Image Data Input Detection Method and System for Neural Network Vision System

Abstract

The present invention discloses an image data input detection method and system for a neural network visual system. Given a neural network model and its training image data set, the training image data set is input into the neural network model and intermediate results are collected to obtain implicit features of the neural network; a Gaussian mixture model is used to fit the intermediate results to obtain model parameters, and the path frequency calculation probability of the training image data set is collected; the image data to be tested is input into the neural network model, and the intermediate results are collected according to the method of step one; the Gaussian mixture model in step two is used to calculate the generation probability and inter-layer transfer probability of the intermediate results, and a joint probability estimation model is used to perform fast probability estimation to verify whether the input image data to be tested is valid.

Description

Technical Field

[0001] The present invention relates to an image data input detection method and system for a neural network vision system, belonging to the technical fields of input verification of deep neural network models, intelligent software quality assurance, image data processing, etc. Background Art

[0002] In recent years, software systems based on deep neural networks have been widely applied in various fields of production and life, bringing great convenience to people's lives. An important field of application of deep learning models is machine vision, which in turn includes sub-tasks such as image classification, object detection, and semantic segmentation. Although software systems based on deep neural network technology have achieved good performance in these tasks, neural network models often do not reach 100% accuracy. Moreover, the prediction process of deep neural network models is difficult to interpret, so it is impossible to judge whether the prediction results of vision models are correct through simple methods.

[0003] Vision systems based on neural network models have been applied in fields with high security requirements such as face recognition and video surveillance, and incorrect prediction results may bring serious consequences. In order to ensure the security of machine vision systems based on deep neural networks, some abnormal input detection technologies for image data have gradually emerged. Such methods can estimate the probability of causing abnormal predictions of the neural network vision system for a given image data sample. If the estimated abnormal prediction probability is greater than a predetermined threshold, the sample is rejected and a warning is given to ensure the reliable operation of the machine vision intelligent system.

[0004] The applicability and reliability of current existing image input detection technologies still have some limitations. Some methods are discriminative-based, modeling the input image detection problem as a binary classification problem of normal data and abnormal data. The construction process of these methods requires the participation of some abnormal input images, and obtaining and annotating picture samples increases the cost of application. In addition, due to the participation of abnormal data, discriminative methods may cause overfitting to a specific data distribution, and it is difficult to ensure the same generalization performance for image data from different distributions. Some other methods have the characteristics of being model-independent. These methods use specific strategies to retrain the deep neural network model, which not only increases the cost and difficulty of applying the detection technology, but also the retrained model often leads to a decline in performance, which is unacceptable in many application scenarios. Summary of the Invention

[0005] Objective of the Invention: Aiming at the problems and deficiencies in the prior art, the present invention provides an image data input detection method and system based on intermediate layer joint modeling for a neural network vision system. The present invention can be applied to common machine vision intelligent systems based on deep neural networks in real life, collect and analyze the intermediate results generated during the inference process of the deep neural network system, and detect the input image sample data that the deep neural network model cannot correctly predict.

[0006] The advantages of the present invention are as follows: First, the usage scenario is not strictly limited by the complexity of the neural network model and the usage scenario, especially applicable in scenarios where abnormal image data is difficult to collect. Second, the method has a high accuracy in detecting the validity of input image data, can effectively distinguish between valid and invalid input images, and thus avoid meaningless prediction results. Third, the method has a small time cost for detecting input images, can meet the requirement of real-time verification of correctness, and can be deployed in a running machine vision model for input detection.

[0007] Technical Solution: An image data input detection method for a neural network vision system includes the following steps:

[0008] Step 1: Neural network implicit feature extraction. Given a neural network model and its training image dataset, input the training image dataset into the neural network model and collect the intermediate results to obtain the neural network implicit features.

[0009] First, according to the given neural network model to be tested, rewrite its forward propagation sub-process to export intermediate results during the inference process; then, use the given training image dataset, input the image dataset into the neural network model and collect the implicit features of the intermediate layer.

[0010] Step 2: Representation space joint probability modeling. Use a Gaussian mixture model to fit the intermediate results, obtain the model parameters, and collect the path frequencies of the training image dataset to calculate the probability.

[0011] Building a joint probability model requires two steps. First, use the intermediate layer implicit features generated in Step 1 to establish a generative model based on a probability graph. Then, map the intermediate layer features to a discrete space to obtain the transition probability of the image data sample between adjacent intermediate layers;

[0012] Step 3: Joint probability estimation model. Input the image data to be tested into the neural network model, and collect the intermediate results in the same way as in Step 1. Use the Gaussian mixture model in Step 2 to calculate the generation probability and inter-layer transfer probability of the intermediate results, and use the joint probability estimation model for rapid probability estimation to verify whether the input image data to be tested is valid.

[0013] The neural network is a deep neural network, which refers to a machine learning model that uses neurons to form hierarchical connections for image data feature extraction and prediction, including an input layer, hidden layers, and an output layer. Each layer contains a large number of neurons, and the neurons are interconnected between layers. Image data is transmitted from the input layer to the output layer, and a prediction result is output, such as the category to which the image data belongs; the intermediate layer result is the hidden feature data output by the neurons in the hidden layer between the input layer and the output layer of the neural network; the neuron is a structure that performs arithmetic operations on the input data using built-in functions, etc., and outputs the arithmetic result; the input image data refers to a single or a batch of image data samples that conform to the input format of the deep neural network model.

[0014] In step 1, the neural network model to be tested refers to a pre-trained model that contains complete model information (such as model architecture, neuron parameters, etc.) and has complete operation permissions (such as collecting model intermediate results, modifying model parameters, etc.).

[0015] In step 1, passing the training image data set into the neural network model and collecting intermediate results means covering the forward propagation process of the neural network to make the neural network retain the hidden features output by the neurons in the intermediate layer, and collecting the intermediate layer hidden features of the training image data set. Since the neurons in the neural network model are arranged layer by layer, the intermediate results can also be divided into multiple layers accordingly. In theory, all the intermediate layer hidden features can be collected and utilized, but due to actual hardware limitations, and the results obtained by sampling have a relatively high accuracy, it is not necessary to use all the intermediate results. In actual use, a uniform sampling strategy or a sampling strategy according to the model structure can be adopted to sample the intermediate results. In practical applications, if the amount of intermediate layer feature data is still large, redundant information can be removed by dimensionality reduction. For example, global pooling can achieve good results, and the PCA dimensionality reduction algorithm can also be used in conjunction with the present invention.

[0016] In step 2, the Gaussian mixture model refers to a multi-dimensional Gaussian mixture clustering algorithm, which has a fitting effect on multi-dimensional complex data distributions. The common Expectation-Maximization (EM) algorithm can be used to find the approximate solution of the Gaussian mixture model, obtain several Gaussian components and corresponding weights, and then obtain the sample probability by a generative method. These model parameters are the basic materials for the subsequent framework for evaluating the probability of new samples.

[0017] In step 2, the so-called joint probability modeling of the representation space refers to the joint probability modeling of the hidden features in the representation space of the deep neural network, and its purpose is to model the inference process of the neural network. The representation space refers to the data space of the intermediate layer hidden features, which contains the intermediate representation features of the model input data.

[0018] In the second step, the fitting of the intermediate layer latent features respectively is to establish a probability distribution model of the intermediate layer latent features using a Gaussian mixture model (GMM). The expectation-maximization (EM) algorithm is used to establish a Gaussian mixture model on the output of the i-th intermediate layer, and the parameters Θ i of K i Gaussian components and the weight Φ i of each Gaussian component are obtained. The parameters Φ i and Θ i in the probabilistic graphical model are the basic materials for evaluating the abnormal probability of image data.

[0019] In the second step, for the collection of the training image dataset, the path frequency is calculated for probability, and the probability calculation method is as follows: all samples in the training set generate path data {(z i , z i-1 ) m |m ∈ Z m}, where m is the size of the training set. The probability p(z i-1 |z i ) of the transition from z i to z i-1 is statistically obtained, providing basic parameters for subsequent evaluation of sample probabilities.

[0020] In the second step, the transition probability between adjacent intermediate layers refers to the transition probability between adjacent layers of the discrete component z i corresponding to the intermediate layer feature x i on the training image dataset; the discrete component z i refers to the cluster to which the input image data belongs as z i according to the clustering result of the GMM of the i-th layer for the intermediate layer feature x i . Further, the probability P(z i-1 |z i ) of the transition from z i to z i-1 is calculated on the training image dataset, that is, the transition probability from the (i - 1)-th layer to the i-th layer. The possible values of z i are K i types, and the transition probability from the (i - 1)-th layer to the i-th layer is a matrix of size K i-1 ×K i .

[0021] In the third step, the joint probability estimation model refers to estimating the joint probability of the intermediate layer output sequence {x 1 , x 2 , …, x m} through a probabilistic graphical model. The joint probability P(x 1 , x 2 , …, x m) has an exponential growth in terms of m for its time complexity. Therefore, the present invention uses a fast forward algorithm based on dynamic programming. Let

[0022] α i (z i )≡P(z i ,x 1 ,x 2 ,…,x i ),

[0023] Then α i (·) can be generated through the following recursive process:

[0024]

[0025] where K i-1 represents the number of Gaussian components in the (i - 1)-th layer, P(x i |z i ) is given by the GMM of the i-th layer, and P(z i |z i-1 ) is the transition probability. Thus, we obtain the joint probability of the intermediate layer output with a linear time complexity.

[0026] In the third step, verifying whether the input image data to be tested is valid means judging whether the joint probability of the image data is greater than a preset threshold; the threshold is a value between 0 and 1. The closer it is to 0, the more it tends to the precision rate of the method, and the closer it is to 1, the more it tends to the recall rate of the method; a feasible strategy in practice is to set a threshold that makes most image data (such as 90%) normal according to the joint probability of the training image data set.

[0027] An image data input detection system for a neural network vision system, comprising:

[0028] Neural network hidden feature extractor: Given a neural network model and its training image data set, input the training image data set into the neural network model and collect the intermediate results to obtain the neural network hidden features.

[0029] First, according to the given neural network model to be tested, rewrite its forward propagation sub-process so that it can export intermediate results during the inference process; then, use the given training image data set, input the image data set into the neural network model and collect the hidden features of the intermediate layer.

[0030] Spatial joint probability modeling tool: Use a Gaussian mixture model to fit the intermediate results, obtain the model parameters, and collect the probability calculated by the path frequency of the training image data set.

[0031] Building a joint probability model requires two steps. First, a generative model based on a probabilistic graph is established using the intermediate hidden features generated by a neural network hidden feature extractor. Then, the intermediate features are mapped to a discrete space to obtain the transition probability of image data samples between adjacent intermediate layers.

[0032] Joint probability estimation model: Input the image data to be tested into the neural network model, and use the neural network hidden feature extractor to collect its intermediate results. Calculate the generation probability and inter-layer transition probability of the intermediate results using the Gaussian mixture model in the representation space joint probability modeling tool, and use the joint probability estimation model for rapid probability estimation to verify whether the input image data to be tested is valid.

[0033] A computer device, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the above computer program, it implements the image data input detection method for a neural network visual system as described above.

[0034] A computer-readable storage medium stores a computer program for executing the image data input detection method for a neural network visual system as described above.

[0035] Beneficial effects: The present invention can make up for the deficiency that ordinary neural network visual models are difficult to identify abnormal image data, thus ensuring the safety of intelligent systems. Compared with the existing neural network model input data verification technology, the present invention does not require training with abnormal data and is not prone to overfitting to specific data. The present invention can efficiently detect and evaluate the validity of input image data, and utilize the evaluated validity to realize real-time screening of test images, improving the actual deployment effect of the neural visual system. Description of the Drawings

[0036] Figure 1 It is the schematic diagram of the method principle of the embodiment of the present invention;

[0037] Figure 2 It is the schematic diagram of the neural network hidden feature extractor of the embodiment of the present invention;

[0038] Figure 3 It is the flow chart of the representation space joint probability modeling tool of the embodiment of the present invention;

[0039] Figure 4 It is the functional schematic diagram of the joint probability estimation model of the embodiment of the present invention;

[0040] Figure 5 It is the schematic diagram of the probabilistic graph model of the embodiment of the present invention. Detailed Embodiments

[0041] The present invention will be further illustrated below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. After reading the present invention, those skilled in the art's various equivalent modifications of the present invention all fall within the scope defined by the appended claims of this application.

[0042] As Figure 1 shown, the image data input detection method for a neural network visual system consists of two stages: training and deployment. In the training stage, first, using a hidden feature extractor, the training image data is input into the neural network model to generate the hidden features produced by each intermediate layer. Subsequently, a Gaussian mixture model is trained on the intermediate layer features using a representation space joint modeling tool, and the transition probability between adjacent layers is calculated using the clustering results of the Gaussian mixture model. In the deployment stage, for any image data to be tested, first, it is input into the neural network to obtain the intermediate layer features, and then the joint probability of the input intermediate layer features is calculated using the joint probability estimation model. Finally, the joint probability value is compared with a predetermined threshold to decide whether to accept or reject the image data to be tested. The entire framework includes three modules corresponding to three steps: the neural network hidden feature extractor, the representation space joint probability modeling tool, and the joint probability estimation model.

[0043] An image data input detection system for a neural network visual system includes:

[0044] Step 1: Given a neural network model and its training image data set, input the training image data set into the neural network model and collect the intermediate results to obtain the neural network hidden features.

[0045] As Figure 2 shown, covering the forward propagation process of the neural network visual model to retain the hidden features output by the intermediate layer neurons, inputting the training image data set into the neural network, and extracting the hidden features {x 1 , x 2 , …, x m} generated by each intermediate layer of the model, where m is the number of intermediate layers. The intermediate features of these image data can theoretically be all collected and utilized, but due to actual hardware limitations, and the results obtained by sampling have a relatively high accuracy, it is not necessary to use all the intermediate results. In actual use, a uniform sampling strategy or a sampling strategy according to the model structure can be adopted for sampling. In actual applications, if the amount of intermediate layer feature data is still large, redundant information in the intermediate layer output of the image data can be removed by dimensionality reduction. For example, global pooling can obtain good results, and the PCA dimensionality reduction algorithm can also be used in conjunction with the present invention. Save the processed intermediate layer results for subsequent training.

[0046] In this step, it is required that the neural network vision system to be tested contains complete model information (such as model architecture, neuron parameters, etc.) and has complete operation permissions (such as collecting intermediate model results, modifying model parameters, etc.). It is not applicable to black-box models that only provide API services.

[0047] Step 2: Use joint probability modeling of the representation space to model the inference process of the neural network vision model on the training set image data.

[0048] As Figure 3 shown, train a Gaussian mixture model layer by layer on the intermediate layer data obtained in Step 1. The Gaussian mixture models of each layer can reasonably select parameters according to the scale of the image data and the computing power of the device, such as the number of Gaussian components. Retain the Gaussian mixture models on the intermediate layer data of each layer, and perform a clustering operation on the training set image data to obtain the intermediate layer output clustering result z of the image data i and calculate the transition probability P(z i |z i-1 ) based on this result for saving and using with the image data to be tested online.

[0049] Joint probability modeling of the representation space refers to the joint probability modeling of the latent features in the representation space of a deep neural network, aiming to model the inference process of the neural vision network on image data. The representation space refers to the data space of the intermediate latent features, including the intermediate representation features of the model input data. This method uses a probability graph model as Figure 5 shown to represent the intermediate layer joint probability. The intermediate layers of the neural network correspond to the layers in the probability graph model, where x i is the output of the i-th intermediate layer of the neural network model, z i is the hidden state of the i-th layer of the probability graph model, and Φ i and Θ i are the parameters of the method.

[0050] Fitting the intermediate layer latent features separately is to use a Gaussian mixture model (GMM) to establish a probability distribution model of the intermediate layer latent features. Use the expectation-maximization (EM) algorithm to establish a Gaussian mixture model on the output of the i-th intermediate layer to obtain the parameters Θ i of K i Gaussian components and the weight Φ i of each Gaussian component. The parameters Φ i and Θ i in the probability graph model are the basic materials for evaluating the anomaly probability of image data.

[0051] The transition probability between adjacent intermediate layers refers to the discrete component z i corresponding to the intermediate layer feature x iTransition probability between adjacent layers; discrete component z i Refers to the intermediate layer feature x of the input data in the i-th layer according to the clustering result of the GMM in the i-th layer i The cluster to which it belongs is z i . Calculate z on the training data set i-1 to z i The probability P(z i |z i-1 ) of the transition, that is, the transition probability from the (i - 1)-th layer to the i-th layer. z i Can take K i values, and the transition probability from the (i - 1)-th layer to the i-th layer is a matrix of size K i-1 ×K i .

[0052] Step 3: Use the joint probability estimation model to analyze the effectiveness of the image data to be tested and report.

[0053] As Figure 4 shown, input one or a batch of image data to be tested into the neural network model, perform the inference process and extract its intermediate results, perform dimensionality reduction processing on the intermediate layer results and save them. Note that the intermediate layer of the model selected in this step should be consistent with the operation in Step 1, and the dimensionality reduction operation should also be consistent. Input the intermediate layer output data into the Gaussian mixture model corresponding to this intermediate layer to obtain the probability P(x i |z i ) of the intermediate layer results on each Gaussian component, where z i represents the Gaussian component of the image data in the i-th layer. Estimate the joint probability of the intermediate layer output sequence {x 1 ,x 2 ,…,x m} through the probabilistic graphical model. Directly calculating the joint probability P(x 1 ,x 2 ,…,x m ) has an exponential growth in time complexity with respect to m, so a fast forward algorithm based on dynamic programming is used. Let

[0054] α i (z i )≡P(z i ,x 1 ,x 2 ,…,x i ),

[0055] Then α i (·) can be generated through the following recursive process:

[0056]

[0057] Finally, obtain the intermediate layer feature sequence {x 1,x 2 ,…,x m The joint probability of {...}. Determine whether the joint probability of the image data to be tested is greater than a preset threshold; the threshold is a value between 0 and 1. The closer it is to 0, the more it tends to the precision rate of the method, and the closer it is to 1, the more it tends to the recall rate of the method. A practical strategy is to set a threshold that makes most image samples (such as 90%) normal according to the joint probability of the training image data set.

[0058] An image data input detection system for a neural network vision system, including:

[0059] Neural network implicit feature extractor: Given a neural network model and its training image data set, input the training image data set into the neural network model and collect intermediate results to obtain neural network implicit features.

[0060] First, according to the given neural network model to be tested, rewrite its forward propagation sub-process to export intermediate results during the inference process; then, use the given training image data set, input the image data set into the neural network model and collect the implicit features of the intermediate layer.

[0061] Representation space joint probability modeling tool: Use a Gaussian mixture model to fit the intermediate results, obtain model parameters, and collect the path frequencies of the training image data set to calculate probabilities.

[0062] Building a joint probability model requires two steps. First, use the intermediate layer implicit features generated by the neural network implicit feature extractor to establish a generative model based on a probability graph. Then map the intermediate layer features to a discrete space to obtain the transition probability of the image data sample between adjacent intermediate layers;

[0063] Joint probability estimation model; Input the image data to be tested into the neural network model, and use the neural network implicit feature extractor to collect its intermediate results. Use the Gaussian mixture model in the representation space joint probability modeling tool to calculate the generation probability and inter-layer transition probability of the intermediate results, and use the joint probability estimation model for rapid probability estimation to verify whether the input image data to be tested is valid.

[0064] Obviously, those skilled in the art should understand that each step of the above image data input detection method for a neural network vision system or each module of the image data input detection system for a neural network vision system can be implemented by a general-purpose computing device. They can be concentrated on a single computing device or distributed on a network composed of multiple computing devices. Optionally, they can be implemented by program code executable by the computing device. Thus, they can be stored in a storage device and executed by the computing device. And in some cases, the steps shown or described can be executed in a different order than here, or they can be separately fabricated into individual integrated circuit modules, or multiple modules or steps among them can be fabricated into a single integrated circuit module to implement. In this way, the embodiments of the present invention are not limited to any specific combination of hardware and software.

Claims

1. An image data input detection method for a neural network visual system, characterized in that, it includes the following steps: Step 1: Neural network implicit feature extraction; Given a neural network model and its training image dataset, input the training image dataset into the neural network model and collect the intermediate results to obtain the neural network implicit features; Step 2: Representation space joint probability modeling; Use the Gaussian mixture model to fit the intermediate results, obtain the model parameters, and collect the path frequencies of the training image dataset to calculate the probability; Step 3: Joint probability estimation model; Input the image data to be tested into the neural network model, and collect the intermediate results in the same way as in Step 1; Use the Gaussian mixture model in Step 2 to calculate the generation probability and inter-layer transition probability of the intermediate results, and use the joint probability estimation model for fast probability estimation to verify whether the input image data to be tested is valid; In the said Step 2, constructing the joint probability model requires two steps. First, establish a generative model based on a probability graph using the intermediate layer implicit features generated in Step 1; Then map the intermediate layer features to the discrete space to obtain the transition probability of the image data samples between adjacent intermediate layers; The said representation space refers to the data space of the intermediate layer implicit features, which contains the intermediate representation features of the model input data; The respective fitting of the intermediate layer implicit features is to use a Gaussian mixture model to establish a probability distribution model for the intermediate layer implicit features; the Gaussian mixture model is established on the output of the u-th intermediate layer using the expectation-maximization algorithm to obtain the parameters Θ 1 of K i Gaussian components and the weight Φ i of each Gaussian component; the parameters Φ i and Θ i in the probabilistic graphical model are the basic materials for evaluating the anomaly probability of image data.

2. The image data input detection method for a neural network visual system as claimed in claim 1, characterized in that, the said neural network is a deep neural network, which refers to a machine learning model for image data feature extraction and prediction formed by hierarchical connection of neurons, including an input layer, an implicit layer, and an output layer; Each layer contains neurons, and the neurons are connected to each other between the layers, and the image data is transmitted from the input layer to the output layer to output the prediction result; The said intermediate layer result is the implicit feature data output by the neurons in the implicit layer between the input layer and the output layer of the neural network; The said neuron is a structure that performs arithmetic operations on the neuron input using built-in functions, etc., and outputs the operation result; The said input image data refers to a single or a batch of image data samples that conform to the input format of the deep neural network model.

3. The image data input detection method for a neural network visual system as claimed in claim 1, characterized in that, in the said Step 1, according to the given neural network model to be tested, rewrite the forward propagation sub-process of the neural network model so that the neural network model can export intermediate results during the inference process; Then use the given training image dataset, input the image dataset into the neural network model and collect the intermediate layer implicit features.

4. The image data input detection method for a neural network visual system as claimed in claim 1, characterized in that, adopt the uniform sampling strategy or the sampling strategy according to the model structure to sample the intermediate results.

5. The image data input detection method for a neural network visual system as claimed in claim 1, characterized in that, In the second step, the transition probability between adjacent intermediate layers refers to the transition probability between the discrete components z i corresponding to the intermediate layer feature x i between adjacent layers; the discrete component z i refers to the cluster to which the input image data belongs among the intermediate layer features x i at the i-th layer according to the clustering result of the GMM at the i-th layer, and the cluster is z i ; calculate the probability of the transition from z i-1 to z i on the training image dataset P(z i |z i-1 ), that is, the transition probability from the (i - 1)-th layer to the i-th layer; z i can take K i values. The transition probability from the (i - 1)-th layer to the i-th layer is a matrix of size K i-1 ×K i .

6. The image data input detection method for a neural network visual system as claimed in claim 1, characterized in that, In the third step, the joint probability estimation model is used to estimate the joint probability of the intermediate layer output sequence {x 1 , x 2 , …, x m} through a probabilistic graphical model. The time complexity of directly calculating the joint probability P(x 1 , x 2 , …, x m ) grows exponentially with respect to m. Therefore, a fast forward algorithm based on dynamic programming is used; let α i (z i )≡P(z i ,x 1 ,x 2 ,…,x i ), Then α i (·) is generated through the following recursive process: where K i-1 represents the number of Gaussian components in the (i - 1)-th layer, and P(x i |z i ) is given by the GMM in the i-th layer, and P(z i |z i-1 ) is the transition probability.

7. A system for implementing an image data input detection method for a neural network vision system according to any one of claims 1-6, characterized in that, comprising: Neural network hidden feature extractor: Given a neural network model and its training image dataset, input the training image dataset into the neural network model and collect the intermediate results to obtain the neural network hidden features; First, according to the given neural network model to be tested, rewrite its forward propagation sub-process to export intermediate results during the inference process; then, use the given training image dataset, input the image dataset into the neural network model and collect the hidden features of the intermediate layer; Representation space joint probability modeling tool: Use a Gaussian mixture model to fit the intermediate results, obtain the model parameters, and collect the path frequencies of the training image dataset to calculate the probability; Constructing a joint probability model requires two steps. First, use the hidden features of the intermediate layer generated by the neural network hidden feature extractor to establish a generative model based on a probability graph; then map the intermediate layer features to a discrete space to obtain the transition probability of the image data between adjacent intermediate layers; Joint probability estimation model; Input the image data to be tested into the neural network model, and collect the intermediate results according to the neural network hidden feature extractor; Use the Gaussian mixture model in the representation space joint probability modeling tool to calculate the generation probability and inter-layer transition probability of the intermediate results, and use the joint probability estimation model for rapid probability estimation to verify whether the input image data to be tested is valid.

8. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the image data input detection method for a neural network vision system according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program for executing the image data input detection method for a neural network vision system according to any one of claims 1-6.

Images