Die penetration detection method and device, computer equipment, readable storage medium and program product

Abstract

The invention relates to a mold penetrating detection method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the steps of performing convolution processing on a to-be-detected three-dimensional model image to obtain a convolution feature map; performing weighting processing on a feature matrix in the weighting processing convolution feature map through an attention mechanism to obtain an attention feature map; on the basis of a memory cross-model detection model, performing cross-model area detection processing on the attention feature map to obtain a cross-model area detection result of the three-dimensional model image; and carrying out mold penetrating detection processing on the mold penetrating area detection result in combination with the graphic auxiliary mold penetrating detection index to obtain a mold penetrating detection result. According to the method, the attention feature graph is constructed by combining the convolution processing with the weight adjustment of the attention mechanism, and the model-penetrating detection processing of the three-dimensional model image is completed through the model-penetrating mode of the historical data by combining the memory model-penetrating detection model, so that the model-penetrating detection efficiency is effectively improved on the basis of ensuring the model-penetrating detection accuracy.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, apparatus, computer equipment, computer-readable storage medium, and computer program product for detecting through-molding. Background Technology

[0002] With the development of computer and modeling technologies, 3D modeling has emerged. 3D modeling refers to using 3D software to construct models with 3D data in a virtual 3D space. However, when 3D models are applied, such as in games, clipping issues can easily occur. Clipping refers to the phenomenon where characters or objects created through 3D modeling collide and overlap, resulting in parts of a 3D model passing through or penetrating other models. This problem, when it occurs in game scenes, disrupts the realism of the game environment and severely impacts the player's gaming experience.

[0003] However, current methods for detecting clipping mainly identify the coordinates of model feature points in an image and then perform detection based on the positional relationship of these coordinate points. However, this method is computationally complex and has low efficiency in processing large batches of clipping detection. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, computer equipment, computer-readable storage medium, and computer program product that can improve the efficiency of mold penetration detection while ensuring the accuracy of detection, in order to address the above-mentioned technical problems.

[0005] Firstly, this application provides a method for detecting penetration, including:

[0006] The 3D model image to be detected is convolved to obtain a convolutional feature map;

[0007] The feature matrix in the convolutional feature map is weighted using an attention mechanism to obtain an attention feature map.

[0008] The attention feature map is processed by the memory clipping detection model to detect clipping regions, and the clipping region detection result of the three-dimensional model image is obtained. The memory clipping detection model is trained on the initial long short-term memory network model through historical three-dimensional model images.

[0009] By combining the graphic-aided clipping detection index, the clipping region detection results are processed to obtain the clipping detection results of the three-dimensional model image to be detected.

[0010] Secondly, this application also provides a penetration detection device, comprising:

[0011] The convolution processing module is used to perform convolution processing on the 3D model image to be detected to obtain a convolution feature map;

[0012] The attention adjustment module is used to weight the feature matrix in the convolutional feature map through an attention mechanism to obtain an attention feature map;

[0013] The memory detection module is used to perform clipping region detection processing on the attention feature map based on the memory clipping detection model to obtain the clipping region detection result of the three-dimensional model image. The memory clipping detection model is trained on an initial long short-term memory network model through historical three-dimensional model images.

[0014] The clipping detection module is used to combine graphic-aided clipping detection indicators to perform clipping detection processing on the clipping region detection results, so as to obtain the clipping detection result of the three-dimensional model image to be detected.

[0015] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0016] The 3D model image to be detected is convolved to obtain a convolutional feature map;

[0017] The feature matrix in the convolutional feature map is weighted using an attention mechanism to obtain an attention feature map.

[0018] The attention feature map is processed by the memory clipping detection model to detect clipping regions, and the clipping region detection result of the three-dimensional model image is obtained. The memory clipping detection model is trained on the initial long short-term memory network model through historical three-dimensional model images.

[0019] By combining the graphic-aided clipping detection index, the clipping region detection results are processed to obtain the clipping detection results of the three-dimensional model image to be detected.

[0020] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0021] The 3D model image to be detected is convolved to obtain a convolutional feature map;

[0022] The feature matrix in the convolutional feature map is weighted using an attention mechanism to obtain an attention feature map.

[0023] The attention feature map is processed by the memory clipping detection model to detect clipping regions, and the clipping region detection result of the three-dimensional model image is obtained. The memory clipping detection model is trained on the initial long short-term memory network model through historical three-dimensional model images.

[0024] By combining the graphic-aided clipping detection index, the clipping region detection results are processed to obtain the clipping detection results of the three-dimensional model image to be detected.

[0025] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0026] The 3D model image to be detected is convolved to obtain a convolutional feature map;

[0027] The feature matrix in the convolutional feature map is weighted using an attention mechanism to obtain an attention feature map.

[0028] The attention feature map is processed by the memory clipping detection model to detect clipping regions, and the clipping region detection result of the three-dimensional model image is obtained. The memory clipping detection model is trained on the initial long short-term memory network model through historical three-dimensional model images.

[0029] By combining the graphic-aided clipping detection index, the clipping region detection results are processed to obtain the clipping detection results of the three-dimensional model image to be detected.

[0030] The aforementioned clipping detection method, apparatus, computer equipment, computer-readable storage medium, and computer program product first perform convolution processing on the 3D model image to be detected, obtaining a convolutional feature map containing the basic features of the 3D model image. Then, an attention mechanism is used to weight the feature matrix in the convolutional feature map to obtain an attention feature map, which automatically focuses on key areas prone to clipping. Next, a memory-based clipping detection model is used to process the attention feature map for clipping region detection, obtaining the clipping region detection result of the 3D model image. This allows for the identification of not only the current clipping pattern but also the recall and recognition of clipping patterns from historical data. Finally, combined with image-aided clipping detection metrics, the clipping region detection result is processed to obtain the final clipping detection result. This application uses convolutional processing combined with an attention mechanism to extract and weight features from the input information to construct an attention feature map, thereby obtaining an attention feature map that focuses attention on key areas prone to clipping. Then, it combines a memory clipping detection model to detect clipping regions in 3D model images by using clipping patterns from historical data. Finally, it combines graphic-aided clipping detection indicators to determine clipping, thereby effectively improving clipping detection efficiency while ensuring accuracy. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a diagram illustrating the application environment of the penetration detection method in one embodiment;

[0033] Figure 2 This is a flowchart illustrating a penetration detection method in one embodiment;

[0034] Figure 3 This is a schematic diagram of the clipping phenomenon in one embodiment;

[0035] Figure 4 This is a schematic diagram of the structure of a feature extraction network based on an attention mechanism in one embodiment;

[0036] Figure 5 This is a schematic diagram of the overall process of the penetration detection method in one embodiment;

[0037] Figure 6 This is a schematic diagram illustrating the detection principle of a penetration detection method in one embodiment;

[0038] Figure 7 This is a schematic diagram illustrating the pattern-breaking detection process with the pattern-breaking area marked in one embodiment;

[0039] Figure 8 This is a flowchart illustrating the penetration detection method in another embodiment;

[0040] Figure 9 This is a structural block diagram of the penetration detection device in one embodiment;

[0041] Figure 10 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0043] The penetration detection method provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on a cloud or other network server. When a target user on terminal 102 wants to perform clipping detection on a 3D model, they can submit the image of the 3D model to be detected to server 104, which receives the image. The server then performs convolution processing on the image to be detected to obtain a convolutional feature map; it further performs weighted processing on the feature matrix in the weighted convolutional feature map using an attention mechanism to obtain an attention feature map; based on a memory clipping detection model, it performs clipping region detection processing on the attention feature map to obtain the clipping region detection result of the 3D model image. The memory clipping detection model is trained on an initial long short-term memory network model using historical 3D model images; combined with a graphics-assisted clipping detection index, it performs clipping detection processing on the clipping region detection result to obtain the clipping detection result of the 3D model image to be detected. The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle systems, and projection devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted displays. Head-mounted displays can be virtual reality (VR) devices, augmented reality (AR) devices, and smart glasses. The server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0044] In one exemplary embodiment, such as Figure 2 As shown, a method for detecting penetration is provided, which is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 201 to 207. Wherein:

[0045] Step 201: Perform convolution processing on the 3D model image to be detected to obtain a convolution feature map.

[0046] The 3D model image to be detected is the object of this application. This application is used to detect whether a 3D model in a scene exhibits clipping behavior. A 3D model is a polygonal representation of an object, typically displayed using a computer or other video equipment. The displayed object can be a real-world entity or a fictional object. Anything existing in the physical world can be represented by a 3D model. Clipping behavior can be determined by referring to... Figure 3The image shows the phenomenon of overlapping and penetrating between characters, objects, or characters and objects in a 3D model due to collisions. Convolution processing, on the other hand, refers to the process of extracting features from a 3D model image based on a trained convolutional neural network. In a convolutional neural network, convolution is a special linear transformation where the convolution kernel slides across the input data, calculating the dot product of the overlapping portions to obtain the extracted features. By performing convolution processing on the 3D model image to be detected, different features in the input 3D model image can be identified, such as edges, textures, and corners, thus obtaining a convolutional feature map. In the scheme of this application, since the input is a 3D model image, the convolution processing is 3D convolution, that is, performing convolution processing on the input 3D model image to obtain a 3D convolutional feature map.

[0047] For example, when a target object wants to perform clipping detection on multiple specified 3D model images to determine whether clipping occurs in these 3D model images, the clipping detection method of this application can be used for clipping detection. First, the target object can collect multiple 3D model images in scenes where clipping may occur. Then, through 102, these 3D model images are submitted to server 104. After obtaining the 3D model images, server 104 uses a pre-trained convolutional neural network on the server to perform 3D convolution processing on these 3D model images to obtain the convolutional feature map corresponding to each 3D model image. The first few layers of the convolutional neural network should focus on extracting basic features, such as edges and textures. As the network layers deepen, the later layers should be able to recognize more complex patterns, such as the specific shape and size of the clipping region. In one embodiment, this application is applied to clipping detection processing of 3D game resources. At this time, game testers can pre-mark scenes where clipping may occur. Then, during the game test run, multiple consecutive 3D model images in these scenes are collected, and the collected 3D model images are submitted to the server. The server performs convolution processing on the input image using a trained convolutional neural network to obtain a convolutional feature map, which is then used to perform clipping detection.

[0048] Step 203: The feature matrix in the convolutional feature map is weighted using an attention mechanism to obtain the attention feature map.

[0049] Attention mechanisms are a widely used technique in deep learning that mimics the human attention process, enabling models to focus on important parts of the input data, thereby improving overall performance and efficiency. This mechanism mainly involves two aspects: determining which part of the input to focus on; and allocating limited information processing resources to the important parts. In this application, a domain-specific attention mechanism is used. This attention network is specifically designed for clipping detection in 3D models, focusing attention on key areas prone to clipping and learning useful feature representations from them to generate attention feature maps. An attention feature map is a feature map obtained by weighting the feature matrices in the convolutional feature map using the attention mechanism; it is also a feature matrix.

[0050] For example, after obtaining the convolutional feature map, in order to identify key features and potential clipping regions in the 3D model image, an attention mechanism can be applied to the convolutional feature map to weight the feature information. The attention module implementing the attention mechanism is integrated into the convolutional neural network so that it automatically focuses on key regions during feature extraction. At this point, the network architecture based on the attention mechanism can be referenced... Figure 4 As shown, this architecture includes multiple convolutional layers, pooling layers, and at least one attention module. The attention module's role is to weight and adjust the feature information of the input feature map, allowing the network to focus on key features. Finally, an attention feature map is output. In clipping detection, key features refer to the features of regions prone to clipping. Therefore, in the feature matrix contained in the attention feature map, regions prone to clipping are assigned relatively higher weights by the attention mechanism, making their feature information more prominent and easier to identify compared to other regions in the attention feature map. In one embodiment, this application is applied to clipping detection in 3D games. A dedicated attention module for game clipping detection can be designed, and then loaded into the initial convolutional neural network to form a feature extraction network. The feature extraction network is then tested using 3D model images of 3D games (including both clipping and non-clipping cases) from historical data to obtain a network structure that can be used to extract attention features. When applied to other scenarios, images from the corresponding scenario can also be selected as model training data.

[0051] Step 205: Based on the memory clipping detection model, the attention feature map is processed to detect clipping regions, and the clipping region detection results of the three-dimensional model image are obtained. The memory clipping detection model is trained on the initial long short-term memory network model through historical three-dimensional model images.

[0052] The memory-based clipping detection model is a model that detects clipping patterns based on a memory mechanism. It is trained on an initial long short-term memory (LSTM) network model using historical 3D model images. This model stores past clipping cases and retrieves relevant information when needed. Specifically, the memory-based clipping detection model can be implemented using an LTM network model. In the overall model training process, the convolutional and attention layers are first trained using historical 3D model images. Then, the model parameters of the convolutional and attention layers are fixed, and feature extraction is performed on the historical 3D model images using these fixed-parameter convolutional and attention layers to obtain sample data for training the memory-based clipping detection model. The memory-based clipping detection model is trained using historical 3D model images. Through repeated iterative training, the memory units of the memory-based clipping detection model store various clipping patterns contained in these historical 3D model images.

[0053] For example, the attention feature map output from the attention processing can be used as input to the memory-based clipping detection model. This combines key features extracted based on the attention mechanism with clipping patterns from historical 3D model images to detect the most likely clipping regions in the image, yielding clipping region detection results. The memory-based clipping detection model can integrate the input attention feature map according to model parameters to obtain fused features that fuse the memory model parameters. Then, clipping region recognition processing can be performed based on the fused features to obtain clipping region detection results. Through the memory-based clipping detection model, the most likely clipping regions can be efficiently identified for each 3D model image, thus achieving clipping detection for 3D model images. In one embodiment, this application is applied to clipping detection in 3D games. In this case, the input 3D model images are multiple consecutive 3D model images within a scene. Here, the memory-based clipping detection model can achieve dynamic reasoning for multiple 3D model images. The memory-based clipping detection model consists of multiple consecutive memory units, and each memory unit corresponds to a time step. When multiple 3D model images are input, each 3D model image can be assigned a time step in sequence, and the output of the previous time step can be used as the input of the current time step for model calculation. This allows dynamic reasoning to be achieved by combining the clipping information in multiple consecutive 3D model images within the scene, thereby improving the accuracy of clipping region detection.

[0054] Step 207: Combine the graphic-aided clipping detection index to perform clipping detection processing on the clipping region detection results to obtain the clipping detection results of the three-dimensional model image to be detected.

[0055] Among them, the image-assisted clipping detection metric (Penetration Metric) refers to identifying whether clipping occurs in a 3D model image based on the relationship between points on the model. By combining the image-assisted clipping detection metric with the clipping regions identified by the model, clipping in 3D model images can be efficiently identified.

[0056] For example, after detecting the most likely clipping area using a memory-based clipping detection model, a graphics-assisted clipping detection index can be combined. This index is calculated based on the specific point positions in the 3D model image, enabling accurate detection of the clipping state. In a specific embodiment, the graphics-assisted clipping detection index satisfies the formula... Where I is the index quantification function, Let (O) be the number of points on the model, (N) be the space occupied by other models, and (O) be the total number of points on the model. By calculating the graphics-assisted clipping detection index, the detection index value corresponding to the input 3D model image can be determined. If the detection index value is within the normal range, the 3D model image is determined not to have clipping; if the detection index value is within the abnormal range, the 3D model image is determined to have clipping. By combining the attention mechanism with the memory clipping detection model, the clipping region most likely to occur in the 3D model image can be detected efficiently. Combined with the graphics-assisted clipping detection index, the clipping phenomenon occurring within the region can be accurately identified. Compared with directly detecting the complete 3D model image, this method can effectively improve detection efficiency.

[0057] The aforementioned clipping detection method first performs convolution processing on the 3D model image to be detected, obtaining a convolutional feature map containing the basic features of the 3D model image. Then, an attention mechanism is used to weight the feature matrix in the convolutional feature map to obtain an attention feature map, which automatically focuses on key regions. Next, a memory-based clipping detection model is used to process the attention feature map for clipping region detection, obtaining the clipping region detection result of the 3D model image. This allows the method to not only identify the current clipping pattern but also recall and identify clipping patterns from historical data. Finally, a graphics-assisted clipping detection index is used to process the clipping region detection result, yielding the final clipping detection result. This application uses convolutional processing combined with an attention mechanism to extract and weight features from the input information to construct an attention feature map, thereby obtaining an attention feature map that focuses attention on key areas prone to clipping. Then, it combines a memory clipping detection model to detect clipping regions in 3D model images by using clipping patterns from historical data. Finally, it combines graphic-aided clipping detection indicators to determine clipping, thereby effectively improving clipping detection efficiency while ensuring accuracy.

[0058] In an exemplary embodiment, step 201 includes: obtaining the layer weights and layer biases of each convolutional layer; sequentially inputting the three-dimensional model image to be detected into each convolutional layer for convolution processing to obtain the final output convolutional feature map.

[0059] For example, a convolutional neural network includes multiple sequentially arranged convolutional layers, each assigned layer weights and biases. During model training, the layer weights and biases of each convolutional layer can be adjusted to extract more accurate feature information. During convolutional processing, the layer weights and biases of each convolutional layer are first obtained. Then, after the 3D model image to be detected enters each convolutional layer, the feature extraction is sequentially processed using the layer weights and biases. For instance, when the 3D model image to be detected is input to the first convolutional layer, the first convolutional feature map is extracted using the layer weights and biases of the first convolutional layer. This first convolutional feature map is then used as the input to the second convolutional layer, and so on, until all convolutional layers have been processed. Finally, after pooling, the final output convolutional feature map is obtained. In actual processing, the first few convolutional layers of the network should focus on extracting basic features, such as edges and textures. As the network depth increases, subsequent convolutional layers should be able to recognize more complex patterns, such as the specific shape and size of piercing regions. The feature extraction convolutional operation satisfies the following formula:

[0060]

[0061] in, It is the feature map of layer (l). and These are the weights and biases of layer (l). This is the output of the previous layer. In this embodiment, the input 3D model image to be detected is processed by the layer weights and layer biases of each convolutional layer. This allows for the extraction of various feature information contained in the 3D model image in a hierarchical manner, thereby ensuring the efficiency and accuracy of feature extraction.

[0062] In an exemplary embodiment, step 203 includes: activating the convolutional feature map using an attention activation function to obtain attention weights; and performing element-wise multiplication of the weight matrix of the attention weights and the feature matrix of the convolutional feature map to obtain an adjusted attention feature map.

[0063] For example, in the attention processing, attention weights can first be determined using an attention activation function, such as sigmoid. The attention mechanism is mainly used to weight the input feature maps, allowing the network to focus on key features. The process of activating the convolutional feature maps using an attention activation function to obtain the attention weights satisfies the following formula:

[0064]

[0065] in, It is the weight matrix for attention weights. It is an activation function. and These represent the weights and biases of the convolutional layer, respectively, and X is the input feature map. After determining the attention weights, the attention weights and attention feature map can be multiplied element-wise to obtain the final output attention feature map. The weighted attention feature map satisfies the formula:

[0066]

[0067] in, It is the feature matrix of the weighted attention feature map. This indicates element-wise multiplication, which is the weight matrix of the attention weights. and convolutional feature maps Multiplying the corresponding elements in the feature matrix yields the attention-adjusted attention feature map. In one embodiment, this application is used for clipping detection of 3D game resources. In this case, a dedicated attention module for game clipping detection can be pre-designed. When extracting the attention feature map, multiple consecutive 3D model images from the game scene can be used as input data to extract convolutional feature maps. Then, for each convolutional feature map, weights are adjusted using an attention mechanism to obtain the desired attention feature map. In this embodiment, attention weights are determined using an attention activation function, and then element-wise multiplication is used to integrate the attention weights and convolutional feature maps. This effectively focuses on the key regions of the convolutional feature maps, resulting in the attention feature map.

[0068] In an exemplary embodiment, the method further includes: acquiring historical 3D model images; identifying model bounding boxes in the historical 3D model images; determining the vertex set of the historical 3D model images based on the model bounding boxes; determining the 3D model volume in the historical 3D model images based on the vertex set; performing graphics-assisted clipping detection processing based on the vertex set and the 3D model volume to obtain graphics-assisted clipping detection metrics; obtaining clipping labels for each historical 3D model image based on the graphics-assisted clipping detection metrics; performing convolution processing on the historical 3D model images through convolutional layers to obtain convolutional feature maps for each historical 3D model image; identifying the predicted labels of the convolutional feature maps; comparing the clipping labels of the historical 3D model images with the predicted labels of the convolutional feature maps to determine the model loss; and updating the parameters of the convolutional layers through backpropagation based on the model loss.

[0069] Historical 3D model images refer to case images of 3D models collected in advance according to business needs. For example, when applying clipping detection in games, historical 3D model images of 3D games can be collected. When applying clipping detection in clothing, historical 3D model images of 3D human figures and clothing can be collected. These historical 3D model images can be images where clipping has occurred or images where clipping has not occurred. Clipping annotation processing involves annotating these images to indicate whether clipping issues exist.

[0070] For example, before performing convolution processing on 3D model images, it is necessary to train the convolutional neural network used for the convolution processing. At this point, historical 3D model images can be collected in advance, and then sample data for training the convolutional model can be constructed by annotating these historical 3D model images. The annotation process for clipping cases in the historical 3D model images can specifically include recording the specific location where clipping occurs, the relevant attributes of the model, and the environmental conditions at the time of clipping. During annotation, for each model, we need to label whether a clipping problem exists and the specific type of clipping. This will serve as the basis for subsequent training and validation.

[0071] Before performing clipping detection, various basic data related to clipping detection need to be prepared. For example, the historical 3D model images used as samples need to be labeled to determine the clipping labels for each historical 3D model image. First, the bounding boxes of each model in the historical 3D model images need to be identified, which is determined by the formula... Confirmed, among which Represents the set of vertices of the model, ( () represents a single vertex. The volume of the model is then calculated through integration, with the formula V = ∫ BB dV, where (BB) represents the bounding box of the model and (dV) represents a small volume element. Based on the vertex set and the volume of the 3D model, image-assisted clipping detection is performed to obtain an image-assisted clipping detection index. The calculation process of the image-assisted clipping detection index satisfies the formula... Where I is the index quantification function, Let (O) be a point on the model, (N) be the space occupied by other models, and (O) be the total number of points on the model. By calculating the graphic-aided clipping detection index, the detection index value corresponding to the input 3D model image can be determined. If the detection index value is within the normal range, the 3D model image is determined not to have clipped; if the detection index value is within the abnormal range, the 3D model image is determined to have clipped. Finally, clipping labels for each historical 3D model image are obtained based on the graphic-aided clipping detection index. After detecting historical 3D model images with clipping phenomena, manual verification and other methods can be used to record the specific location of the clipping, the relevant attributes of the model, and the environmental conditions at the time of the clipping on these historical 3D model images.

[0072] Then, the historical 3D model images with clipping labels are used to update the parameters of the convolutional layers, ensuring that the updated convolutional layers can effectively extract the features of the clipping regions contained in the 3D model images. In one embodiment, the parameter update process of the convolutional layers can be performed by convolving the historical 3D model images to obtain convolutional feature maps for each historical 3D model image. Through layer-by-layer convolution processing, the convolutional feature maps of each historical 3D model image are first extracted. Then, the predicted labels of the convolutional feature maps are identified; here, the weights of the output convolutional feature maps can also be adjusted using an attention mechanism, and then clipping detection is directly performed on the weighted convolutional feature maps. The attention mechanism uses a weighting method that simulates human focus, adjusting the weights to areas prone to clipping. At this point, key areas of focus can be identified, and by detecting whether clipping occurs in these areas, the predicted labels of each convolutional feature map can be identified. The clipping labels of the historical 3D model images and the predicted labels of the convolutional feature maps are compared to determine the model loss. After obtaining the model loss, the parameters of the convolutional layers are updated based on the model loss through backpropagation. During model training, the selected loss calculation function specifically satisfies the following formula:

[0073]

[0074] Training a convolutional neural network involves forward propagation and backward propagation. In forward propagation, input data passes through the network and predictions are generated. The loss is then calculated using the loss function described above. In backward propagation, the gradient of the loss function is used to update the network weights. The formula for weight updates satisfies the following equation:

[0075]

[0076] in, This is the updated weight. It's the learning rate. This is the gradient of the loss function with respect to the weights. By inputting historical 3D model images with clipping labels into a convolutional neural network in batches, the network is iteratively trained, and the model parameters are updated until a convolutional layer model with satisfactory recognition accuracy is obtained. In this embodiment, historical 3D model images are labeled to construct sample data for model training. Finally, these samples are used to update the parameters of the convolutional layer, resulting in a convolutional neural network layer that can be used for clipping region feature extraction, ensuring the accuracy of convolution and attention processing.

[0077] In an exemplary embodiment, the attention feature map includes an attention feature map sequence extracted from a 3D model image sequence. Step 205 includes: adjusting the feature matrix of each attention feature map in the attention feature map sequence using the model parameters of the memory clipping detection model to obtain a fused feature map sequence; performing activation processing on the fused feature map sequence using an activation function to obtain the probability distribution matrix of the clipping region of each fused feature map in the fused feature map sequence; and parsing the probability distribution matrix of the clipping region to obtain the clipping region detection result of each 3D model image in the 3D model image sequence.

[0078] In this context, a 3D model image sequence refers to multiple 3D model images arranged in sequence. For example, for a 3D game scene, during the interaction of 3D models, 3D model images of the scene can be extracted in time sequence to form a 3D model image sequence. The attention feature maps extracted from this sequence can then form an attention feature map sequence.

[0079] For example, in the memory integration stage implemented through a memory-based pattern detection model, the goal is to combine the trained convolutional layers and attention mechanisms with the memory-based pattern detection model. This allows the model to not only identify the current pattern but also recall and recognize patterns from historical data, achieving accurate pattern detection. The memory-based pattern detection model can store past pattern cases and retrieve relevant information when needed, while the input can be a sequence of attention feature maps. Dynamic inference is performed by combining the attention feature map sequence with attention feature map sequences before and after the time sequence. By first fusing the attention feature map sequence and the model parameters of the memory-based pattern detection model, a fused feature map sequence composed of fused feature maps is obtained. The weighted attention feature maps are then processed. By adjusting the feature matrix of each attention feature map in the attention feature map sequence using the model parameters of the memory clipping detection model, the attention feature maps and the model parameters of the memory clipping detection model are optimized. Combine them to obtain the fused feature map. + Then, an activation function is applied to the fused feature map sequence to obtain the probability distribution of the clipping regions in each fused feature map; the softmax function can be used as the activation function here. The result This represents the probability distribution of clipping regions in a single fused feature map. Finally, by analyzing the probability distribution of clipping regions, the clipping region detection results for each 3D model image in the 3D model image sequence are obtained.

[0080] For the analysis process of clipping probability, we can first determine the product of the attention feature map and the probability distribution of each 3D model image in the sequence of 3D model images to obtain the clipping pattern recognition result of each 3D model image. Then, based on the clipping pattern recognition result, we determine the region with the highest clipping probability in each 3D model image and take the region with the highest clipping probability as the clipping region detection result of each 3D model image. At the same time, the clipping region detection process also needs to combine the input information of the memory unit of the memory clipping detection model at different time steps. That is, we need to first determine the unit state of the memory unit of the memory clipping detection model at different time steps; based on the unit state of the memory unit at different time steps, we adjust the feature information of each attention feature map in the attention feature map sequence to obtain the input information of the memory unit of the memory clipping detection model at different time steps. This is the input to the memory unit in the memory-based mold-detection model at time step (t). It contains historical information and contextual data related to the mold-detection task, helping the model consider previous states and information when making decisions at the current time step. , It is an attention feature map. This is an array composition module.

[0081] Finally, based on the input information of the memory unit at different time steps, the clipping detection constraint information is determined; based on the clipping pattern recognition results and the clipping detection constraint information, clipping region recognition processing is performed to determine the region with the highest clipping probability in each 3D model image.

[0082] After applying the penetration detection constraint, the overall penetration region determination process satisfies the formula. ,in, The most likely clipping region can be determined using the above formula for each input in the 3D model image sequence. Then, clipping detection is performed based on this region, which effectively improves the efficiency of clipping detection. In this embodiment, by using the 3D model image sequence as input for clipping detection and combining it with a memory-based clipping detection model for dynamic inference, the accuracy of clipping detection can be effectively improved.

[0083] In an exemplary embodiment, the method further includes: acquiring historical 3D model images; performing clipping annotation processing on the historical 3D model images to obtain clipping labels for each historical 3D model image; predicting the clipping state of the historical 3D model images based on an initial long short-term memory network model to obtain network prediction labels; comparing the clipping labels and network prediction labels to obtain the memory network loss of the initial long short-term memory network model; obtaining model parameter gradient data based on the memory network loss; and performing model parameter update processing on the initial long short-term memory network model according to the model parameter gradient data to obtain a memory clipping detection model.

[0084] For example, before performing clipping detection using a memory-based clipping detection model, the model also needs to be trained. The training process for this model can also use historical 3D model images as training data. For instance, when applied to a 3D game scene, 3D model images from that scene are selected as training data. Clipping labels are obtained for each historical 3D model image through clipping annotation. These labeled historical 3D model images can then be used as training samples. The specific model training process involves first predicting the clipping state of historical 3D model images based on an initial Long Short-Term Memory (LSTM) network model to obtain network prediction labels. This requires training with a pre-trained convolutional neural network and attention module, extracting attention feature maps from the historical 3D model images. These attention feature maps can be input into the initial LTM network model to obtain corresponding network prediction labels. Then, by comparing the clipping labels and the network prediction labels, the memory network loss of the initial LTM network model is obtained. The memory module is trained using historical 3D model images to recognize and store pattern recognition. The loss function of the memory module satisfies the following formula:

[0085] in, It is the loss function of the Long Short-Term Memory network model. It's a real label. It is a predicted label, and This represents the total number of time steps. Model parameter gradient data is obtained based on the memory network loss; finally, based on this gradient data, the initial Long Short-Term Memory (LSTM) network model is updated to obtain the memory penetration detection model. That is, the model is optimized and iterated based on the detection results to improve the accuracy and efficiency of penetration detection. The optimization of the model parameters is as follows:

[0086]

[0087] in, These are the updated model parameters. It's the learning rate. This represents the gradient of the total loss function with respect to the model parameters. By inputting training data in batches and iteratively training the model, a memory clipping detection model that meets the usage requirements is finally output, thus effectively ensuring the detection accuracy of the memory clipping detection model. In this embodiment, the initial long short-term memory network model is trained using historical 3D model images to obtain a usable memory clipping detection model. By fully learning various clipping patterns in historical data, the accuracy and efficiency of clipping detection can be effectively improved.

[0088] In an exemplary embodiment, the method further includes: when the clipping detection result indicates that clipping occurs in the three-dimensional model image, reconstructing the three-dimensional model in the three-dimensional model image to obtain a reconstructed three-dimensional model image; generating a comparison image of the effect of the three-dimensional model image and the reconstructed three-dimensional model image; and providing feedback on model reconstruction information based on the comparison image.

[0089] For example, 3D model reconstruction processing refers to remodeling the original 3D model image and modifying the detected clipping parts. After clipping detection is completed, the obtained clipping detection result includes two cases: the 3D model image has clipping and the 3D model image does not have clipping. When the clipping detection result indicates that the 3D model image has clipping, the clipping parts need to be repaired. At this time, the 3D model in the 3D model image can be reconstructed to obtain a reconstructed 3D model image. After the 3D reconstruction is completed, a comparison image of the 3D model image and the reconstructed 3D model image is generated, and the model reconstruction information is fed back to the model provider. This allows the model provider to understand the model quality through effect comparison. In this embodiment, by reconstructing the 3D model image when clipping is detected, the corresponding reconstruction information is fed back to help the model provider understand the clipping status of the model, thereby facilitating the rapid optimization of the model.

[0090] In one embodiment, this application can be specifically applied to a game testing scenario. After receiving a game testing request, the game testing server can first search the database for the target game corresponding to the game testing request, and simultaneously extract the game testing information contained in the game testing request, such as the specified test scenario, and the character models and motion models used in the test. Then, based on the test information in the game testing request, it loads the test scenario of the target game and runs the game in the test scenario. At the same time, it acquires images of the 3D model to be detected in the test scenario, and then uses the acquired 3D model images as input data to perform clipping detection, thereby realizing the connection between the game testing scenario and the clipping detection task, effectively improving the processing efficiency of the game testing task.

[0091] This application also provides an application scenario, which is illustrated by using the above-mentioned clipping detection method in this application scenario. The clipping detection method specifically includes:

[0092] When developing 3D games, users need to test the designed game scenes. Game testing includes multiple test items, such as integration testing, interface testing, and functional testing. For 3D games, checking for clipping issues in the test scenes is also an important test item. The clipping detection method described in this application can be used to detect clipping in game scenes. Identifying scenes where clipping occurs allows game developers to optimize the corresponding scenes and ensure the game's visual quality.

[0093] The overall process of mold penetration detection can be referred to Figure 5 As shown, the process includes four steps: data preparation, attention network training, memory integration, and clipping detection. The data preparation stage requires collecting resource models and scene data from the 3D game, including normal models and models with clipping issues. This data, after annotation, can be used for model training. Attention network training and memory integration are used to train the convolutional neural network model and the memory clipping detection model, respectively. By using annotated 3D model images as training samples, the convolutional neural network model and the memory clipping detection model are trained sequentially. The two trained models are then used to detect clipping in real-world game scenes.

[0094] The schematic diagram of the solution in this application can be referred to. Figure 6 As shown, in the data collection phase, a large number of 3D game resource models are first collected, including character models, object models, and environment models. These models need to cover various possible game scenarios to ensure that the detection system can be widely applied to the 3D game field. Then, the collected models and images are labeled using a graphics-assisted clipping detection metric to obtain training data that can be used for model training. During model training, the attention network is first trained using the obtained samples, which involves updating the parameters of the convolutional layers within the attention network. Predictive labels for the training data are extracted through the convolutional and attention layers. Then, by comparing the predicted labels with the actual clipping labels, the model loss is determined. The model parameters of the convolutional layers are then updated using a backpropagation scheme to complete the parameter update process for the attention network. Through repeated iterative training, convolutional layers and the attention network that can be used for prediction are obtained.

[0095] Then, a memory clipping detection model is trained based on this attention network. In this process, the model parameters of the attention network and convolutional layers need to be fixed, and attention feature maps of each training data are extracted through the attention network. These attention feature maps are then used as input data and fed into the initial long short-term memory network model to obtain the corresponding network prediction labels. By comparing the network prediction labels with the actual clipping labels, the model loss is determined, and the model parameters of the initial long short-term memory network model are updated. Through repeated iterative training, a memory clipping detection model that can be used for prediction is obtained.

[0096] After both the attention network and the memory-based clipping detection model are trained, efficient and accurate clipping detection can be performed using these two models. At this point, 3D model images from actual game scenarios and interactive scenes can be used as input data, sequentially fed into the attention network and the memory-based clipping detection model. This completes the clipping detection process within the scene. Specifically, for an interactive scene, a sequence of 3D model images can be extracted as input data. This input data is first fed into the attention network, undergoing convolution processing and weight adjustment of the attention mechanism to obtain a corresponding attention feature map sequence. This attention feature map sequence is then fed into the memory-based clipping detection model, which detects clipping regions. During this process, the memory-based clipping detection model combines learned historical clipping patterns with dynamic reasoning for annotation processing, such as... Figure 7 As shown, for each consecutive image in the attention feature map sequence, the region most likely to clip through is labeled. In the first image of the sequence, the detected clipping region is the region labeled 1, while in the second and third images, the detected clipping regions are the regions labeled 2. Then, combined with image-aided clipping detection metrics, the clipping region detection results are processed to obtain the clipping detection result for the 3D model image to be detected. If clipping occurs in the region most likely to clip through, it indicates that clipping has occurred in the corresponding game scene, requiring adjustments to both the scene and the model.

[0097] In one embodiment, such as Figure 8 As shown, the penetration detection method of this application includes:

[0098] Step 801: Obtain the sequence of 3D model images to be detected. Step 803: Perform convolution processing on the sequence of 3D model images to be detected to obtain convolutional feature maps. Step 805: Obtain the layer weights and biases of each convolutional layer. Step 807: Based on the layer weights and biases of each convolutional layer, sequentially perform convolution processing on the sequence of 3D model images to be detected for feature extraction to obtain a sequence of convolutional feature maps. Step 809: Activate the sequence of convolutional feature maps using an attention activation function to obtain an attention weight sequence. Step 811: Perform element-wise multiplication of the attention weight sequence and the sequence of convolutional feature maps to obtain an attention feature map sequence. Step 813: Adjust the feature matrix of each attention feature map in the attention feature map sequence using the model parameters of the memory-based clipping detection model to obtain a fused feature map sequence. Step 815: Activate the fused feature map sequence using an activation function to obtain the probability distribution matrix of the clipping regions in each fused feature map in the fused feature map sequence. Step 817: Analyze the probability distribution matrix of the clipping region to obtain the clipping region detection results for each 3D model image in the 3D model image sequence. Step 819: Combine the image-aided clipping detection index to perform clipping detection processing on the clipping region detection results to obtain the clipping detection results for the 3D model image to be detected.

[0099] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0100] Based on the same inventive concept, this application also provides a mold penetration detection device for implementing the mold penetration detection method described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the mold penetration detection device provided below can be found in the limitations of the mold penetration detection method above, and will not be repeated here.

[0101] In one exemplary embodiment, such as Figure 9 As shown, a penetration detection device is provided, comprising:

[0102] The convolution processing module 902 is used to perform convolution processing on the 3D model image to be detected to obtain a convolution feature map.

[0103] The attention adjustment module 904 is used to perform weighted processing on the feature matrix of the weighted convolutional feature map through the attention mechanism to obtain the attention feature map.

[0104] The memory detection module 906 is used to perform clipping region detection processing on the attention feature map based on the memory clipping detection model to obtain the clipping region detection results of the three-dimensional model image. The memory clipping detection model is trained on the initial long short-term memory network model through historical three-dimensional model images.

[0105] The clipping detection module 908 is used to combine graphic-aided clipping detection indicators to process the clipping detection results of the clipping area detection results, and obtain the clipping detection results of the three-dimensional model image to be detected.

[0106] In one embodiment, the convolution processing module 902 is specifically used to: obtain the layer weights and layer biases of each convolutional layer; and based on the layer weights and layer biases of each convolutional layer, sequentially input the three-dimensional model image to be detected into each convolutional layer for convolution processing to obtain the final output convolutional feature map.

[0107] In one embodiment, the attention adjustment module 904 is specifically used to: activate the convolutional feature map through the attention activation function to obtain attention weights; and perform element-wise multiplication of the weight matrix of the attention weights and the feature matrix of the convolutional feature map to obtain the adjusted attention feature map.

[0108] In one embodiment, the system further includes a model training module, configured to: acquire historical 3D model images; identify model bounding boxes in the historical 3D model images; determine the vertex set of the historical 3D model images based on the model bounding boxes; determine the 3D model volume in the historical 3D model images based on the vertex set; perform image-assisted clipping detection processing based on the vertex set and the 3D model volume to obtain an image-assisted clipping detection index; obtain clipping labels for each historical 3D model image based on the image-assisted clipping detection index; perform convolution processing on the historical 3D model images through convolutional layers to obtain convolutional feature maps for each historical 3D model image; identify the predicted labels of the convolutional feature maps; compare the clipping labels of the historical 3D model images with the predicted labels of the convolutional feature maps to determine the model loss; and update the parameters of the convolutional layers through backpropagation based on the model loss.

[0109] In one embodiment, the attention feature map includes an attention feature map sequence extracted from a 3D model image sequence. The memory detection module 906 is specifically used to: adjust the feature matrix of each attention feature map in the attention feature map sequence using the model parameters of the memory clipping detection model to obtain a fused feature map sequence; activate the fused feature map sequence using an activation function to obtain the probability distribution matrix of the clipping region for each fused feature map in the fused feature map sequence; and analyze the probability distribution matrix of the clipping region to obtain the clipping region detection result for each 3D model image in the 3D model image sequence.

[0110] In one embodiment, the memory detection module 906 is specifically used to: determine the product of the feature matrix and the probability distribution matrix of the attention feature map of each three-dimensional model image in the three-dimensional model image sequence, and obtain the clipping pattern recognition result of each three-dimensional model image; determine the region with the highest clipping probability of each three-dimensional model image based on the clipping pattern recognition result, and take the region with the highest clipping probability as the clipping region detection result of each three-dimensional model image.

[0111] In one embodiment, the memory detection module 906 is specifically used to: determine the input information of the memory unit of the memory penetration detection model at different time steps; determine the penetration detection restriction information based on the input information of the memory unit at different time steps; and perform penetration region recognition processing based on the penetration pattern recognition result and the penetration detection restriction information to determine the region with the highest penetration probability in each three-dimensional model image.

[0112] In one embodiment, the memory detection module 906 is specifically used to: determine the unit state of the memory unit of the memory detection model at different time steps; and adjust the feature information of each attention feature map in the attention feature map sequence based on the unit state of the memory unit at different time steps to obtain the input information of the memory unit of the memory detection model at different time steps.

[0113] In one embodiment, a memory model training module is further included, used for: acquiring historical 3D model images; performing clipping annotation processing on the historical 3D model images to obtain clipping labels for each historical 3D model image; predicting the clipping state of the historical 3D model images based on an initial long short-term memory network model to obtain network prediction labels; comparing the clipping labels and network prediction labels to obtain the memory network loss of the initial long short-term memory network model; obtaining model parameter gradient data based on the memory network loss; and updating the model parameters of the initial long short-term memory network model according to the model parameter gradient data to obtain a memory clipping detection model.

[0114] In one embodiment, the system further includes a model reconstruction module, configured to: reconstruct the 3D model in the 3D model image when the clipping detection result indicates that clipping has occurred in the 3D model image, to obtain a reconstructed 3D model image; generate a comparison image of the 3D model image and the reconstructed 3D model image; and provide feedback on model reconstruction information based on the comparison image.

[0115] In one embodiment, the system further includes an image acquisition module, configured to: acquire a game test request; locate the target game corresponding to the game test request; load the test scene of the target game based on the test information in the game test request; and acquire an image of the 3D model to be tested in the test scene.

[0116] Each module in the aforementioned mold penetration detection device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0117] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 10 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data related to mold detection. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a mold detection method.

[0118] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0119] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0120] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0121] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps in the above method embodiments.

[0122] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0123] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0124] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0125] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for detecting mold penetration, characterized in that, The method includes: The 3D model image to be detected is convolved to obtain a convolutional feature map; The feature matrix in the convolutional feature map is weighted using an attention mechanism to obtain an attention feature map. The attention feature map is processed by the memory clipping detection model to detect clipping regions, and the clipping region detection result of the three-dimensional model image is obtained. The memory clipping detection model is trained on the initial long short-term memory network model through historical three-dimensional model images. By combining the graphic-aided clipping detection index, the clipping region detection results are processed to obtain the clipping detection results of the three-dimensional model image to be detected.

2. The method according to claim 1, characterized in that, The convolutional feature map obtained by performing convolutional processing on the 3D model image to be detected includes: Obtain the layer weights and layer biases of each convolutional layer; Based on the layer weights and layer biases of each convolutional layer, the 3D model image to be detected is sequentially input into each convolutional layer for convolution processing to obtain the final output convolutional feature map.

3. The method according to claim 2, characterized in that, The step of weighting the feature matrix in the convolutional feature map using an attention mechanism to obtain the attention feature map includes: The convolutional feature map is activated using an attention activation function to obtain attention weights; The attention weight matrix and the feature matrix of the convolutional feature map are multiplied element-wise to obtain the adjusted attention feature map.

4. The method according to claim 3, characterized in that, The method further includes: Obtain historical 3D model images; Identify the model bounding box of the historical 3D model image; Based on the model bounding box, determine the vertex set of the historical 3D model image; The volume of the 3D model in the historical 3D model image is determined based on the vertex set of the historical 3D model image; Based on the vertex set and the volume of the 3D model, a graphics-assisted clipping detection process is performed to obtain a graphics-assisted clipping detection index. Based on the aforementioned graphic-assisted clipping detection index, clipping labels are obtained for each historical 3D model image; The historical 3D model images are convolved by convolutional layers to obtain convolutional feature maps for each historical 3D model image; the predicted labels of the convolutional feature maps are identified; the clipping labels of the historical 3D model images and the predicted labels of the convolutional feature maps are compared to determine the model loss; based on the model loss, the parameters of the convolutional layers are updated through backpropagation.

5. The method according to any one of claims 1 to 4, characterized in that, The attention feature map includes a sequence of attention feature maps, which is extracted from a sequence of three-dimensional model images. The method of performing clipping region detection processing on the attention feature map based on the memory clipping detection model to obtain the clipping region detection result of the 3D model image includes: By adjusting the feature matrix of each attention feature map in the attention feature map sequence using the model parameters of the memory penetration detection model, a fused feature map sequence is obtained. The probability distribution matrix of the clipping region of each fusion feature map in the fusion feature map sequence is obtained by activating the fusion feature map sequence with an activation function. The probability distribution matrix of the clipping region is analyzed to obtain the clipping region detection results for each 3D model image in the 3D model image sequence.

6. The method according to claim 5, characterized in that, The step of analyzing the probability distribution matrix of the clipping region to obtain the clipping region detection results for each 3D model image in the 3D model image sequence includes: In the sequence of three-dimensional model images, the product of the feature matrix and the probability distribution matrix of the attention feature map of each three-dimensional model image is determined to obtain the pattern recognition result of each three-dimensional model image. Based on the pattern recognition results, the region with the highest probability of pattern penetration in each 3D model image is determined, and the region with the highest probability of pattern penetration is taken as the pattern penetration region detection result of each 3D model image.

7. The method according to claim 6, characterized in that, The regions with the highest probability of clipping in each 3D model image, determined based on the clipping pattern recognition results, include: Determine the input information of the memory unit of the memory penetration detection model at different time steps; The penetration detection limitation information is determined based on the input information of the memory unit at different time steps; Based on the pattern recognition results and the pattern detection restriction information, pattern recognition processing is performed to determine the region with the highest pattern recognition probability in each 3D model image.

8. The method according to claim 7, characterized in that, The input information of the memory unit of the memory penetration detection model at different time steps includes: Determine the unit state of the memory unit in the memory penetration detection model at different time steps; Based on the unit state of the memory unit at different time steps, the feature information of each attention feature map in the attention feature map sequence is adjusted to obtain the input information of the memory unit of the memory penetration detection model at different time steps.

9. The method according to claim 1, characterized in that, The method further includes: Obtain historical 3D model images; The historical 3D model images are subjected to clipping annotation processing to obtain clipping labels for each historical 3D model image; Based on the initial long short-term memory network model, the clipping state of the historical 3D model image is predicted to obtain the network predicted label; By comparing the wear-through label and the network prediction label, the memory network loss of the initial long short-term memory network model is obtained; The gradient data of the model parameters are obtained based on the memory network loss. Based on the gradient data of the model parameters, the initial long short-term memory network model is updated to obtain the memory penetration detection model.

10. The method according to any one of claims 1 to 9, characterized in that, The method further includes: When the clipping detection result indicates that the three-dimensional model image has clipping, the three-dimensional model in the three-dimensional model image is reconstructed to obtain a reconstructed three-dimensional model image. A comparison image showing the effects of generating the 3D model image and reconstructing the 3D model image; Feedback on model reconstruction information is based on the aforementioned effect comparison chart.

11. The method according to any one of claims 1 to 10, characterized in that, Before performing convolution processing on the 3D model image to be detected to obtain the convolutional feature map, the process also includes: Get game test request; Find the target game corresponding to the game test request; Based on the test information in the game test request, load the test scenario of the target game. Images of the 3D model to be tested are acquired in the test scenario.

12. A device for detecting mold penetration, characterized in that, The device includes: The convolution processing module is used to perform convolution processing on the 3D model image to be detected to obtain a convolution feature map; The attention adjustment module is used to weight the feature matrix in the convolutional feature map through an attention mechanism to obtain an attention feature map; The memory detection module is used to perform clipping region detection processing on the attention feature map based on the memory clipping detection model to obtain the clipping region detection result of the three-dimensional model image. The memory clipping detection model is obtained by training an initial long short-term memory network model with historical three-dimensional model images. The clipping detection module is used to combine graphic-aided clipping detection indicators to perform clipping detection processing on the clipping region detection results, so as to obtain the clipping detection result of the three-dimensional model image to be detected.

13. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 11.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 11.

15. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 11.

Images