embodiment of the present invention, a semantic segmentation network based on an encoder-decoder structure is used to achieve pixel-level region classification.

[0028] Specifically, the encoder consists of four downsampling stages, each composed of two residual convolutional blocks and one channel attention module. The input to stage 1 is the 3-channel HSV image obtained in step S1, with a 3×3 kernel size and 64 output feature map channels. Stage 2 performs a 2x downsampling using convolutions with a stride of 2, increasing the number of output channels to 128. Stages 3 and 4 expand the number of channels to 256 and 512 respectively, reducing the spatial resolution to 1 / 4 and 1 / 8 of the original image. The channel attention module uses a Squeeze-and-Excitation (SE) structure, adaptively adjusting the weights of each channel through global average pooling and two fully connected layers, achieving a compression ratio of... The value is set to 16. This attention mechanism enhances the response of feature channels related to changes in livor mortis color while suppressing background noise channels.

[0029] The decoder recovers spatial resolution through four upsampling stages. Each stage uses bilinear interpolation for a 2x upsampling and then uses skip connections to concatenate and fuse the feature maps from the encoder and decoder for the corresponding stages. The final output layer uses a 1×1 convolution to map the feature maps to three category channels, corresponding to the livor mortis region, normal skin region, and background region, respectively. A Softmax activation function is then applied to output the probability value of each pixel belonging to each category. Preferably, the segmentation threshold is set to 0.5, meaning pixels with a probability value greater than 0.5 are classified into the corresponding category. In this embodiment, the segmentation network is trained using a weighted combination of the Dice loss function and the cross-entropy loss function, with a weight ratio of 0.6:0.4. The training dataset contains 2000 labeled livor mortis image samples.

[0030] Preferably, the segmentation network is trained using a transfer learning strategy. The encoder part uses a ResNet-34 model pre-trained on the ImageNet dataset for weight initialization to accelerate convergence and improve segmentation accuracy. During training, a phased training strategy is employed: the first phase freezes the encoder weights and trains only the decoder for 20 epochs with a learning rate of [missing information]. The second stage involves unfreezing all parameters and fine-tuning them end-to-end, training for 80 epochs, and reducing the learning rate to [a certain value]. The segmentation accuracy was evaluated on the test set as follows: the Dice coefficient for the livor mortis region reached 0.92, the pixel accuracy reached 0.96, and the IoU (Intersection over Union) reached 0.87. It should be noted that for individuals with darker skin tones, the color contrast between livor mortis and normal skin may be reduced. In this case, the segmentation network can improve its robustness to segmenting such samples by increasing the proportion of dark-skinned samples in the training set.

[0031] After segmentation, extract the mask from the livor mortis area. ,in Indicates position It belongs to the livor mortis area. This represents the non-livor mortis region. Based on this mask, a color analysis network is used to extract color features from the livor mortis region. The color analysis network consists of three parallel statistical feature extraction branches, processing the H, S, and V channels respectively. For each channel within the livor mortis region, the network calculates the following statistical features: channel mean. Channel standard deviation Channel skewness and channel kurtosis Taking the H channel as an example, its mean value is calculated as follows: ,in: For pixels The hue value at that location ranges from 0 to 360. This represents the total number of pixels in the lividity area. The average values ​​of the saturation (S channel) and brightness (V channel) channels. and Calculated in the same way. Standard deviation skewness and kurtosis Calculated according to statistical standards. In addition, the color analysis network also calculates the hue histogram distribution characteristics of the livor mortis region, dividing the hue range of 0 to 360 into 36 equal intervals, statistically analyzing the pixel percentage within each interval, and obtaining a 36-dimensional hue histogram feature vector. .

[0032] Finally, all the above features are concatenated to construct a feature vector of livor mortis color. :

[0033] ,in: The feature vector has 48 dimensions, including 12 dimensions of statistical features (4 for each of the 3 channels) and a 36-dimensional hue histogram feature. This feature vector comprehensively describes the distribution characteristics of the livor mortis region in the three dimensions of hue, saturation, and brightness, providing an informative input representation for subsequent time-related modeling.

[0034] Step S3: Modeling the correlation between the evolution of livor mortis color and postmortem time. This step is the core innovation of this invention. Livor mortis color feature vector. The input is fed into a network that models the relationship between the evolution of livor mortis color and the time elapsed after death. This network adopts a three-branch parallel fusion architecture, which estimates the time elapsed after death from three complementary perspectives: the temporal characteristics of color evolution, the spatial distribution morphology characteristics, and the mechanical response characteristics.

[0035] The core task of color analysis is to establish a non-linear mapping relationship between the evolution of livor mortis color and the passage of time after death. In forensic medicine, the color change of livor mortis exhibits a gradual pattern, from pale red in the early stages of death to dark red, then to purplish-red, and finally to dark purple. This invention establishes a time-color mapping curve through joint analysis of the mean hue and the mean saturation.

[0036] Preferably, the color analysis branch first calculates the hue mean. Mapped to labels representing the stages of livor mortis color evolution. This embodiment defines four stages: pale pink stage (…). (corresponding to approximately 0 to 4 hours after death), dark red period ( The reddish-brown area corresponds to approximately 4 to 12 hours after death, and the purplish-red period ( (corresponding to approximately 12 to 24 hours after death) and the dark purple period ( (This corresponds to more than 24 hours after death). It should be noted that the above hue range is a typical reference value under standard environmental conditions, and in practical applications, it needs to be calibrated according to the statistical distribution of the training dataset.

[0037] Based on the stage division, the color analysis branch utilizes a regression network consisting of three fully connected layers to establish a fine mapping between color features and postmortem time. The input features are... The resulting 6-dimensional vector has 64, 32, and 1 neurons in its three fully connected layers, respectively. The first two layers use ReLU activation, while the last layer uses Softplus to ensure non-negative output. The color analysis branch outputs a time estimate of the color branch. The unit is h.

[0038] This invention further proposes a time-mapping function for the evolution of livor mortis color, used to describe the changing trend of the mean hue over time after death. This function is modeled based on a Sigmoid variant:

[0039] ,

[0040] in: The time elapsed since death, in hours; The initial value of the lividity hue in the early stage of death is taken as 15° in this embodiment, which corresponds to a light red hue. This is the final hue value when the color of livor mortis tends to stabilize. In this embodiment, the value is taken as 285°, which corresponds to a dark purple hue. The hue evolution rate constant reflects the rate of color change, and is taken as 0.12h under standard ambient temperature of 20℃. This value was obtained by fitting a training sample of 200 cases with known post-mortem elapsed time. The midpoint of the hue evolution, i.e., when the hue value reaches... The time after death is taken as 10 hours in this embodiment. The function describes an S-shaped curve, with slow hue changes in the early and late post-mortem stages, and more rapid changes in the middle stages, consistent with the color evolution of livor mortis observed in forensic medicine. The hue analysis branch applies the inverse function of this mapping function to the observed hue mean. This allows us to obtain the initial analytical values ​​for the color branch time estimate, which are then finely adjusted by the regression network.

[0041] It is worth emphasizing that the time-color mapping function proposed in this invention differs fundamentally from traditional linear regression models. Traditional methods typically assume a simple linear relationship between hue values ​​and postmortem time, ignoring the nonlinear characteristics of significantly different rates of livor mortis color evolution at different stages. The sigmoid variant function used in this invention naturally captures this nonlinear pattern: in the early postmortem stage, blood is still settling, and color changes are relatively slow; in the middle stage, hemoglobin gradually deoxygenates and is reduced, and the rate of color change reaches its peak; in the late stage, color evolution tends to saturate, and the rate of change decreases again. Furthermore, the color analysis branch introduces saturation features as an auxiliary criterion, because livor mortis has low saturation (pale color) in the early stages of formation, and as blood further deposits and diffuses in the tissue, saturation gradually increases and tends to stabilize. The joint analysis of hue and saturation provides richer temporal evolution information than a single hue feature, helping to improve the accuracy of time inference. After network training, the color analysis branch can output continuous time estimates of the input color features with an accuracy of 0.1 hours.

[0042] The distribution characteristics branch assesses the time elapsed after death from the perspective of the spatial distribution morphology of livor mortis. As time progresses after death, livor mortis exhibits a spatial evolution from scattered distribution to gradual merging, specifically manifested as increased merging degree, gradually blurred boundaries, and a gradual expansion of the diffusion range.

[0043] This invention proposes a quantitative index for corpuscular fusion. Defined as:

[0044] ,in: This represents the area of ​​the largest connected region in the livor mortis region, expressed in pixels. The total area of ​​all connected regions of livor mortis, in pixels; The number of connected components in the livor spot region; The maximum number of connected components is a preset reference value, which is set to 50 in this embodiment. The value of ranges from 0 to 1. A value closer to 1 indicates a higher degree of fusion of livor mortis, corresponding to a longer post-mortem period. When the livor mortis completely fuses into a single connected domain... , ,at this time .

[0045] Boundary clarity index This is obtained by calculating the gradient statistical features of the edges of the livor mortis region. Preferably, the Sobel operator is first used to calculate the gradient magnitude of the pixels at the edge of the livor mortis mask in the luminance V channel, and then the mean of the gradient magnitudes of all edge pixels is calculated. and standard deviation The boundary clarity index is defined as follows:

[0046] ,in: This represents the average magnitude of the edge gradient. The standard deviation of the edge gradient magnitude; To prevent extremely small constants with a denominator of zero, the value is taken as... . The value ranges from 0 to 1. A larger value indicates a clearer boundary of livor mortis and a shorter post-mortem time; conversely, as livor mortis merges and hemoglobin diffuses, the boundary tends to become blurred. The value gradually decreases.

[0047] diffusion range index The extent of livor mortis is assessed by calculating the ratio of the area of ​​the convex hull of the livor mortis region to the actual area.

[0048] ,in: This represents the area of ​​the convex hull of the livor mortis region, expressed in pixels. The value ranges from 0 to 1, with a larger value indicating a denser and more complete distribution of livor mortis. In the early stages of livor mortis formation, the scattered livor spots are sparsely distributed, and the convex hull area is much larger than the actual area. The value is relatively small; as livor mortis merges and spreads, The value gradually increases.

[0049] Distribution feature branches will affect the degree of fusion Boundary clarity and diffusion range Compositional distribution feature vector The input is fed into a regression network consisting of two fully connected layers (16 neurons and 1 neuron, with ReLU and Softplus activation functions), and the output is an estimate of the distribution branch time. The unit is h.

[0050] The finger pressure test is another important innovative feature of this invention. In forensic practice, the finger pressure test is a classic method for determining the transmissibility (early stage) and fixation (late stage) of livor mortis, but traditional methods rely on visual observation by forensic personnel and lack objective quantitative standards. This invention achieves an objective assessment of the finger pressure test through quantitative difference analysis of images before and after finger pressure.

[0051] The acupressure test branch receives two images of the same livor mortis site, one before and one after acupressure, denoted as follows: and First, preprocessing step S1 and livor mortis region segmentation step S2 are performed on both images to obtain corresponding livor mortis region masks. Then, the hue difference is calculated within the overlapping livor mortis regions of the two images. Difference in brightness :

[0052] ,

[0053] ,

[0054] in: This represents the intersection of the cadaveric lesion regions masked from two images. The total number of pixels in the intersection region; and Images before and after acupressure, showing the location. The hue value at that location; and Images before and after acupressure, showing the location. The brightness value at that location.

[0055] This invention sets a fixed threshold. and As a quantitative standard for determining the fixation of livor mortis. Preferably, , (The difference in brightness after normalization to 0 and 1). When and At that time, it was determined to be fixed livor mortis, and the time stage after death was marked as the late stage (usually corresponding to more than 12 hours after death); when or When the condition is diagnosed as metastatic livor mortis, the postmortem time stage is marked as the early stage (usually within 12 hours after death). The above threshold was determined by ROC curve analysis of 150 labeled samples with known postmortem time, under the principle of maximizing Youden's index.

[0056] Furthermore, the branch of acupressure testing is integrated. and Calculate the finger pressure response index : ,in: The range of values ​​is limited to between 0 and 1 by truncation. A value close to 1 indicates that the livor mortis is highly fixed, with almost no change before and after finger pressure, corresponding to a longer time after death; A value close to 0 indicates highly transferable livor mortis, significant fading upon pressure, and a relatively short post-mortem period. The pressure test branch will... Using stage labels as auxiliary features, the estimated finger pressure branch time is output through linear regression. .

[0057] The time estimates from the three branches are integrated through a learnable weighted fusion layer. The fusion layer employs an attention-weighted mechanism, adaptively adjusting the weights based on the confidence level of each branch's output.

[0058] ,

[0059] in: This is a preliminary estimate of the time elapsed after death, in hours. , and The fusion weights of the three branches satisfy the following conditions: Furthermore, all weights are non-negative. Preferably, the weights are automatically learned based on the reliability of the features in each branch through a Softmax-normalized attention network. In the initial training phase, the weights are initialized to... , , This reflects the dominant role of color features in time inference.

[0060] IV. Step S4: Environmental Factor Correction and Multi-site Fusion. Ambient temperature has a significant impact on the formation rate and evolution process of livor mortis. At higher temperatures, increased blood flow and accelerated hemoglobin degradation lead to faster livor mortis formation and fixation; conversely, the opposite is true at lower temperatures. Therefore, it is necessary to perform temperature correction on the preliminary estimate of postmortem time.

[0061] The ambient temperature correction module designed in this invention uses a nonlinear temperature-time response function for correction. Let the ambient temperature be... (Unit: °C), Standard reference temperature is ℃, temperature correction factor Defined as:

[0062] ,

[0063] in: The temperature sensitivity coefficient reflects the degree to which temperature changes affect the rate of livor mortis evolution. In this embodiment, it was obtained by fitting data from 300 samples under different temperature conditions. ℃ This value causes the correction factor to decrease to approximately 0.70 times its original value for every 10°C increase in temperature. When hour, The shortened time estimate after correction indicates that high temperature accelerates the evolution of livor mortis; when hour, The corrected time estimate is prolonged, indicating that low temperature slows down the evolution of livor mortis. The effective range for temperature correction is 0℃ to 40℃.

[0064] The corrected time estimate for a single location is:

[0065] ,

[0066] in: For the first Preliminary post-mortem time estimates for each sampling site; For the temperature-corrected first Time estimates for each part.

[0067] Multi-site fusion branching complex cervical ( ), back ( ) and lower limbs ( The corrected time estimates for the three body parts. Since the image quality of livor mortis images may vary across different body parts, this invention employs an adaptive weighted fusion strategy based on image quality scores. Image quality score for each body part. Based on image sharpness Percentage of livor mortis area and mean color saturation Comprehensive calculation:

[0068] ,

[0069] in: For the first The Laplacian variance normalized value of each part of the image reflects the image sharpness, and the value ranges from 0 to 1. For the first The proportion of the area of ​​livor mortis in each part of the image to the total area of ​​the image, with a value ranging from 0 to 1; For the first The average saturation value of the livor mortis region in each body part, ranging from 0 to 1. The final time estimate for multi-body fusion is:

[0070] ,

[0071] in: This is the final estimated time elapsed after death after correction, in hours. This weighted fusion strategy gives greater weight to areas with higher image quality and more prominent livor features in the final estimate, thereby improving the accuracy and robustness of the overall inference.

[0072] In a preferred embodiment of the present invention, the multi-part fusion branch also incorporates an outlier detection mechanism. Before calculating the final fusion time estimate, the system first assesses the dispersion between the corrected time estimates for each part. Specifically, it calculates the maximum absolute difference between any two of the three part estimates. ,when Exceeding the preset abnormal threshold Time (in this embodiment) (h) If the system identifies abnormal data, it automatically removes the abnormal area with the largest deviation from the estimated values ​​of the other two areas, and only performs weighted fusion on the remaining two areas. This anomaly detection mechanism can effectively handle situations where individual areas have abnormal livor mortis characteristics due to factors such as local lesions, external injuries, or clothing obscuring the image, further ensuring the reliability of the fusion results. Furthermore, when only images of livor mortis from some areas can be acquired (e.g., due to the posture of the body or limitations of the scene), the system automatically degrades to a weighted fusion mode for the available areas, ensuring that valid inference results can still be output under non-ideal acquisition conditions.

[0073] V. Step S5: Estimation of the time interval of death and generation of the report, based on the final estimated value of the corrected elapsed time after death. This step calculates the estimated time of death interval and generates an auxiliary identification report. Considering the inherent uncertainty in estimating the time of death, outputting point estimates combined with confidence intervals better meets the needs of forensic practice.

[0074] This invention determines the confidence level by evaluating the consistency of the estimated values ​​of each part in multi-part fusion. Consistency level Defined as the inverse function of the coefficient of variation of the corrected time estimates for each part:

[0075] ,

[0076] in: This is the mean of the corrected time estimates for the three locations; This represents the standard deviation of the corrected time estimates for the three locations. The value ranges from 0 to 1. A higher value indicates that the estimated values ​​of each part are more consistent and the confidence level of the inference is higher.

[0077] Based on consistency The upper and lower bounds of the time interval of death are calculated as follows:

[0078] ,

[0079] ,

[0080] in: The basic time margin is set to 2 hours in this embodiment. This value is determined based on the conventional error range for estimating postmortem time in forensic practice. The minimum margin factor is set to 0.2 to ensure that a certain degree of uncertainty is retained even when all parts are completely identical. Therefore, when When, the half-width of the time interval is h; when When, the half-width of the time interval is h. The inference result is based on The interval-based output effectively conveys the uncertainty information of the inference results.

[0081] In one embodiment of the present invention, the system also supports flexibly adjusting the precision level of the time interval according to the needs of the case. When the appraiser selects the high-precision mode, the basic time margin is... The time frame is reduced to 1.5 hours, suitable for cases requiring high time accuracy; when the conservative mode is selected, The time frame has been increased to 3 hours to reduce the risk of missed coverage, making it suitable for preliminary screening scenarios where time accuracy requirements are relatively lenient. The accuracy level can be manually set by the investigator based on the specific circumstances of the case, or it can be automatically recommended by the system based on the confidence level. When the confidence level is high, the system defaults to recommending the high-accuracy mode; when the confidence level is low, the system defaults to recommending the conservative mode, thus achieving a dynamic balance between inference accuracy and reliability.

[0082] Regarding report generation, the system output includes the following three parts: The first part is the livor mortis labeled image, where the segmented livor mortis areas are marked on the original acquired image with a semi-transparent colored overlay, and each area is labeled with a text label indicating the color stage classification and finger pressure test results. The second part is a summary of the estimated time of death, including the independent time estimate for each body part, temperature correction coefficient, fusion weight, final time estimate, and time interval. The third part is supplementary information for the expert opinion, outputting the basis for the inference, analysis of influencing factors, and confidence rating in structured text format, with the confidence rating divided into high ( ),middle( ) and low ( The report is divided into three levels. The entire report generation process is automated, providing forensic experts with objective and traceable technical support documents.

[0083] All neural network models involved in the above methods require pre-training. In one embodiment of the present invention, the construction process of the training dataset is as follows: Autopsy case data approved by ethical standards are obtained from cooperating forensic institutions. Each case includes standardized images of multiple livor mortis sites, ground truth values ​​of post-mortem time annotated by forensic personnel, environmental temperature records, and finger pressure test image pairs. Preferably, the training dataset contains no fewer than 1500 samples, covering a time range of 0 to 72 hours post-mortem, with environmental temperatures ranging from 5°C to 38°C. The dataset is divided into a training set, a validation set, and a test set in a 7:1.5:1.5 ratio.

[0084] During training, data augmentation strategies included random brightness adjustments (ranging from 0.8 to 1.2 times the original value), random contrast adjustments (ranging from 0.85 to 1.15 times), random horizontal flips, and random rotations (ranging from -15° to +15°). It should be noted that the hue channel was not randomly offset to avoid disrupting the intrinsic mapping between livor mortis color and postmortem time.

[0085] The semantic segmentation network is trained using the Adam optimizer, with an initial learning rate set to [value missing]. The learning rate decays using a cosine annealing strategy, with 120 epochs of training and a batch size of 8. The association modeling network is also trained using the Adam optimizer with an initial learning rate of... The loss function uses a weighted combination of smooth L1 loss and stage-specific cross-entropy loss:

[0086] ,

[0087] in: This represents the actual time elapsed since death, in hours (h). The stage classification probability is output by the color analysis branch; For the true labels of stage classification; The weighting coefficient for the stage classification loss is set to 0.3 in this embodiment. The smoothed L1 loss exhibits the characteristic of being a linear function when the error is large and a quadratic function when the error is small, which makes the model training process more robust to outlier samples.

[0088] The overall data flow is as follows: Raw image data undergoes preprocessing in step S1 to obtain an HSV multi-channel image; after segmentation in step S2, a 48-dimensional color feature vector is extracted; after three-branch parallel processing in step S3, a preliminary time estimate is obtained; after temperature correction and multi-site fusion in step S4, a final time estimate is obtained; and after interval calculation in step S5, a complete inference report is output. The data interfaces between each step are strictly defined, with the output of the previous step directly serving as the key input for the next, forming an end-to-end automated processing pipeline. Preferably, during the inference stage, the system's processing time for a single image does not exceed 2 seconds (tested on an NVIDIA RTX 3060 graphics card platform), meeting the practical needs of rapid inference at forensic scenes.

[0089] like Figure 2As shown, the intelligent inference system for death time of forensic livor mortis images of the present invention includes a standardized acquisition and preprocessing module 1 for livor mortis images, a livor mortis region segmentation and color feature extraction module 2, a livor mortis color evolution and post-death time correlation modeling module 3, an environmental factor correction and multi-site fusion module 4, and a death time interval estimation and report generation module 5. Each module corresponds one-to-one with steps S1 to S5 in the method embodiment.

[0090] The cadaver image standardization acquisition and preprocessing module 1 includes a standardization light source submodule, a color calibration submodule, and a color space conversion submodule. The standardization light source submodule comprises a ring-shaped LED array and an illumination uniformity control unit, with a color temperature set within the range of 5000K to 6500K, preferably 5500K, and illumination uniformity not less than 90%. The color calibration submodule integrates a standard 24-color chart and a color correction matrix calculation unit, automatically completing the calibration process upon initiation of each acquisition task. The color space conversion submodule performs white balance correction and standard RGB-to-HSV conversion, outputting a 3-channel HSV feature image. Preferably, this module is integrated into a portable acquisition terminal, which is equipped with an image sensor with a resolution of not less than 12 megapixels and supports RAW and PNG image storage formats.

[0091] The livor mortis region segmentation and color feature extraction module 2 includes a semantic segmentation submodule and a color analysis submodule. The semantic segmentation submodule deploys the aforementioned encoder-decoder structure segmentation network to achieve pixel-level classification of the livor mortis region, normal skin region, and background region. The color analysis submodule, based on the livor mortis mask obtained from segmentation, calculates the statistical features and hue histogram features of the H, S, and V channels in parallel, outputting a 48-dimensional livor mortis color feature vector. In one embodiment of the invention, this module is deployed in an edge computing unit, preferably using an embedded GPU platform to meet the needs of real-time on-site processing.

[0092] The Module 3, which models the correlation between livor mortis color evolution and postmortem time, is the core inference engine of the system. It includes a color analysis branch, a distribution feature branch, a pressure test branch, and a fusion layer. The color analysis branch contains a stage classification network and a time regression network. Based on hue and saturation information in the color features, it establishes a time-color mapping curve and outputs a time estimate for the color branch. The distribution feature branch contains a fusion degree calculation unit, a boundary sharpness calculation unit, and a diffusion range calculation unit. Based on the spatial distribution morphology characteristics of the livor mortis area, it outputs a time estimate for the distribution branch. The pressure test branch contains an image registration unit and a difference analysis unit. By comparing the hue and brightness differences before and after pressure testing, it determines the fixation of the livor mortis and outputs a time estimate for the pressure test branch and a stage label. The fusion layer employs an attention weighting mechanism to adaptively weight and fuse the outputs of the three branches to obtain a preliminary estimate of the postmortem time. The parallel processing of the three branches and the collaborative working mechanism of the fusion layer enable the system to comprehensively determine the postmortem time from three complementary dimensions: color evolution, spatial distribution, and mechanical response, achieving synergistic efficiency at the information fusion level.

[0093] The environmental factor correction and multi-site fusion module 4 includes a temperature correction submodule and a multi-site weighted fusion submodule. The temperature correction submodule receives ambient temperature values ​​from an external temperature sensor or manually input, calculates a temperature correction coefficient based on the exponential temperature-time response function described in the method embodiment, and corrects the preliminary time estimates for each site. The multi-site weighted fusion submodule receives the temperature-corrected time estimates for the neck, back, and lower limbs respectively, performs adaptive weighted fusion based on the image quality scores of each site, and outputs the final post-mortem time estimate. Preferably, the temperature sensor is integrated into the acquisition terminal to achieve automatic acquisition and recording of ambient temperature.

[0094] The Time of Death Interval Estimation and Report Generation Module 5 includes an interval calculation submodule and a report generation submodule. The interval calculation submodule calculates the confidence level and the upper and lower bounds of the time interval based on the consistency of estimates from multiple body parts. The report generation submodule automatically generates a structured report containing images with livor mortis annotations, a summary of estimated time of death, and supplementary information from expert opinions. It supports PDF and standardized data format output, facilitating archiving and subsequent review.

[0095] Data communication between the above modules is achieved through a unified message bus. The output of the livor mortis image standardization acquisition and preprocessing module 1 is transmitted to the livor mortis region segmentation and color feature extraction module 2 via the message bus. The output of the livor mortis region segmentation and color feature extraction module 2 is transmitted to the livor mortis color evolution and postmortem time correlation modeling module 3. The output of the livor mortis color evolution and postmortem time correlation modeling module 3 is transmitted to the environmental factor correction and multi-site fusion module 4. The output of the environmental factor correction and multi-site fusion module 4 is transmitted to the death time interval estimation and report generation module 5, forming a complete end-to-end automated processing pipeline. At the same time, the confidence information of the death time interval estimation and report generation module 5 can be fed back to the livor mortis image standardization acquisition and preprocessing module 1 to guide whether additional sites need to be acquired or sites with poor image quality need to be re-acquired, forming a closed-loop optimization mechanism.

[0096] In a preferred embodiment of the present invention, the system's hardware deployment adopts a collaborative architecture between the edge and the cloud. The edge includes a portable acquisition terminal and an embedded inference unit, responsible for image acquisition, preprocessing, segmentation, and preliminary inference; the cloud deploys a high-performance computing server, responsible for the regular updating and training of the model and the refined analysis of complex cases. Data transmission between the edge and the cloud is conducted through a secure encrypted channel. In scenarios where the network is unavailable, the edge can independently complete the entire inference process. The system also integrates a data management submodule, which encrypts and manages the storage and versions of all acquired images, inference results, and identification reports. It supports retrieval and querying by case number, time range, and identification personnel, meeting the requirements of data traceability and the integrity of the evidence chain in forensic identification work. In addition, the system provides a standardized application programming interface, supporting data interface with existing forensic information management systems to achieve seamless integration of the identification process.

[0097] To verify the effectiveness of the method of the present invention, a systematic performance evaluation was conducted on a test dataset. The test dataset contained 300 independent samples, covering a time range of 0 to 72 hours post-mortem, with an ambient temperature range of 5°C to 38°C, and included cases of different genders, ages, and body types.

[0098] Regarding the accuracy of estimating elapsed time after death, the method of this invention has a mean absolute error (MAE) of 1.8 h and a median absolute error of 1.4 h across the entire time range. Looking at different stages, the MAE is 1.2 h in the early stage (0-12 h post-mortem), 1.9 h in the middle stage (12-24 h post-mortem), and 2.6 h in the late stage (24-72 h post-mortem). This result indicates that the method of this invention has higher estimation accuracy in the early post-mortem stage, which is consistent with the forensic pattern that livor mortis color changes significantly in the early stages. Further analysis of the coefficient of determination (R²) for each stage shows a value of 0.94 for the early stage, 0.88 for the middle stage, and 0.79 for the late stage, with an overall R² of 0.91, indicating that the color-time mapping relationship established by the method of this invention has a high goodness of fit.

[0099] In a comparative experiment with existing technologies, the method of this invention was compared with the following baseline methods: (a) a traditional visual judgment method based on experienced forensic personnel, where the mean of postmortem time estimates was taken from three experts with over 10 years of forensic experience; (b) a linear regression method using only a single color feature; and (c) a traditional forensic method using only temperature-rigor mortis assessment. On the same test set, the MAE of the traditional visual judgment method was 3.5 h, the MAE of the linear regression method using only a single color feature was 2.8 h, and the MAE of the temperature-rigor mortis assessment method was 3.1 h. The MAE of the method of this invention was 1.8 h, which is approximately 48.6% lower than the traditional visual judgment method and approximately 35.7% lower than the linear regression method using only a single color feature, demonstrating a significant performance improvement. It is worth noting that the inference results of the method of the present invention are completely consistent among different appraisers (because the system output is a deterministic result), while the consistency among three experts in the traditional visual judgment method is only ICC=0.72, indicating that the method of the present invention has a significant advantage in terms of the objectivity and repeatability of the appraisal conclusion.

[0100] In the ablation experiment, the MAE after removing the color analysis branch, distribution feature branch, and finger pressure test branch were 2.4h, 2.1h, and 2.0h, respectively, all higher than the 1.8h of the complete three-branch model, indicating that each branch contributes positively to the final inference accuracy. In particular, the removal of the color analysis branch resulted in the most significant increase in MAE (from 1.8h to 2.4h), confirming the dominant role of color features in time inference. After removing the ambient temperature correction module, under non-standard temperature conditions (…), ℃ or The MAE at extreme temperatures increased from 2.3h to 3.8h, indicating that the temperature correction module is particularly crucial for improving inference accuracy under extreme temperature conditions. Further analysis of the correction effect of the temperature correction module in different temperature ranges revealed the following: In the low-temperature range of 5℃ to 10℃, the corrected MAE decreased from 4.2h to 2.5h, a reduction of 40.5%; in the high-temperature range of 30℃ to 38℃, the corrected MAE decreased from 3.6h to 2.1h, a reduction of 41.7%; within the standard temperature range of 15℃ to 25℃, the change in MAE before and after correction was minimal (from 1.7h to 1.6h), which is expected, as the correction coefficient is close to 1 under standard temperature conditions. The introduction of the multi-site fusion strategy reduced the MAE from the single-site optimum of 2.2h to 1.8h, a reduction of approximately 18.2%. The livor mortis on the back, due to its larger area and more prominent features, typically received the highest weight in multi-site fusion (average weight approximately 0.45), while the average weights for the neck and lower limbs were approximately 0.28 and 0.27, respectively.

[0101] Regarding the calibrability of the time interval estimation, the reliability of the interval estimation was evaluated with a target coverage rate of 95%. Test results show that the time interval output by the method of this invention covers 93.7% of the actual elapsed time after death, close to the theoretical target value of 95%, indicating that the interval estimation has good calibration performance. The average interval width is 3.2 hours, significantly narrower than the typical time intervals (usually 6 to 12 hours) provided by traditional forensic personnel, providing more accurate time reference information for forensic practice.

[0102] The above experimental results fully demonstrate that the intelligent time of death estimation method based on the color evolution analysis of livor mortis images proposed in this invention is significantly superior to existing technical solutions in terms of estimation accuracy, robustness, and practicality, and can provide objective and quantitative technical support for forensic time of death identification.

[0103] In practical application testing, the method of this invention underwent a six-month trial evaluation at forensic identification institutions in five different regions. During the trial, a total of 120 actual cases were processed. Feedback results showed that the acceptance rate of the system's output results by forensic personnel was 91.7%. Specifically, in 110 cases, the time interval of death inferred by the system was consistent with the time of death ultimately determined by the forensic personnel or was adopted as an important reference. In the remaining 10 cases, the main reasons for the significant deviation in the system's inference results were abnormal livor mortis caused by special causes of death (such as cherry-red livor mortis due to carbon monoxide poisoning) and extreme environmental conditions (such as frozen environments or high-temperature decomposition environments). For these special circumstances, the system has added a risk warning function to the report, reminding forensic personnel to pay attention to abnormal factors that may affect the accuracy of the inference. The method of this invention can automatically identify the above-mentioned abnormal situations and output early warning information, suggesting that forensic personnel combine it with other forensic indicators for comprehensive judgment, reflecting the design concept of human-machine collaboration. From a practical point of view, the system takes an average of 3 minutes from image acquisition to report output. It includes all steps such as image preprocessing, region segmentation, feature extraction, time inference, temperature correction, multi-site fusion and report generation. Compared with the traditional identification process, which takes 10 to 20 minutes just for the observation and recording of livor mortis, it significantly improves the identification efficiency.

[0104] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for intelligently estimating the time of death from forensic livor mortis images, characterized in that, Includes the following steps: Step S1, Standardized acquisition and preprocessing of livor mortis images: Under the control of a standardized light source and color calibration device, high-resolution images of the livor mortis area of ​​the lower part of the corpse are acquired, and white balance correction and color space conversion are performed on the acquired high-resolution images of the livor mortis area. The high-resolution images of the livor mortis area are converted from the RGB color space to the HSV color space to obtain a multi-channel livor mortis feature image containing hue channel, saturation channel and lightness channel. Step S2, Lividity region segmentation and color feature extraction: Automatic skin region segmentation is performed on the multi-channel livor mortis feature image. The livor mortis region is separated from the non-livor mortis region by pixel-level classification based on semantic segmentation network. The hue distribution features, saturation distribution features and brightness distribution features are extracted from the segmented livor mortis region using a color analysis network to construct the livor mortis color feature vector. Step S3, Modeling the Correlation Between Lividity Color Evolution and Postmortem Time: The livor mortis color feature vector is input into the correlation modeling network for time inference. The correlation modeling network includes a color analysis branch, a distribution feature branch, and a finger pressure test branch. The color analysis branch quantifies the gradual change of livor mortis color from a light red stage through a dark red stage and a purplish-red stage to a dark purple stage based on the numerical change of hue channel values. Each color evolution stage is quantitatively divided by a preset numerical range of hue channel mean values, and a nonlinear mapping curve between hue channel values ​​and postmortem time is established. The distribution feature branch evaluates the degree of livor mortis fusion, boundary clarity, and diffusion range changes. The finger pressure test branch determines the transferability and fixation of livor mortis through image difference analysis before and after finger pressure to determine the postmortem time stage. The outputs of the three branches are fused through a fusion layer to obtain a preliminary estimate of the postmortem time. Step S4, Environmental Factor Correction and Multi-site Fusion: The environmental temperature correction module is used to perform nonlinear temperature correction on the preliminary estimate of postmortem time based on the ambient temperature of the body. The inference results of multiple livor mortis features on the neck, back and lower limbs are combined through the multi-site fusion branch, and an adaptive weighted fusion strategy is used to obtain the corrected estimate of postmortem time.

2. The intelligent method for inferring the time of death from forensic livor mortis images according to claim 1, characterized in that, In step S1, the color temperature range of the standardized light source is 5000K to 6500K, the illumination uniformity is not less than 90%, the color calibration device includes a standard color card, the standard color card is photographed before acquiring the livor mortis image to generate a color correction matrix, and the resolution of the acquired image is not less than 12 million pixels.

3. The intelligent method for inferring the time of death from forensic livor mortis images according to claim 1, characterized in that, In step S2, the semantic segmentation network adopts an encoder-decoder structure. The encoder contains four downsampling stages, each of which contains a residual convolutional block and an attention module. The decoder fuses multi-scale features through skip connections. The pixel-level classification results of the segmentation output include livor mortis region category, normal skin category, and background category.

4. The intelligent method for inferring the time of death from forensic livor mortis images according to claim 1, characterized in that, In step S4, the temperature correction range of the ambient temperature correction module is from 0℃ to 40℃. When the ambient temperature is higher than the standard reference temperature, the estimated time after death is shortened, and when the ambient temperature is lower than the standard reference temperature, the estimated time after death is extended. The standard reference temperature is set to 20℃.

5. The intelligent method for inferring the time of death from forensic livor mortis images according to claim 1, characterized in that, In step S3, the color analysis branch maps the hue channel values ​​to preset labels for the evolution stages of livor mortis color. The labels for the evolution stages of livor mortis color include light red, dark red, purplish red, and dark purple. Each stage is quantitatively divided by a preset range of hue channel mean values. The color analysis branch establishes a nonlinear mapping relationship between the hue mean and saturation mean and the time elapsed after death through a regression network, and outputs the color branch time estimate.

6. The intelligent method for inferring the time of death from forensic livor mortis images according to claim 1, characterized in that, In step S3, the distribution feature branch quantifies the degree of fusion by calculating the proportion of the connected domain area of ​​the livor mortis region, assesses the boundary clarity by calculating the mean and variance of the edge gradient of the livor mortis region, and assesses the diffusion range by calculating the ratio of the convex hull area of ​​the livor mortis region to the actual area. The degree of fusion, boundary clarity and diffusion range are combined into a distribution feature vector, which is input into the fully connected regression layer to obtain the distribution branch time estimate.

7. The intelligent method for inferring the time of death from forensic livor mortis images according to claim 1, characterized in that, In step S3, the finger pressure test branch receives two images of the same livor mortis area before and after finger pressure, calculates the hue difference and brightness difference in the two images within the livor mortis area, and determines that the livor mortis is fixed and the time stage after death is marked as the late stage when both the hue difference and brightness difference are less than the preset fixation threshold.

8. The intelligent method for inferring the time of death from forensic livor mortis images according to claim 1, characterized in that, In step S4, the multi-site fusion branch obtains the estimated time elapsed after death for the neck livor mortis region, back livor mortis region, and lower limb livor mortis region, respectively. The estimated values ​​for each site are adaptively weighted and fused based on the image quality score of each site's livor mortis. The image quality score is calculated based on image sharpness, livor mortis area ratio, and color saturation.

9. The intelligent method for inferring the time of death from forensic livor mortis images according to claim 8, characterized in that, It also includes step S5, estimation of time interval of death and generation of report: Based on the corrected postmortem estimating value and its confidence level, the time interval of death is estimated, and an inference result report containing the livor mortis labeled image, the time interval of death estimate and the auxiliary information of the identification opinion is generated. The calculation method of the time interval of death is: taking the corrected postmortem estimating value as the center, the upper and lower bounds of the time interval are determined according to the confidence level. The confidence level is calculated based on the consistency of the estimated values ​​of each part in the multi-part fusion branch. The higher the consistency level, the narrower the time interval.

10. A forensic livor mortis image intelligent time-of-death estimation system, used to implement the forensic livor mortis image intelligent time-of-death estimation method as described in claim 9, characterized in that, include: The standardized acquisition and preprocessing module for livor mortis images is used to acquire high-resolution images of the livor mortis area of ​​the lower part of the body under the control of a standardized light source and color calibration device, and to perform white balance correction and color space conversion to obtain multi-channel livor mortis feature images. The livor mortis region segmentation and color feature extraction module is used to automatically segment the skin region of multi-channel livor mortis feature images to separate the livor mortis region, and to extract color features using a color analysis network to construct a livor mortis color feature vector. The mortuary stain color evolution and post-mortem time correlation modeling module includes a color analysis branch, a distribution feature branch, and a finger pressure test branch, which is used to perform time correlation modeling on the mortuary stain color feature vector and output a preliminary estimate of the time elapsed after death. The environmental factor correction and multi-site fusion module is used to perform nonlinear temperature correction based on ambient temperature, and to perform adaptive weighted fusion by integrating the inference results of multiple livor mortis features from the neck, back and lower limbs to obtain the corrected estimate of postmortem time. The death time interval estimation and report generation module is used to calculate the death time interval estimate based on the corrected postmortem estimator and its confidence level, and generate an inference result report that includes livor mortis-annotated images, death time interval estimates, and auxiliary information of expert opinions.

Images