Femoral species identification system and method based on scanning electron microscope and artificial intelligence
By combining scanning electron microscopy and artificial intelligence image recognition technology, a ResNet50-based identification model was constructed, enabling efficient and accurate species identification of damaged or fragmented bone samples, thus overcoming the shortcomings of traditional forensic methods.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-19
- Publication Date
- 2026-03-27
AI Technical Summary
Traditional forensic methods struggle to accurately identify species when dealing with severely damaged or fragmented bone samples, and microscopy techniques require destructive processing of the samples.
By combining scanning electron microscopy (SEM) with artificial intelligence image recognition, a recognition model is constructed using a ResNet50 network. High-resolution images of skeletal microstructures are obtained using SEM, and two-stage cascaded prediction is performed to achieve automated classification.
It significantly improves the accuracy and efficiency of species identification, overcoming the limitations of traditional methods, especially when dealing with damaged or fragmented samples.
Smart Images
Figure CN121747149A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of species identification technology, specifically relating to a femoral species identification system and method based on scanning electron microscopy and artificial intelligence. Background Technology
[0002] Species identification is a crucial aspect of forensic anthropology and plays a vital role in assisting in the investigation of criminal cases. Traditional methods distinguish human from non-human skeletons by visually observing morphological characteristics, but these methods have limitations when dealing with severely damaged or fragmented bone samples. While microscopy can observe microscopic structures, it requires destructive sample processing. Scanning electron microscopy (SEM) overcomes this limitation, allowing direct observation of the original bone surface and providing high-resolution images. Combined with energy dispersive spectroscopy (EDS) for component analysis, it offers more comprehensive technical support for species identification.
[0003] The development of artificial intelligence (AI) technology has brought new opportunities to forensic medicine. Image recognition technology based on machine learning can reduce subjective errors, automate and standardize the identification process, and improve the accuracy and efficiency of results.
[0004] However, there is currently a lack of applications that combine the two. Therefore, this study combines SEM technology with AI image recognition to develop an efficient and accurate solution for femoral species identification. Summary of the Invention
[0005] In view of the technical defects mentioned in the background art, the purpose of this invention is to provide a femoral species identification system and method based on scanning electron microscopy and artificial intelligence. It aims to overcome the shortcomings of traditional methods by combining SEM technology with AI image recognition, thereby improving identification accuracy and significantly increasing identification efficiency.
[0006] To achieve the above objectives, in a first aspect, embodiments of the present invention provide a femoral species identification system based on scanning electron microscopy and artificial intelligence, the system comprising:
[0007] The data module is used to acquire datasets and perform data preprocessing; wherein, the dataset includes bone slice image data of femoral bones of different species acquired using scanning electron microscopy, and the images in the dataset are screened, classified and amplified.
[0008] The model building module is used to build an identification model for species identification; wherein, the identification model is trained based on the ResNet50 network and preprocessed bone fragment image data; the identification model includes a two-class image recognition model for initial classification and a multi-class image recognition model for further subdivision, and the two models are trained independently;
[0009] The species identification module is used to input the femoral image data to be identified into the identification model after the identification model training is completed, and perform two-stage cascaded prediction to obtain the identification results corresponding to each image; wherein, the femoral image data is acquired by using a scanning electron microscope.
[0010] As a specific implementation of this application, when scanning with a scanning electron microscope, images are taken at multiple preset points to obtain image data at corresponding magnifications; wherein each point is associated with a preset magnification value.
[0011] As a specific implementation of this application, ImageNet pre-trained weights are used during training, and parameters are fine-tuned during model training to reduce overfitting and improve small sample performance.
[0012] As one specific implementation of this application, both models use ResNet50 as their backbone network, while also customizing the classification layer:
[0013] Add a GlobalAveragePooling2D layer for feature integration, add a fully connected layer Dense(1024, activation='relu'), and the last layer Dense(len(class_names), activation='softmax') for multi-class output. The training and output are based on the types of data provided during training.
[0014] As a preferred implementation of this application, image preprocessing is also performed using the OpenCV library during model training.
[0015] Secondly, embodiments of the present invention also provide a method for femoral species identification based on scanning electron microscopy and artificial intelligence, applied to the femoral species identification system based on scanning electron microscopy and artificial intelligence described in the first aspect, the method comprising the following steps:
[0016] Acquiring the dataset and preprocessing the data; wherein, the dataset includes bone fragment images of femoral bones of different species obtained using scanning electron microscopy, and the images in the dataset are filtered, classified and amplified.
[0017] A species identification model is constructed; wherein, the identification model is trained based on a ResNet50 network and preprocessed bone fragment image data; the identification model includes a two-class image recognition model for initial classification and a multi-class image recognition model for further subdivision, and the two models are trained independently;
[0018] After the recognition model training is completed, the femoral image data to be identified is input into the recognition model for two-stage cascaded prediction to obtain the identification results corresponding to each image; wherein, the femoral image data is acquired by using a scanning electron microscope.
[0019] The technical solution provided by this invention combines SEM technology with AI image recognition, effectively integrating the high-resolution imaging capability of SEM with the efficient analysis advantages of AI image recognition. This provides an innovative approach and strong technical support for species identification in the field of forensic anthropology. Compared with the limitations of traditional forensic methods in handling bone fragments or severely damaged samples, this solution directly obtains images of the microstructure of bones through SEM and uses AI for automated image recognition and classification, significantly overcoming the shortcomings of traditional methods. This not only improves the accuracy of identification but also greatly increases the efficiency of identification. Attached Figure Description
[0020] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below.
[0021] Figure 1 This is an architecture diagram of a femoral species identification system based on scanning electron microscopy and artificial intelligence provided in an embodiment of the present invention;
[0022] Figure 2 This is a schematic diagram of a model training process provided in an embodiment of the present invention;
[0023] Figure 3 These are the ROC curves of the two types of image recognition models provided in this embodiment of the invention;
[0024] Figure 4 The ROC curves of the five-category image recognition model provided in this embodiment of the invention;
[0025] Figure 5 This is a flowchart of a femoral species identification method based on scanning electron microscopy and artificial intelligence provided in an embodiment of the present invention. Detailed Implementation
[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0027] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.
[0028] Scanning electron microscopy (SEM) is an observation method that falls between transmission electron microscopy and optical microscopy.
[0029] Energy-dispersive X-ray spectroscopy (EDS) is a detection instrument used for micro-area composition analysis. It determines the types and contents of elements by analyzing the characteristic X-ray wavelengths and intensities emitted by the sample.
[0030] It should be noted that, unless otherwise stated, the technical terms used in this embodiment have the common meaning as understood in the relevant technical field.
[0031] Please refer to Figure 1 This invention provides a femoral species identification system based on scanning electron microscopy and artificial intelligence, the system comprising:
[0032] The data module is used to acquire datasets and perform data preprocessing; wherein, the dataset includes bone slice image data of femoral bones of different species acquired using scanning electron microscopy, and the images in the dataset are screened, classified and amplified.
[0033] The model building module is used to build an identification model for species identification; wherein, the identification model is trained based on the ResNet50 network and preprocessed bone fragment image data; the identification model includes a two-class image recognition model for initial classification and a multi-class image recognition model for further subdivision, and the two models are trained independently;
[0034] The species identification module is used to input the femoral image data to be identified into the identification model after the identification model training is completed, and perform two-stage cascaded prediction to obtain the identification results corresponding to each image; wherein, the femoral image data is acquired by using a scanning electron microscope.
[0035] In this embodiment, a two-category (human and non-human) image recognition model is used to initially distinguish between human and non-human skeletons; a multi-category image recognition model is illustrated using a five-category (horse, cow, pig, monkey, sheep) image recognition model as an example.
[0036] It should be noted that the data in the dataset is used as sample data for model training.
[0037] 1.1 Sample Preparation
[0038] Human femoral samples were derived from donated samples received by the Anatomical Laboratory of Guizhou Medical University. Rhesus monkey femoral samples were provided by the collaborating institution, Kunming Institute of Zoology, Chinese Academy of Sciences (Ethical Review Number IACUC-PE-2023-07-002). Bovine, equine, ovine, and porcine femoral samples were purchased from local specialized breeding bases. Inclusion criteria for species samples: Human samples: Cause of death not involving femoral injury, age range of 28 - 60 years old, male gender; Non-human samples: Healthy before death, without femoral injury and related diseases, consistent species origin, age range corresponding to that of human samples (cattle: 2 - 7 years old, horse: 3 - 10 years old, sheep: 2 - 7 years old, pig: 2 - 5 years old, rhesus monkey: 5 - 15 years old), male gender (Ethical Review Number 2100880). Exclusion criteria for species samples: Human samples: CT scans excluded femoral injuries, femoral diseases, highly decomposed corpses, osteoporosis and other bone diseases; Non-human samples: Femoral injury, femoral disease, inconsistent species origin. This study was approved by the Ethics Committee of Guizhou Medical University (Ethical Review Number 2021 Review No. 66).
[0039] 1.2 Sample Preparation
[0040] Human (H), bovine (N), equine (M), ovine (S), porcine (P), and rhesus monkey (Q) femurs were selected. The surrounding soft tissues were removed, and the longitudinal length of the femur (the highest point of the femoral head to the lowest end of the medial condyle of the femur) was measured. The midpoint was determined and marked. From the midpoint, marks were made and cuts were made at 10 mm positions in both the proximal part (P) and distal end (D) directions (the cutting line was perpendicular to the longitudinal axis of the femur) to obtain the middle bone ring, numbered 0. From the 0 - numbered bone ring, marks were made and cuts were made at 20 mm positions in both the proximal and distal directions to obtain the proximal bone ring, numbered 1, and the distal bone ring, numbered 2. Each bone ring was marked with four quadrants A / P / L / M (A for anterior, P for posterior, M for medial, L for lateral) and cut to obtain bone blocks marked with A, P, L, and M. Using the cross - section of the femoral bone block as the observation surface, a series of low - to - high - grit metallographic sandpapers (800 grit, 1200 grit, 1500 grit, 2000 grit, 4000 grit) were successively used on a metallographic sample polishing machine (MPD - 1) at a rotational speed of 500 r / min for polishing and grinding to produce thin bone slices with a length of 5 mm, a width of 4 mm, and a thickness of 1 mm. During all cutting and grinding processes, 0.9% normal saline was sprayed to prevent the bone slices from drying and cracking, which would affect subsequent observations. The polished bone slice samples were placed in an ultrasonic cleaner for 1 hour and 30 minutes. Then, the bone slice samples were dehydrated in a gradient of alcohol (30%, 50%, 70%, 80%, 90%, 100% concentrations for 30 minutes each), air - dried for 30 minutes, and then placed in a blast dryer set at 65°C for 16 hours of drying.
[0041] 1.3 SEM Scanning
[0042] The bone fragment sample was fixed to the stage using conductive adhesive, ensuring firm adhesion, and then placed in the gold plating apparatus for 60 seconds. After plating, the bone fragment sample was scanned using a SEM with a 15 kV accelerating voltage and backscattered electron (BSE) mode. A full-scan function was used to acquire image data under a 540x microscope. Then, images were taken at 2000x, 10000x, 15000x, and 40000x magnification at five selected locations on the bone fragment sample (top, bottom, left, right, and center). The elemental composition and content of the bone fragment sample were analyzed using energy dispersive X-ray spectroscopy (EDS) on the SEM. The relative content of each element in the bone fragment sample was calculated, and the differences in the relative content of each element were analyzed.
[0043] When using a scanning electron microscope, images are taken at multiple preset points to obtain image data at corresponding magnifications; each point is associated with a preset magnification value.
[0044] Reference Figure 2 The training process of its model includes:
[0045] Femoral bones from different species were acquired and processed into bone fragments. Scanning electron microscopy was used to acquire bone fragment image data. The acquired image data underwent preliminary screening and classification according to species. Augmentation processing was performed to amplify the image data. A model training platform was built based on the ResNet50 network. When the input data consisted of two pre-classified femoral image data folders (human (H) and non-human (NH)), the trained model was a two-class image recognition model. The same approach was used for five-class (i.e., multi-class image recognition) models. Then, the two trained models were used for two-stage cascaded prediction.
[0046] Two-class image recognition models are primarily used to classify images into two main categories: humans (Class_H) and non-humans (Class_NH). This initial classification process is based on the prediction results of the two-class image recognition model. It calculates the probability distribution of the input image across each category and selects the category with the highest probability as the prediction result. If the prediction result of the two-class image recognition model is Class_NH, a five-class image recognition model is then used for more detailed classification. The five-class image recognition model can further subdivide the image into five subcategories: horse, cow, pig, monkey, and sheep (Class_M, Class_N, Class_P, Class_Q, Class_S).
[0047] When applying the model, ImageNet pre-trained weights are used during training, and parameters are fine-tuned during model training to reduce overfitting and improve small-sample performance.
[0048] Both models use ResNet50 as their backbone network, while also having a custom classification layer.
[0049] Add a GlobalAveragePooling2D layer for feature integration, add a fully connected layer Dense(1024, activation='relu'), and the last layer Dense(len(class_names), activation='softmax') for multi-class output. The training and output are based on the types of data provided during training.
[0050] Furthermore, during model training, image preprocessing is performed using the OpenCV library.
[0051] During implementation, a convolutional neural network (CNN) was built using the TensorFlow framework, and ResNet50 was introduced to utilize its powerful deep learning capabilities to process image data. At the same time, image preprocessing was performed using the OpenCV library to further optimize the system's performance, ultimately achieving an efficient and accurate image recognition system that meets the application needs of various complex scenarios.
[0052] To verify the system's performance, the research group loaded a batch of test images from a designated image folder. These images were first resized to the required input size for the model (224×224 pixels) and converted to an array format suitable for model processing. Subsequently, the system sequentially performed classification predictions on each image. For images initially classified as Class_NH, the system further subdivided them using a five-class image recognition model, ultimately outputting the most accurate class label. To visually demonstrate the classification results, a visualization method was used to display the images and their predicted labels on the same canvas page, with each image labeled with its predicted class below, facilitating observation of the model's classification performance.
[0053] In this study, the performance of the classification model was comprehensively evaluated through the calculation and visualization of a series of performance metrics. The accuracy, precision, recall, and F1 score were calculated using the accuracy, precision, recall, and F1 score functions, respectively. Precision, recall, and F1 score were all calculated using a weighted average to comprehensively consider the impact of differences in the number of samples from different classes on the metrics. Here, TP represents TruePositive, TN represents TrueNegative, FP represents FalsePositive, and FN represents FalseNegative.
[0054] Subsequently, the study further analyzed the classification performance of the model, and evaluated the classification effect of the model on different categories in more detail by calculating the ROC curve and AUC value of the model for each category.
[0055] In the process of plotting the ROC curve, the ROC curve is represented in orange and its AUC value is marked. At the same time, a blue dashed line is drawn to represent the case of random guessing, with an AUC value of 0.5, which is used as a comparison benchmark.
[0056] The ROC curve plot shows the performance of a classification model, with the horizontal axis representing the false positive rate and the vertical axis representing the true positive rate. Figure 3 The figures show the ROC curves for two-class image recognition models. The solid orange line represents the model's micro-average ROC curve, with an area under the curve (AUC) of 0.9994. The dashed blue line represents the baseline for random guessing, with an AUC of 0.5.
[0057] Figure 4 The table shows the ROC curves for the five-category image recognition model. The solid orange line represents the model's micro-average ROC curve, with an area under the curve (AUC) of 0.9859. The dashed blue line represents the baseline for random guessing, with an AUC of 0.5. Overall, the model's ROC curve is close to the upper left corner, indicating that it can distinguish between positive and negative samples well at different thresholds.
[0058] As shown in Table 1, the performance of two different image recognition models was evaluated, and their accuracy, precision, recall, and F1 score were recorded. The two-class image recognition model achieved an accuracy of 98.95%, a precision of 98.96%, a recall of 98.95%, and an F1 score of 98.95%. The five-class image recognition model achieved an accuracy of 92.35%, a precision of 92.73%, a recall of 92.35%, and an F1 score of 92.38%.
[0059] Table 1 Performance Evaluation of Two-Class and Five-Class Image Recognition Models
[0060] Evaluation parameters of two-class image recognition models Five-category image recognition model evaluation parameters Accuracy 0.9895 0.9235 Precision 0.9896 0.9273 Recall 0.9895 0.9235 F1 Score 0.9895 0.9238
[0061] The above-mentioned solution effectively integrates the high-resolution imaging capabilities of SEM with the efficient analytical advantages of AI image recognition by combining SEM technology with AI image recognition. This provides an innovative approach and strong technical support for species identification in the field of forensic anthropology. Compared with the limitations of traditional forensic methods in handling bone fragments or severely damaged samples, this solution directly obtains images of the microscopic structure of bones through SEM and uses AI for automated image recognition and classification, which significantly overcomes the shortcomings of traditional methods. This not only improves the accuracy of identification but also greatly increases the efficiency of identification.
[0062] Furthermore, in another embodiment, based on the above technical solution, to address the challenges of distinguishing certain morphologically similar species, the training process also integrates SEM image features and EDS elemental composition data (such as the relative content differences of C, O, Ca, and P observed in this study) to construct the model. Species identification is performed by combining the relative elemental content component, and this is incorporated into the identification model to improve its versatility and accuracy. This is expected to provide more accurate and efficient support for research in areas such as crime detection, species identification, and bioarchaeology.
[0063] Reference Figure 5 Based on the same inventive concept, this invention also provides a method for femoral species identification based on scanning electron microscopy and artificial intelligence, applied to the femoral species identification system based on scanning electron microscopy and artificial intelligence described in the first aspect. The method includes the following steps:
[0064] S101, Acquire dataset and perform data preprocessing; wherein, the dataset includes bone slice image data of femoral bones of different species obtained using scanning electron microscopy, and the images in the dataset are screened, classified and amplified.
[0065] S102, Construct a species identification model; wherein, the identification model is trained based on a ResNet50 network and preprocessed bone fragment image data; the identification model includes a two-class image recognition model for preliminary classification and a multi-class image recognition model for further subdivision, and the two models are trained independently;
[0066] S103, after completing the training of the recognition model, the femoral image data to be identified is input into the recognition model for two-stage cascaded prediction to obtain the identification results corresponding to each image; wherein, the femoral image data is acquired by using a scanning electron microscope.
[0067] Furthermore, the method also includes:
[0068] During training, ImageNet pre-trained weights are used, and parameters are fine-tuned during model training to reduce overfitting and improve few-sample performance.
[0069] Both models use ResNet50 as their backbone network, while also having a custom classification layer.
[0070] Add a GlobalAveragePooling2D layer for feature integration, add a fully connected layer Dense(1024, activation='relu'), and the last layer Dense(len(class_names), activation='softmax') for multi-class output. The training and output are based on the types of data provided during training.
[0071] Furthermore, during model training, image preprocessing is performed using the OpenCV library.
[0072] It should be noted that for a more detailed description of the workflow of the method embodiments, please refer to the foregoing embodiments section, which will not be repeated here.
[0073] The entire solution effectively integrates the high-resolution imaging capabilities of SEM with the efficient analytical advantages of AI image recognition by combining SEM technology with AI image recognition. This provides an innovative approach and strong technical support for species identification in the field of forensic anthropology. Compared with the limitations of traditional forensic methods in handling bone fragments or severely damaged samples, this solution directly obtains images of the microscopic structure of bones through SEM and uses AI for automated image recognition and classification, which significantly overcomes the shortcomings of traditional methods. This not only improves the accuracy of identification but also greatly increases the efficiency of identification.
[0074] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.
[0075] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices or units, or may be electrical, mechanical or other forms of connection.
[0076] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0077] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0078] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A femur species identification system based on scanning electron microscopy and artificial intelligence, characterized by, The system comprises: a data module for obtaining a data set and data preprocessing; wherein the data set comprises bone slice image data of femurs of different species obtained by using a scanning electron microscope, and the images in the data set are subjected to screening, classification and expansion processing; a model construction module for constructing an identification model for species identification; wherein the identification model is trained based on a ResNet50 network and the preprocessed bone slice image data; the identification model comprises a two-class image identification model for preliminary classification and a multi-class image identification model for further subdivision, and the two models are independently trained; a species identification module for, after the training of the identification model is completed, inputting femur image data to be identified into the identification model for two-stage cascade prediction to obtain identification results corresponding to each image; wherein the femur image data is obtained by using a scanning electron microscope.
2. The femur species identification system based on scanning electron microscopy and artificial intelligence according to claim 1, wherein, When scanning by using the scanning electron microscope, a plurality of preset points are used for image data acquisition at corresponding magnifications; wherein each point is associated with a preset magnification value.
3. The femur species identification system based on a scanning electron microscope and artificial intelligence according to claim 1, wherein During training, ImageNet pre-training weights are used to fine-tune parameters in model training to reduce overfitting and improve small sample performance.
4. The femur species identification system based on scanning electron microscopy and artificial intelligence according to any one of claims 1 to 3, characterized in that, The backbone networks of the two models both use ResNet50, and classification layers are customized: A GlobalAveragePooling2D layer is added for feature integration, a fully connected layer Dense(1024, activation='relu') is added, and the last layer Dense(len(class_names), activation='softmax') is used for multi-classification output, which is trained and output according to the data categories provided during training.
5. The femur genus identification system based on a scanning electron microscope and artificial intelligence according to claim 4, characterized by, During model training, image preprocessing is also performed by using an OpenCV library.
6. A method for femur species identification based on scanning electron microscopy and artificial intelligence, characterized in that, The method is applied to the femur species identification system based on a scanning electron microscope and artificial intelligence in claim 1, and the method comprises the following steps: obtaining a data set and data preprocessing; wherein the data set comprises bone slice image data of femurs of different species obtained by using a scanning electron microscope, and the images in the data set are subjected to screening, classification and expansion processing; constructing an identification model for species identification; wherein the identification model is trained based on a ResNet50 network and the preprocessed bone slice image data; the identification model comprises a two-class image identification model for preliminary classification and a multi-class image identification model for further subdivision, and the two models are independently trained; after the training of the identification model is completed, inputting femur image data to be identified into the identification model for two-stage cascade prediction to obtain identification results corresponding to each image; wherein the femur image data is obtained by using a scanning electron microscope.
7. The method for femur species identification based on scanning electron microscopy and artificial intelligence according to claim 6, characterized in that, During training, ImageNet pre-training weights are used to fine-tune parameters in model training to reduce overfitting and improve small sample performance.
8. The femur species identification method based on scanning electron microscopy and artificial intelligence according to claim 6 or 7, characterized in that, The backbone networks of the two models both use ResNet50, and classification layers are customized: Add the GlobalAveragePooling2D layer for feature integration, add the fully connected layer Dense(1024, activation='relu'), and the last layer Dense(len(class_names), activation='softmax') for multi-classification output, according to the data categories provided during training and output.
9. The method for femur species identification based on scanning electron microscopy and artificial intelligence according to claim 8, characterized in that, During model training, image preprocessing is also performed through the OpenCV library.