Article classification detection method

Abstract

The invention discloses an article classification detection method, which belongs to the technical field of automatic classification and comprises the following steps of: first-stage detection: extracting a potential area; and second-stage detection: filtering potential areas. The object detection method aims at solving the problem that an existing object detection method is too many in object area image parameters and prone to over-fitting. According to the method, the GoogleNet network model with the Inception module is adopted, the sparsity of a network structure can be kept, the high calculation performance of a dense matrix can be utilized, the problem that overfitting is likely to be caused while the depth and width of the network are increased is solved, and interference of natural factors (such as rainwater and light) inside or outside the environment can be removed from a shot area image. Meanwhile, the method fully combines the advantages of the GoogleNet and the GAN, takes the GoogleNet as a generator of the GAN, and effectively improves the capability of the GoogleNet for extracting the environment image features through the confrontation game of the generator and a discriminator.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic classification, and particularly relates to an article classification detection method. BACKGROUND

[0002] At present, there are various types of articles, including different materials, shapes, sizes and functions. Such diversity increases the complexity of detecting article classification, because each article may require different processing methods and standards. And the automatic classification system may not be able to accurately identify some special material or shape of the article, resulting in classification errors or omissions.

[0003] With the rapid development of artificial intelligence technology, especially the convolutional neural network in deep learning in image processing, combined with related hardware, it is possible to solve this problem. For example, the GoogleNet network proposes a model that can maintain the sparsity of the network structure and utilize the high computing performance of the dense matrix, creatively proposes the Inception structure, solves the problem of overfitting caused by increasing the depth and width of the network, and gets a lot of attention in the field of artificial intelligence.

[0004] However, the traditional GoogleNet network still has some problems in article recognition. For example, using a deeper network will result in too many parameters, and the limited training data set is prone to overfitting, and the cost of creating a data set is very large, and the deeper the network, the more likely to appear gradient dispersion problem, it is difficult to optimize the model. Therefore, in the case of a small number of training samples, GoogleNet is difficult to fully extract the effective information in the room for article feature extraction, and is easily disturbed by environmental factors (rain, snow, light) and even image clarity.

[0005] Based on this, the present application designs an article classification detection method to solve the above problems. SUMMARY

[0006] In view of the above-mentioned shortcomings of the prior art, the present application provides an article classification detection method. Based on

[0007] The indoor environment article recognition method of GoogleNet-Gan, and fully combines the advantages of GoogleNet and GAN, which will

[0008] GoogleNet as the generator of GAN, and through the confrontation game of the generator and the discriminator, effectively improves the ability of GoogleNet to extract environmental image features, thereby ensuring the detection accuracy.

[0009] To achieve the above purpose, the present application realizes the following technical scheme:

[0010] An article classification detection method, comprising the following steps:

[0011] First stage detection: potential region extraction; specifically comprising the following steps:

[0012] Step a, using a deep learning network method or an SSIM method to extract the features of the test image;

[0013] Step b, calculate the similarity distance of each pixel point using the cosine distance, and if the similarity distance is less than the threshold, set it as a candidate point;

[0014] Step c, merge the candidate points, calculate the circumscribed rectangle frame that can completely contain all points for each group of merged candidate points, obtain the rectangle frame coordinates, and obtain the candidate region;

[0015] Second stage detection: potential region filtering; comprising the following steps:

[0016] Step d, for each candidate region, extract its neat environment template image and test image;

[0017] Step e, use the learned similarity network to perform similarity discrimination on the neat environment template image and the test image, and if the similarity is greater than the threshold, it is considered to be interference, and the candidate region is filtered;

[0018] Step f, output the final rectangular frame region, which is the inconsistent region.

[0019] Further, the deep learning network method in step a adopts a deep learning network method based on GoogleNet-GAN.

[0020] Further, the deep learning network method based on GoogleNet-GAN comprises the following steps:

[0021] Step 1, data preprocessing: taking neat environment template images under different light conditions with rain, the images are processed in grayscale to obtain a 200*300 matrix, denoted as x n , n represents the nth image;

[0022] Step 2, model construction: using a GAN network as the main framework, composed of a generator and a discriminator, the generator G uses a GoogleNet network model; during the training of the generator G and the discriminator D in the GAN network, the parameters of one side are fixed and the parameters of the other side are updated through alternating iteration.

[0023] Further, the model construction process of step 2 comprises the following steps:

[0024] Step 2.1, the generator G uses the GoogleNet network model, which includes the Inception module; the GoogleNet network model traverses the image through the superposition of convolutional layers and pooling layers to extract abstract high-resolution images.

[0025] Dimensional features:

[0026]

[0027] Where: O is the side length of the output feature map, I is the side length of the grayscale value matrix of the neat environment template image, F is the side length of the convolution kernel, P is the zero padding value, and S is the convolution kernel movement step size.

[0028] The Inception module of the GoogleNet network model performs the following steps:

[0029] Step 2.1.1, the convolution layer performs convolution calculation;

[0030] Step 2.1.2: Introduce a batch normalization layer before the activation function of each layer. Perform batch normalization on the tensor after convolution so that the variables entering the activation function are distributed at the maximum gradient position with mean 0 and unit variance. Then perform an affine transformation and pass it to the activation function.

[0031] Step 2.1.3. Add an activation function after the batch normalization layer to perform a nonlinear transformation on the linear function passed from the upper layer.

[0032] Step 2.1.4: Construct a loss function to guide the network to train a model that can ignore rain and light interference;

[0033] Step 2.2: The discriminator D uses the Unet network model to downsample the output from 200*300 to 0 / 1.

[0034] Step 2.3, set the loss function;

[0035] Step 2.4: Perform model training.

[0036] Furthermore, step 2.1.1, the convolution layer performs convolution calculation: input template image X p，q , p∈[0,4], q∈[0,4], X is the input image data, p and q are subscripts used to index specific locations or regions in the image, and the convolution kernel is w m，n , m∈[0,1], n∈[0,1], m is the size of the convolution kernel in the height direction, n is the size of the convolution kernel in the width direction, then the convolution calculation formula is:

[0037]

[0038] wherein y m，n is the value at the position with coordinates (m, n) on the output feature map after convolution calculation, g(.) is an activation function, "*" is a convolution operation, and b is a convolution bias.

[0039] Further, step 2.1.3, an activation function is added after the batch normalization layer to perform nonlinear transformation on the linear function from the upper layer.

[0040] The activation function includes a Sigmoid activation function, a Tanh activation function, and a Relu activation function.

[0041] The expression of the Sigmoid activation function is:

[0042]

[0043] wherein s(x) is the value of the Sigmoid function, the value range is (0, 1), and e is a natural constant.

[0044] The expression of the Tanh activation function is:

[0045]

[0046] wherein T(x) is the value of the Tanh function.

[0047] The derivatives of the Sigmoid activation function and the Tanh activation function tend to zero as the independent variable increases, resulting in gradient disappearance, and the formula is:

[0048]

[0049] The expression of the Relu activation function is:

[0050]

[0051] The derivative expression of the Relu activation function is:

[0052]

[0053] Further, step 2.1.4, a loss function is constructed to guide the network to train a model that can ignore rain and light interference.

[0054] The loss function is:

[0055]

[0056] wherein y i is the true value of the i-th sample, y i p is the predicted value of the i-th sample, and n is the number of samples.

[0057] Further, step 2.3, setting the loss function:

[0058]

[0059] where, denotes the minimization with respect to the distribution G; denotes the maximization with respect to the distribution D; is the expectation operator, denoting the mathematical expectation of the successor function over samples x sampled from the real data distribution; D(x) is the probability that the discriminator considers x to come from the real data distribution; G(z) is the generator function; z is usually a random noise vector; is the mathematical expectation of the successor function over random noise vectors z sampled from the noise distribution pz.

[0060] Further, in step 2.4, the number of iterations is set to 32, and the learning rate is 0.015.

[0061] Further, in step d, for each candidate region, a deep learning network based on GoogleNet-GAN is used to extract its neat environment template image and test image.

[0062] In order to better achieve the purpose of the present application, the present application also provides

[0063] Compared with the prior art, the present application has the beneficial effects that: the present application proposes an indoor environment object recognition method based on GoogleNet-GAN. In view of the problem that the existing object detection method has too many parameters of the object region image and is prone to overfitting. The present application uses a GoogleNet network model with Inception module, which can not only maintain the sparsity of the network structure, but also utilize the high computing performance of the dense matrix, solve the problem of overfitting caused by increasing the depth and width of the network, and remove the natural factors (such as rain, light) interference in the photographed region image. At the same time, the present method fully combines the advantages of GoogleNet and GAN, uses GoogleNet as the generator of GAN, and through the adversarial game of the generator and the discriminator, effectively improves the feature extraction ability of GoogleNet.

[0064] The present application has strong anti-interference and robustness, and can recognize in the case of multi-light interference and rain interference, and has good recognition performance in complex environment.

[0065] The application can solve the problem of article classification difficulty, according to the characteristics of a large number of article pictures, a deep learning network based on GoogleNet-GAN is proposed, the interference information of light and rain is eliminated through the deep learning network, the feature information of the neat environment and the chaotic environment of the article is extracted, and the similarity of the two is calculated to obtain the abnormal area, and the real messy area is screened out, which has good recognition performance in complex environment, effectively solves the problem of article area rectification difficulty, and can adapt to different data amount situations.

[0066] The application extracts the potential article area, and performs similarity discrimination on each candidate area to filter out the article area meeting the condition. BRIEF DESCRIPTION OF DRAWINGS

[0067] In order to more clearly illustrate the technical solutions of the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0068] Figure 1 The principle diagram of the first stage detection of the application;

[0069] Figure 2 The principle diagram of the GoogleNet network model of the application;

[0070] Figure 3 The principle diagram of the Unet model under sampling of the application;

[0071] Figure 4 The principle diagram of the second stage detection of the application;

[0072] Figure 5 The external factor and article area similarity distance diagram;

[0073] Figure 6 The structure diagram of the similarity network of the application;

[0074] Figure 7 The flowchart of the article classification detection method of the application. DETAILED DESCRIPTION

[0075] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0076] Embodiment one: please refer to the drawings of the specification Figure 7 A kind of article classification detection method, overall using the combination method of image feature comparison and similarity network filtering, specifically including the following steps: setting template, setting parameter, first stage detection, second stage detection, configuration filtering and output detection result;

[0077] Step one, setting template: setting template image, corresponding mask image and filtering configuration parameter file under normal circumstances for different article environment;Different mask area configuration parameter example is as follows:

[0078] Region pixel value Region first stage threshold Region second stage threshold Region filter ratio threshold 50 0.2 0.1 0.1 100 0.4 0.05 0.005

[0079] Step two, setting parameter: set the basic first stage detection parameter, mainly including:

[0080] 1, setting comparison discrimination threshold, greater than the threshold, it is considered that exception, the threshold is smaller, the more detection quantity is;

[0081] 2, setting local maximum value positioning, the smaller the detection parameter, the more detection quantity is;

[0082] 3, setting comparison area threshold, greater than the threshold, it is considered that exception.

[0083] Step three, first stage detection: potential area extraction;Specifically including the following steps:

[0084] Step a, using deep learning network method (such as GoogleNet, ResNet (residual network) or YoLo (You only Look once) etc.) or SSIM method (structural similarity index) to extract the features of the test image for the test image;

[0085] Wherein, SSIM method is: this method is mainly according to image structure similarity calculation, obtains similarity heat map, and similarity heat map is represented the pixel level similarity degree of test image and template image.

[0086] The deep learning network method is: such method mainly utilizes the powerful feature extraction ability of deep convolutional neural network, selects a pre-trained model (such as GoogleNet, ResNet, YoLo, etc.) on a large data set such as ImageNet (image network) and CoCo (common objects in context) for deep feature extraction, and cascades different receptive field levels to obtain deep features of different receptive fields, and cascades different down-sampled feature maps, here the size adjustment and splicing method is used for cascading different layers.

[0087] Step b, calculate the similarity distance of each pixel point deep learning by using cosine distance, if the similarity distance is less than a certain threshold (comparison and discrimination threshold, local maximum positioning, comparison area threshold set in the previous step), set as a candidate point;

[0088] Step c, merge the candidate points, calculate the minimum rectangular frame (i.e. circumscribed rectangular frame) that can completely contain all points for each group of candidate points after merging, obtain the rectangular frame coordinates, and obtain the candidate region.

[0089] Due to the interference of natural conditions such as light and rain, etc., the image may contain interference information, which affects the recognition of the environment area. Therefore, a deep learning network based on GoogleNet-GAN is further constructed, the neat environment template images at different time points are collected, the deep learning network learns the shadow and light, etc. to eliminate the misjudgment caused by light.

[0090] When the deep learning network method in step a adopts the deep learning network based on GoogleNet-GAN, the following steps are included:

[0091] Step 1, data preprocessing: taking a place as an example, the neat environment template images are taken at different time periods, in order to remove the interference of rain, light, etc., the camera position is fixed, the neat environment template images are taken at different time points, and the neat environment template images under different light and rain are taken, the image size is fixed, the clearest image is selected and gray processing is performed, a 200*300 matrix is obtained, denoted as x1, x2, x3, …, x g ; g represents the gth image;

[0092] Step 2: Model construction: Using the GAN network (Generative Adversarial Network) as the main framework, the GAN network is a generative model based on game theory scenarios. The model consists of a generator and a discriminator. The discriminator determines whether the data source is real data. During the training of the generator G and the discriminator D in the GAN network, through alternating iterations, one is fixed and the parameter weights of the other are updated. In this process, both parties optimize their own networks. The generator G continuously optimizes its ability to predict the future network, and the discriminator D continuously improves its ability to distinguish between predicted results and real network structures, thus forming a competitive confrontation until the two sides reach a dynamic balance. The specific steps include:

[0093] Step 2.1, such as Figure 2 As shown in the figure, the generator G uses the GoogleNet network model, which includes the Inception module. In order to maintain the sparsity of the network structure and utilize the high computing performance of dense matrices, the GoogleNet team proposed a basic neuron structure called "Inception". Based on this architecture, a model that is resistant to external interference such as complex light and rain can be effectively trained.

[0094] The GoogleNet network model is composed of a stack of convolutional layers and pooling layers. The convolutional layers and pooling layers are key steps in feature extraction, while the other layers play a role in connection and classification. The feature map input to the convolutional neural network can be regarded as a matrix composed of pixel values ​​of the captured environment image. By stacking the convolutional layers and pooling layers, the image is traversed to extract abstract high-dimensional features for later differential comparison:

[0095]

[0096] Where: O is the side length of the output feature map, I is the side length of the grayscale value matrix of the neat environment template image, F is the side length of the convolution kernel, P is the zero padding value, and S is the convolution kernel movement step size;

[0097] In the GoogleNet network model, the Inception module performs the following steps:

[0098] Step 2.1.1, convolution layer performs convolution calculation: input template image X p，q , p∈[0,4], q∈[0,4], X is the input image data, p and q are subscripts used to index specific locations or regions in the image, and the convolution kernel is w m，n , m∈[0,1], n∈[0,1], m is the size of the convolution kernel in the height direction, n is the size of the convolution kernel in the width direction, then the convolution calculation formula is:

[0099]

[0100] where y m，n is the value at the position (m, n) on the output feature map after convolution calculation, reflecting the feature information of the corresponding region after convolution operation, g(.) is the activation function, the accumulated result (linear transformation result) of convolution calculation is input into the activation function, introducing a non-linear factor, so that the neural network can learn and represent more complex patterns, "*" is the convolution operation, and b is the convolution bias.

[0101] Step 2.1.2, in the Inception module, in order to solve the gradient disappearance, a batch normalization layer is introduced before the activation function of each layer, and the tensor after convolution is batch normalized, so that the variable entering the activation function is distributed in the gradient maximum position with mean value 0 and unit variance, and then an affine transformation is performed to pass it to the activation function.

[0102] Suppose a batch of input variables x1, x2, …, x k , respectively, find their mean and variance:

[0103]

[0104] where k is the number of data, i.e. the number of samples, representing the total number of data points participating in the standard deviation calculation, i is an index variable used to traverse each data point from 1 to k, the value range is from 1 to k, and each data is pointed to in turn. x i is the value of the i-th data point, i.e. the specific observation value in the data set;

[0105] Then the batch x value is batch normalized:

[0106]

[0107] where, is the value after standardization, also known as the standardized variable. Its value reflects the position of the original data x i relative to the data distribution after adjustment, which usually has a standard scale, which helps to improve the training effect and convergence speed of the model. ε is a very small positive number, usually referred to as a smoothing term or numerical stability term. The main purpose of adding ε is to avoid the case where the standard deviation denominator is zero.

[0108] Then an affine transformation is performed:

[0109] y = α x + β

[0110] where α xThe slope, also called the regression coefficient, determines the degree of inclination of the straight line and reflects the average change in y when x changes by one unit. In regression analysis, it reflects the degree and direction of the influence of the independent variable x on the dependent variable y, and beta is the intercept, which is the value of y when x=0, that is, the y-coordinate of the intersection point of the straight line and the y-axis.

[0111] Step 2.1.3. Add an activation function after the batch normalization layer to perform nonlinear transformation on the linear function passed from the upper layer and solve the nonlinear classification problem.

[0112] The activation function includes Sigmoid activation function, Tanh activation function and Relu activation function.

[0113] The expression of Sigmoid activation function is:

[0114]

[0115] Where s(x) is the value of Sigmoid function, the value range is (0, 1), and e is the natural constant.

[0116] The expression of Tanh activation function is:

[0117]

[0118] Where T(x) is the value of Tanh function, and x is the input value of the activation function.

[0119] The derivative of Sigmoid activation function and Tanh activation function tends to zero as the independent variable increases, resulting in gradient disappearance, and the formula is:

[0120]

[0121] The expression of Relu activation function is:

[0122]

[0123] The derivative expression of Relu activation function is:

[0124]

[0125] Step 2.1.4. Construct a loss function to guide the network to train a model that can ignore rain and light interference;

[0126] The loss function is:

[0127]

[0128] Where y i is the true value of the i-th sample, y i pPi is the predicted value of the i-th sample, n is the number of samples, i.e. the total number of data points participating in the calculation;

[0129] Step 2.2, as shown in the figure, the discriminator D can use a simple convolutional neural network, such as using Unet network model, down-sampling from 200*300 to 0 / 1, improving the network training ability; the data distribution output by the generator gradually approaches the distribution of the training set "real" data under the guidance of the discriminator. Figure 3 Step 2.3, set the loss function:

[0130]

[0131]

[0132] Wherein, Indicates the minimum value of a certain quantity with respect to the distribution G. Indicates the maximum value of a certain quantity with respect to the distribution D. Is the expected operator, which means that for a sample x sampled from the real data distribution, the mathematical expectation of its successor function (log D(x)) is calculated; D(x) is a probability value output, which represents the probability that the discriminator considers that the input x comes from the real data distribution. G(z) is the generator (Generator) function. z is usually a random noise vector, and the generator G maps the noise z to the generated data sample G(z). Is the mathematical expectation of the successor function (log(1-D(G(z)))) of the random noise vector z sampled from the noise distribution pz.

[0133] Step 2.4, model training: according to the data quantity, set appropriate iteration times and learning rate, such as setting the iteration times to 32 times and the learning rate to 0.015 in the present application.

[0134] Step four, second stage detection: latent area filtering; specifically including the following steps:

[0135] Step d, for each candidate area, use the deep learning network based on GoogleNet-GAN to extract its neat environment template image and test image;

[0136] Step e, use the learned similarity network to perform similarity discrimination on the neat environment template image and the test image, and if the similarity is greater than a certain threshold, it is considered that the external factors such as light and rain interfere, and the candidate area is filtered out;

[0137] Step f, output the final rectangular frame area, which is the inconsistent area.

[0138] ​This stage is mainly to distinguish the light and rain from the results of the first stage detection, that is, to further distinguish the light and rain from the areas judged as abnormal in the first stage detection, to eliminate the interference factors of rain and light, and to reduce the misjudgment. The prior art scheme is to use a deep learning scheme, that is, to collect all potential inconsistent areas by the first stage detection, which includes inconsistent articles, inconsistent light and inconsistent rain, etc. The present application expects to consider only the inconsistent articles, and the rest belongs to interference. By designing a similarity network, the network learns the similarity of the two conditions, that is, the similarity of the articles themselves should be relatively low, and the similarity caused by factors such as light and rain should be very high, so as to eliminate the misjudgment of light, rain and other factors. The similarity distance of the image pairs caused by different factors is counted by the similarity network Figure 6 , as shown in Figure 4 , the similarity distance caused by external factors such as light and rain is very small, almost close to 0, while the similarity distance of the image pairs caused by the articles themselves is relatively large, for example, setting the threshold value to 0.5 can eliminate most external factor interference, while also ensuring that the image pairs caused by the articles themselves are not misjudged as external interference.

[0139] As shown in Figure 4 , the chaotic environment image for testing and the neat environment template image are photographed, input into the anti-interference deep learning network based on GoogleNet-GAN, and the images excluding the interference of light, rain and other factors are output; the similarity network is used to compare the features of the corresponding matrix of the output test image and the neat environment template image, the similarity distance of each pixel point is calculated by using the cosine distance, and the abnormal points are selected; a suitable comparison threshold value is set, and the abnormal points below the threshold value are connected and merged to obtain the abnormal area.

[0140] The cosine distance calculation formula is:

[0141]

[0142] d = 1 - cos(θ)

[0143] Wherein, cos(θ) is the cosine similarity, d is the cosine distance, A i is the i-th element of vector A, B i is the i-th element of vector B, n is the dimension of the vector, indicating the number of elements in the vector, and i is an index variable.

[0144] The model training process of the similarity network: the input of the network uses 9 channels, inputs two RGB images and their difference images, and outputs a sigmoid cross-entropy loss, 1 represents that the two input images are not similar, and 0 represents that the input images are similar.

[0145] Step five, configure filtering;

[0146] According to the specific threshold set by each region, different thresholds are set for the wall surface and the ground, and the detection result filtering is carried out according to the threshold, so that the similarity distance graph of external factors and the object region is obtained, for example Figure 5

[0147] Region pixel value Region first stage threshold Region second stage threshold Region filter ratio threshold 50 0.3 0.5 0.1 100 0.5 0.5 0.1 150 0.5 0.5 0.1 200 0.3 0.5 0.1 250 0.3 0.5 0.1 255 0.3 0.5 0.1

[0148] Step six, output the detection result.

[0149] The application provides an indoor environment object recognition method based on GoogleNet-GAN. In view of the problem that the existing object detection method has too many object region image parameters and is prone to overfitting. The application adopts a GoogleNet network model with an Inception module, which can maintain the sparsity of the network structure and utilize the high computing performance of the dense matrix, solve the problem of overfitting caused by increasing the network depth and width, and remove the interference of natural factors (such as rain and light) in or outside the environment in the captured region image. Meanwhile, the method fully combines the advantages of GoogleNet and GAN, uses GoogleNet as the generator of GAN, and through the adversarial game of the generator and the discriminator, effectively improves the feature extraction capability of GoogleNet.

[0150] The application has strong anti-interference and robustness, and can recognize in the case of multi-light interference and rain interference, and has good recognition performance in complex environments.

[0151] The application can solve the problem of object classification difficulty, according to the characteristics of a large number of object pictures, a deep learning network based on GoogleNet-GAN is proposed, the interference information of light and rain is removed through the deep learning network, the feature information of the object in the neat environment and the object in the messy environment is extracted, and the similarity calculation is performed on the two to obtain the abnormal region, and the real messy region is screened out, and the application has good recognition performance in complex environments, effectively solves the problem of object region improvement difficulty, and can adapt to different data amount situations.

[0152] The application extracts the potential object region, and performs similarity discrimination on each candidate region to filter out the object region meeting the condition.

[0153] ​The above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalent features; and these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for detecting object classification, characterized in that: The following steps are involved: The first stage of detection: potential area extraction; specifically includes the following steps: Step a: extracting features of the test image using a deep learning network method or SSIM method; Step b: Calculate the similarity distance of each pixel point using the cosine distance. If the similarity distance is less than the threshold, it is set as a candidate point. Step c: Merge the candidate points, and for each group of merged candidate points, calculate the bounding rectangle that can completely contain all the points, obtain the coordinates of the rectangle, and obtain the candidate area; The second stage of detection: potential area filtering; including the following steps: Step d: For each candidate area, extract its neat environment template image and test image; Step e: Using the learned similarity network, the neat environment template image and the test image are judged for similarity. If the similarity is greater than a threshold, it is considered to be interference and the candidate area is filtered out. Step f: Output the final rectangular frame areas, which are the inconsistent areas.

2. The object classification detection method according to claim 1, characterized in that: The deep learning network method in step a adopts a deep learning network method based on GoogleNet-GAN.

3. The object classification detection method according to claim 2, characterized in that: The deep learning network method based on GoogleNet-GAN includes the following steps: Step 1: Data preprocessing: Take a neat environment template image under rain and different light conditions, process the image in grayscale, and obtain a 200*300 matrix, denoted as x n , n represents the nth image; Step 2: Model construction: Use the GAN network as the main framework, which consists of a generator and a discriminator. The generator G uses the GoogleNet network model. During the training of the generator G and the discriminator D in the GAN network, one is fixed and the parameter weight of the other is updated through alternating iterations.

4. The object classification detection method according to claim 3, characterized in that: The model building process in step 2 is as follows: Step 2.1: Generator G uses the GoogleNet network model, which includes the Inception module. The GoogleNet network model traverses the image by superimposing convolutional layers and pooling layers to extract abstract high-dimensional features: Where: O is the side length of the output feature map, I is the side length of the grayscale value matrix of the neat environment template image, F is the side length of the convolution kernel, P is the zero padding value, and S is the convolution kernel movement step size; The Inception module of the GoogleNet network model performs the following steps: Step 2.1.1, the convolution layer performs convolution calculation; Step 2.1.2: Introduce a batch normalization layer before the activation function of each layer. Perform batch normalization on the tensor after convolution so that the variables entering the activation function are distributed at the maximum gradient position with mean 0 and unit variance. Then perform an affine transformation and pass it to the activation function. Step 2.1.

3. Add an activation function after the batch normalization layer to perform a nonlinear transformation on the linear function passed from the upper layer. Step 2.1.4: Construct a loss function to guide the network to train a model that can ignore rain and light interference; Step 2.2: The discriminator D uses the Unet network model to downsample the output from 200*300 to 0 / 1. Step 2.3, set the loss function; Step 2.4: Perform model training.

5. The object classification detection method according to claim 4, characterized in that: Step 2.1.1, convolution layer performs convolution calculation: input template image X p，q , p∈[0,4], q∈[0,4], X is the input image data, p and q are subscripts used to index specific locations or regions in the image, and the convolution kernel is W m，n , m∈[0,1], n∈[0,1], m is the size of the convolution kernel in the height direction, n is the size of the convolution kernel in the width direction, then the convolution calculation formula is: Among them, y m，n After the convolution calculation, the value at the position with coordinates (m, n) on the output feature map is obtained. g(.) is the activation function, "*" is the convolution operation, and b is the convolution bias.

6. The object classification detection method according to claim 4, characterized in that: Step 2.1.

3. Add an activation function after the batch normalization layer to perform a nonlinear transformation on the linear function passed from the upper layer. Activation functions include Sigmoid activation function, Tanh activation function and Relu activation function; The Sigmoid activation function expression is: Where s(x) is the value of the Sigmoid function, ranging from (0, 1), and e is a natural constant; The Tanh activation function expression is: Where T(x) is the value of the Tanh function; The derivatives of the Sigmoid activation function and the Tanh activation function tend to zero as the independent variable increases, causing the gradient to disappear. The formula is: The expression of the Relu activation function is: The derivative expression of the Relu activation function is:

7. The object classification detection method according to claim 4, characterized in that: Step 2.1.4: Construct a loss function to guide the network to train a model that can ignore rain and light interference; The loss function is: Among them, y i is the true value of the i-th sample, y i p is the predicted value of the i-th sample, and n is the number of samples.

8. The object classification detection method according to claim 4, characterized in that: Step 2.3, set the loss function: in, Indicates finding the minimum value of distribution G; Indicates finding the maximum value with respect to distribution D; is the expectation operator, which means the mathematical expectation of the successor function of a sample x sampled from the true data distribution; D(x) is the probability that the discriminator believes that x comes from the true data distribution; G(z) is the generator function; z is usually a random noise vector; To find the mathematical expectation of the successor function of a random noise vector z sampled from the noise distribution pz.

9. The object classification detection method according to claim 4, characterized in that: In step 2.4, set the number of iterations to 32 and the learning rate to 0.

015.

10. The object classification detection method according to any one of claims 1 to 9, characterized in that: In step d, for each candidate region, a deep learning network based on GoogleNet-GAN is used to extract its neat environment template image and test image.

Images