AI computer vision target detection method based on adaptive compensation mechanism

Abstract

The invention discloses an AI computer vision target detection method based on an adaptive compensation mechanism. The method comprises the following steps of data acquisition and labeling, data expansion, feature extraction model training, feature dimension reduction model training, classifier model training and target detection and recognition. Compared with the prior art, the method has the advantages that 1, the expression of morphological characteristics is enhanced through the generative adversarial network, the quality and details of the generated image are improved, and the training effect and generalization ability of the model are enhanced; 2, the exploration type gradient descent method optimizes the training process, and by predicting and adjusting a gradient updating strategy, the training time is shortened, and the stability and performance of the model are improved; 3, the reverse reconstruction strategy of the auto-encoder effectively ensures the retention of key features, and improves the quality of dimension reduction data and the accuracy of subsequent processing; and 4, a dynamic topology routing mechanism enables the decision tree to be adaptively adjusted according to data characteristics, so that the classification accuracy and the processing speed are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of visual target detection, and specifically to an AI computer vision target detection method based on an adaptive compensation mechanism. Background Art

[0002] With the widespread application of intelligent technologies, such as autonomous driving, intelligent monitoring, and industrial automation, the demand for high-precision and high-efficiency target detection technology is increasing. The core of the target detection task is to accurately identify and locate specific targets in images, such as pedestrians, vehicles, and other important objects, in a variety of complex environments.

[0003] The Chinese invention patent with application number CN201610024820.3 proposes a moving target detection method based on a mixed Gaussian model and superpixel segmentation. The task of moving target detection is to extract objects of interest as much as possible from complex scenes and present the results in a binary image. The method of the present invention first uses a Gaussian mixture model to model the background and obtain the background image of the current frame. Secondly, the current frame is subjected to superpixel segmentation using the SLIC (simple linear iterative clustering) algorithm. Finally, the texture features of the segmented image and the background image are extracted by LTP (local ternary pattern) respectively, and then compared to obtain the moving target. The beneficial effects of the present invention are: it is possible to detect moving targets in real time based on the video images captured by the camera; the detection method is efficient and accurate.

[0004] The Chinese invention patent with application number CN201810907184.8 proposes a method for optical remote sensing image target detection based on a deep convolutional network with dilated residual convolution, which solves the problems of low accuracy and high false alarm rate in optical remote sensing image aircraft and ship detection in the prior art. The implementation steps are as follows: construct a test data set; construct a training data set; build a target detection network based on dilated residual convolution for expanding the feature receptive field; use the training data set to train the target detection network based on dilated residual convolution; use the trained target detection network based on dilated residual convolution to perform target detection on the test data set; output the test results. The network constructed by the present invention uses a dilated residual convolution module and feature fusion, which is more suitable for optical remote sensing image target detection, not only improving the accuracy of ordinary targets, but also significantly improving the accuracy of small target detection in optical remote sensing images. It is used for ground object target detection in optical remote sensing images.

[0005] The Chinese invention patent with application number CN201610352704.4 proposes a method for detecting small infrared targets under non-uniform backgrounds, which mainly solves the problems of small target detection probability and false alarm probability deteriorating seriously under non-uniform backgrounds such as land-air coexistence, cloudy or irregular cloud edges in the prior art. The implementation process is as follows: the current infrared image is divided into blocks using a sliding window method to obtain multiple sub-blocks; a protection area and a local background area are constructed in the neighborhood of the current detection sub-block; the maximum gray value of the pixel of the current detection sub-block and the maximum and mean of the pixel gray value in the local background area are calculated; a saliency detection map is obtained using enhanced local Weber contrast measurement; and finally, the target detection is completed using threshold decision technology. The present invention has the advantages of high small target detection probability and low false alarm probability, and can be used for the detection of small infrared targets under non-uniform backgrounds such as land-air boundaries and cloud edges.

[0006] The existing technology has defects and limitations in the following aspects:

[0007] 1. The data augmentation methods in the existing technologies usually fail to make full use of morphological features, resulting in insufficient quality of generated images and insufficient generalization ability of the model.

[0008] 2. The gradient vanishing or gradient exploding problems often encountered in traditional neural network training limit the training efficiency and network performance.

[0009] 3. Traditional feature dimensionality reduction methods often lead to the loss of key features due to their high dependence and inaccuracy in the encoding and decoding process, affecting the subsequent data processing effect.

[0010] 4. The traditional decision tree model has a static structure and is not suitable for complex and changing data characteristics, resulting in unsatisfactory classification effect and efficiency. Summary of the invention

[0011] The technical problem to be solved by the present invention is to provide an AI computer vision target detection method based on an adaptive compensation mechanism in view of the deficiencies in the prior art proposed in the above background technology.

[0012] To solve the above technical problems, the present invention provides a technical solution: an AI computer vision target detection method based on an adaptive compensation mechanism, which comprises the following steps:

[0013] Step 1 S1, data collection and annotation;

[0014] Step 2 S2, data expansion;

[0015] Step 3 S3, feature extraction model training;

[0016] Step 4 S4, feature dimension reduction model training;

[0017] Step 5 S5, classifier model training;

[0018] Step six S6: target detection and recognition.

[0019] Furthermore, the step 1 S1 uses image data, which is obtained from surveillance cameras in various environments, including but not limited to urban transportation, commercial areas and industrial areas, and the collected image data is stored in jpg data format;

[0020] The collected data is labeled manually, and the labelers classify the objects in the image.

[0021] Furthermore, the step two S2 adopts a generative adversarial network based on morphological adaptive compensation features to generate samples, thereby realizing data expansion. The generative adversarial network based on morphological adaptive compensation features includes two parts: a generator and a discriminator. The generator is responsible for generating realistic image data from noise distribution, and the discriminator is responsible for distinguishing the generated image from the real image.

[0022] Furthermore, the step three S3 uses a 6-layer fully connected neural network for feature extraction, and the structure of the 6-layer fully connected neural network is as follows:

[0023] a-Input layer: receives and transmits sample data after sample expansion;

[0024] b-Hidden layers: The network contains four hidden layers, each consisting of a certain number of neurons;

[0025] b1-first hidden layer: has 128 neurons, used for preliminary feature extraction;

[0026] b2-second hidden layer: has 256 neurons to further abstract features;

[0027] b3-the third hidden layer: has 512 neurons, further deepening feature extraction;

[0028] b4-the fourth hidden layer: has 1024 neurons, further deepening feature extraction;

[0029] b5 - output layer: has 512 neurons.

[0030] Furthermore, the step four S4 adopts an autoencoder algorithm based on inverse reconstruction as a feature dimensionality reduction model. The autoencoder includes two parts: an encoder and a decoder. The encoder is responsible for compressing the input data into a low-dimensional representation, and the decoder reconstructs the data from the low-dimensional representation.

[0031] Furthermore, the step five S5 adopts a deep neural decision tree classification algorithm based on sparse constraints as a classifier. The deep neural decision tree classification algorithm based on sparse constraints includes multiple layers of decision nodes and terminal decision leaf nodes. After receiving input features, each node guides different subsequent nodes or makes classification decisions through a series of decision rules determined by training.

[0032] Furthermore, step six S6 uses the trained model to process new samples, and the input data is a single frame image captured from the surveillance video.

[0033] After adopting the above structure, the present invention has the following advantages: 1. The performance of morphological features is enhanced by generating adversarial networks, and the quality and details of generated images are improved, thereby enhancing the training effect and generalization ability of the model.

[0034] 2. The exploratory gradient descent method optimizes the training process. By predicting and adjusting the gradient update strategy, it reduces the training time and improves the stability and performance of the model.

[0035] 3. The inverse reconstruction strategy of the autoencoder effectively ensures the retention of key features, improves the quality of the reduced dimensionality data and the accuracy of subsequent processing.

[0036] 4. The dynamic topology routing mechanism enables the decision tree to adaptively adjust according to data characteristics, improving classification accuracy and processing speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a flowchart of an AI computer vision target detection method based on an adaptive compensation mechanism.

[0038] Figure 2 It is a training flow diagram of an exploratory gradient descent neural network algorithm for an AI computer vision target detection method based on an adaptive compensation mechanism.

[0039] Figure 3 It is a schematic diagram of the training process of the autoencoder algorithm of the AI ​​computer vision target detection method based on the adaptive compensation mechanism. DETAILED DESCRIPTION

[0040] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0041] Combined with Figure 1-3 , an AI computer vision target detection method based on an adaptive compensation mechanism, which includes the following steps:

[0042] Step 1 S1, data collection and annotation;

[0043] The step 1 S1 uses image data, which is obtained from surveillance cameras in various environments, including but not limited to urban traffic, commercial areas and industrial areas, and the collected image data is stored in jpg data format;

[0044] The collected data is labeled manually, and the labeling personnel classify the targets in the image. The specific common labeling categories include: pedestrians, vehicles, traffic signs, obstacles, a total of 4 categories.

[0045] Step 2 S2, data expansion;

[0046] The step two S2 adopts a generative adversarial network based on morphological adaptive compensation features to generate samples, thereby realizing data expansion. The generative adversarial network based on morphological adaptive compensation features includes two parts: a generator and a discriminator. The generator is responsible for generating realistic image data from noise distribution, and the discriminator is responsible for distinguishing the generated image from the real image. On the basis of the traditional generative adversarial network, the generator of the generative adversarial network proposed in the present invention adopts a morphological feature extraction module to extract and enhance the morphological features of the target, such as edges, textures, etc., from the noise data, so as to improve the details of the generated image and further improve the quality of the generated image.

[0047] Specifically, the training process of the generative adversarial network algorithm based on morphological adaptive compensation features is as follows:

[0048] S201, initializing the parameters of the generative adversarial network model. In one embodiment, the generator G is initialized. c and the discriminator D c Weight and The way is expressed as:

[0049]

[0050]

[0051] In the formula, and are the weights of the generator and the discriminator respectively, ~ means obeying a specific distribution, represents a normal distribution with a mean of 0 and a standard deviation of 0.02.

[0052] S202, perform morphological feature extraction training, and optimize the parameters of the extractor by minimizing the difference between the morphological features and the original image by training the morphological feature extraction module alone. Specifically, define the morphological feature extractor F c , whose training goal is to minimize the real image Its morphological characteristicsc The difference between is expressed as:

[0053]

[0054] In the formula, is the feature loss function, represents the real image data, M c Represents the morphological features extracted from real images.

[0055] In one embodiment, the morphological feature extraction M c The calculation is performed through a multi-layer convolutional neural network, and each layer of convolution increases the edge and texture extraction capabilities, which can be expressed as:

[0056]

[0057] In the formula, * represents the convolution operation, Re() is the ReLU activation function, and are the convolution kernel weight and bias respectively.

[0058] S203, generate an initial noise sample as the input of the generator, expressed as:

[0059]

[0060] In the formula, z c is the noise vector, is a normal distribution with mean 0 and covariance matrix I.

[0061] S204, generator G c Use the noise vector z c and morphological characteristics c Generate an image, represented as:

[0062]

[0063] In the formula, is the image produced by the generator.

[0064] S205, Discriminator D c Evaluate the authenticity of the generated image and output Feedback to the generator, the calculation method of feedback loss is expressed as:

[0065]

[0066] In the formula, is the feedback loss.

[0067] Furthermore, the output of the discriminator optimizes the output of the generator through a dynamic adjustment mechanism. The adjustment amount of the generator weight is calculated as follows:

[0068]

[0069] In the formula, is the adjustment amount for the generator weight, η bf is the learning rate, D′ c () is the gradient of the discriminator with respect to its input. Preferably, η bf Set to 0.01.

[0070] S206. Calculate double loss, including adversarial loss and content loss The generator and discriminator are optimized, and the calculation method is expressed as:

[0071]

[0072]

[0073]

[0074] In the formula, is the adversarial loss, which evaluates the distinction between generated images and real images; is the content loss, assessing the preservation of morphological features; is the total loss function of the generative adversarial network; and are the first and second weight coefficients of the loss function respectively; SSIM() is the structural similarity index, which is used to measure the visual similarity of two images; β c is the structural similarity weight. Preferably, and They are set to 0.1 and 0.6 respectively, and the structural similarity is set to 0.3.

[0075] S207, using the gradient descent method to update the parameters of the generator and the discriminator, the update method is expressed as:

[0076]

[0077]

[0078] In the formula, α c is the learning rate, ← is the parameter update operation, and denote the gradients of the generator and discriminator weights respectively.

[0079] S208, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0080] Step 3 S3, feature extraction model training;

[0081] The step 3 S3 uses a 6-layer fully connected neural network to extract features. The structure of the 6-layer fully connected neural network is as follows:

[0082] a-Input layer: receives and transmits sample data after sample expansion;

[0083] b-Hidden layers: The network contains four hidden layers, each consisting of a certain number of neurons;

[0084] b1-first hidden layer: has 128 neurons, used for preliminary feature extraction;

[0085] b2-second hidden layer: has 256 neurons to further abstract features;

[0086] b3-the third hidden layer: has 512 neurons, further deepening feature extraction;

[0087] b4-the fourth hidden layer: has 1024 neurons, further deepening feature extraction;

[0088] b5 - output layer: has 512 neurons.

[0089] In the prior art, when some schemes use neural networks for feature extraction, in certain neural network structures, they may encounter problems such as gradient vanishing, gradient explosion, or falling into local optimal solutions, which affects the stability of training and the performance of the model. The present invention adopts a neural network based on exploratory gradient descent as a feature extraction model. In the traditional neural network training process, the exploratory gradient descent method is used to not only consider the current gradient information, but also adopt a forward exploration mechanism to predict the possible change trend of the gradient before the gradient is updated, and adjust the learning step size and direction to achieve faster convergence speed and better local minimum, thereby enhancing the generalization ability of the feature.

[0090] Specifically, the training process of the neural network algorithm based on exploratory gradient descent is as follows:

[0091] S301, initializing the parameters of the neural network, including the weight and bias parameters of the neural network. In one embodiment, the initialization method is expressed as:

[0092]

[0093]

[0094] Where W p，i is the weight matrix of the i-th layer of the neural network; b p，i is the bias vector of the i-th layer of the neural network; σ p is the scaling factor of the neural network weights, β p is the scaling factor of the bias of the neural network; represents a normal distribution with a mean of 0 and a standard deviation of 1; ∈ p is a small noise. Preferably, ∈ p is a noise vector whose elements are all 0.001.

[0095] S302, the input data passes through each layer of the neural network, and each layer is subjected to linear transformation and nonlinear activation function processing, and finally the feature representation of the output layer is obtained. The output calculation method of each layer is:

[0096] a p，i =Rez(W p，i ·a p，i-1 +b p,i )

[0097] In the formula, a p，i represents the activation output of the i-th layer of the neural network; a p，i-1 is the output of the i-1th layer of the neural network; Rez() is the adaptive ReLU activation function.

[0098] In one embodiment, let the input of the adaptive ReLU activation function be x zsy , the calculation method of the adaptive ReLU activation function is expressed as:

[0099] Rez(x zsy )=γ p max(0, x zsy ) + ξ p min(0, x zsy )

[0100] In the formula, γ p and p are the adaptive coefficients of the positive and negative parts respectively, max() is the maximum value function, and min() is the minimum value function. Preferably, γ p Set to 0.3, ξ p Set to 0.7.

[0101] S303, using an exploratory gradient descent update strategy, predicting possible gradient changes before each update of the neural network parameters, adjusting the step size and direction of the gradient update according to the prediction results, optimizing the parameter adjustment during the training process, and avoiding falling into the local optimum too early. The calculation method is expressed as:

[0102]

[0103]

[0104] In the formula, ΔW p，i and Δb p,i are the update amounts of the i-th layer weight and i-th layer bias of the neural network respectively; η p is the learning rate of the neural network; is the loss function with respect to parameter W p，i The gradient of is the loss function with respect to parameter b p,i The gradient of is the loss function of the neural network; α p is the exploration intensity coefficient; is the prediction function of the gradient. Preferably, η p Set to 0.03, Cross entropy loss is used.

[0105] In one embodiment, the exploratory gradient prediction function uses historical gradient data to perform linear regression prediction, and the calculation method is expressed as:

[0106]

[0107] In the formula, Represents the loss function at the t-1th iteration with respect to the parameter W p，i The gradient of t is the attenuation factor; α p is the exploration intensity coefficient; Tgz is the total number of iterations. Preferably, β t Set to 0.98, alpha p Set to 2.

[0108] Furthermore, the weight and bias parameters of the neural network are updated by using the error back propagation method.

[0109] S304, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0110] Step 4 S4, feature dimension reduction model training;

[0111] The step 4 S4 adopts an autoencoder algorithm based on inverse reconstruction as a feature dimensionality reduction model, and the autoencoder includes two parts: an encoder and a decoder. The encoder is responsible for compressing the input data into a low-dimensional representation, and the decoder reconstructs the data from the low-dimensional representation.

[0112] Specifically, the training process of the autoencoder algorithm based on inverse reconstruction is as follows:

[0113] S401, initializing the parameters of the autoencoder algorithm, including the weight and bias parameters of the autoencoder. In one embodiment, the initialization method is expressed as:

[0114]

[0115] b r,init =0

[0116] Where W r，init is the initial weight matrix of the autoencoder; b r,init is the initial bias of the autoencoder; n r,in and n r，out Represent the number of neurons in the input and output layers respectively; U(-1,1) represents a random number uniformly distributed in the interval [-1,1].

[0117] S402, the input feature-extracted data is converted into a low-dimensional space representation through the encoder, and the low-dimensional representation is reconstructed into an output with the same dimension as the original data through the decoder. In the forward propagation process, the output calculation method of the encoder is expressed as:

[0118] z r =Sig(W r,enc ·x r +b r，enc )

[0119] And, the reconstructed output of the decoder is:

[0120]

[0121] In the formula, z r is the low-dimensional feature representation of the code; Sig() is the Sigmoid activation function; W r，enc and W r,dec are the weights of the encoder and decoder respectively; b r，enc and b r,dec are the biases of the encoder and decoder respectively; x r is the input data; is the output of the reconstruction.

[0122] S403, calculating the inverse reconstruction loss of the autoencoder, and defining the calculation method of the inverse reconstruction loss function as:

[0123]

[0124] Where, L r is the inverse reconstruction loss function; Ntr is the number of samples input in the current batch; |||| represents the L2 norm; λ r is the regularization parameter; D KL (P r ||Q r ) is the KL divergence.

[0125] In one embodiment, KL divergence measures the original data distribution P r and reconstruct the data distribution Q r The difference between them is used to ensure that the key features of the data are not lost during the dimensionality reduction process. The calculation method is expressed as:

[0126]

[0127] In the formula, p r，i and q r，i are the probability distributions of the original data and the reconstructed data in the i-th dimension, respectively, and Mtr is the feature dimension before feature dimensionality reduction.

[0128] S404, according to the gradient calculated by the loss function, the weights and biases of the encoder and decoder are updated using the gradient descent algorithm, and the update method is expressed as:

[0129]

[0130]

[0131]

[0132]

[0133] Where η r is the learning rate; is the updated encoder weight; is the updated encoder bias; is the updated decoder weight; is the updated decoder bias.

[0134] Furthermore, the gradient For example, the calculation method is expressed as:

[0135]

[0136] In the formula, represents the partial derivative of the loss function with respect to the reconstructed output, is the partial derivative during decoding, is the partial derivative during encoding.

[0137] S405, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1200 times.

[0138] Step 5 S5, classifier model training;

[0139] The step five S5 adopts a deep neural decision tree classification algorithm based on sparse constraints as a classifier. The deep neural decision tree classification algorithm based on sparse constraints includes multiple layers of decision nodes and terminal decision leaf nodes. After receiving the input features, each node guides different subsequent nodes or makes classification decisions through a series of decision rules determined by training. The present invention adopts a dynamic topology routing mechanism to dynamically adjust the structure of the decision tree according to the characteristics of the input data, so that the classifier can adaptively adjust its decision path according to different input data characteristics, thereby improving the accuracy and efficiency of classification.

[0140] Specifically, the training process of the deep neural decision tree classification algorithm based on sparse constraints is as follows:

[0141] S501, initializing the parameters of the deep neural decision tree classification algorithm model based on sparse constraints, including initializing the parameters θ of each node of the decision tree u，i and routing weight ρ u，i , the initialization method is expressed as:

[0142]

[0143] ρ u，i ~Uniform(0,1)

[0144] In the formula, θ u，i represents the decision parameter of the i-th node, represents a normal distribution with a mean of 0 and a standard deviation of 0.01; ρ u，i represents the routing weight of the i-th node, and Uniform(0,1) represents the uniform distribution in the interval [0,1].

[0145] S502: The input features after dimensionality reduction are distributed through each layer of the decision tree according to the parameters of each decision node until they reach the leaf node. Each decision node determines the destination of the input features according to the weight after sparse constraint optimization. Specifically, each input sample X u The propagation is performed through each layer of the decision tree. In one embodiment, the calculation method for determining the transmission direction of the i-th node is expressed as:

[0146]

[0147] In the formula, p u，i is the activation function output of the i-th node of the deep neural decision tree; Sig() is the Sigmoid activation function; b u，i is the bias term of the i-th node in the deep neural decision tree.

[0148] S503. Calculate the error between the predicted output and the true label at the leaf node, and use the cross entropy loss function to measure the classification effect. The calculation method is expressed as:

[0149]

[0150] Where, L u is the cross entropy loss function, y u is the true label, is the predicted label, λ u,reg is the regularization coefficient and |||| is the L2 norm.

[0151] S504, adjusting the decision parameter θ through the back propagation algorithm u，i and routing weight ρ u，i , to minimize the loss function, expressed as:

[0152]

[0153]

[0154] In the formula, α u is the learning rate; and are the partial derivatives of the loss function with respect to the decision parameters and routing weights, respectively.

[0155] S505. Dynamically adjust the structure of the decision tree according to the performance feedback during the training process to adapt to the complex data characteristics and changing classification task requirements, expressed as:

[0156]

[0157] In the formula, f adjust () is the loss change ΔL u Function to adjust routing weight, Represents the adjusted routing weight of the i-th node.

[0158] In one embodiment, the adjustment function is specifically defined as a gradient-based adjustment method, which uses an adaptive factor to dynamically adjust the weight according to the loss change. The calculation method is expressed as:

[0159]

[0160] Where η ρ is the learning rate for the routing weight, γ u is a parameter that controls the adjustment sensitivity. Preferably, η ρ Set to 0.01, γ u Set to 3.

[0161] S506, repeat the above steps until the preset stop iteration condition is met, which means the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0162] Step six S6: target detection and recognition.

[0163] The step six S6 uses the trained model to process the new sample. The input data is a single frame image captured from the surveillance video. Preferably, the resolution of the image is 1080p and has been standardized. In one embodiment, the collected raw data is input into the trained feature extraction and feature dimensionality reduction model for feature processing. Further, the processed features are input into the classifier model for classifier training to obtain the classification results. In this embodiment, the classification categories include: pedestrians, vehicles, traffic signs, obstacles, a total of 4 categories.

[0164] The present invention and its implementation methods are described above, and such description is not restrictive, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by it, and does not deviate from the purpose of the invention, and does not creatively design a structure and implementation method similar to the technical solution, they should all fall within the protection scope of the present invention.

Claims

1. An AI computer vision target detection method based on an adaptive compensation mechanism, characterized in that: It includes the following steps: Step 1 S1, data collection and annotation; Step 2 S2, data expansion; Step 3 S3, feature extraction model training; Step 4 S4, feature dimension reduction model training; Step 5 S5, classifier model training; Step six S6: target detection and recognition.

2. The AI ​​computer vision target detection method based on the adaptive compensation mechanism according to claim 1, characterized in that: The step 1 S1 uses image data, which is obtained from surveillance cameras in various environments, including but not limited to urban traffic, commercial areas and industrial areas, and the collected image data is stored in jpg data format; The collected data is labeled manually, and the labelers classify the objects in the image.

3. The AI ​​computer vision target detection method based on the adaptive compensation mechanism according to claim 1, characterized in that: The step two S2 adopts a generative adversarial network based on morphological adaptive compensation features to generate samples, thereby realizing data expansion. The generative adversarial network based on morphological adaptive compensation features includes two parts: a generator and a discriminator. The generator is responsible for generating realistic image data from noise distribution, and the discriminator is responsible for distinguishing the generated image from the real image.

4. The AI ​​computer vision target detection method based on the adaptive compensation mechanism according to claim 1, characterized in that: The step 3 S3 uses a 6-layer fully connected neural network to extract features. The structure of the 6-layer fully connected neural network is as follows: a-Input layer: receives and transmits sample data after sample expansion; b-Hidden layers: The network contains four hidden layers, each consisting of a certain number of neurons; b1-first hidden layer: has 128 neurons, used for preliminary feature extraction; b2-second hidden layer: has 256 neurons to further abstract features; b3-the third hidden layer: has 512 neurons, further deepening feature extraction; b4-the fourth hidden layer: has 1024 neurons, further deepening feature extraction; b5 - output layer: has 512 neurons.

5. The AI ​​computer vision target detection method based on the adaptive compensation mechanism according to claim 1, characterized in that: The step four S4 adopts an autoencoder algorithm based on inverse reconstruction as a feature dimensionality reduction model. The autoencoder includes two parts: an encoder and a decoder. The encoder is responsible for compressing the input data into a low-dimensional representation, and the decoder reconstructs the data from the low-dimensional representation.

6. The AI ​​computer vision target detection method based on the adaptive compensation mechanism according to claim 1, characterized in that: The step five S5 adopts a deep neural decision tree classification algorithm based on sparse constraints as a classifier. The deep neural decision tree classification algorithm based on sparse constraints includes multiple layers of decision nodes and terminal decision leaf nodes. After receiving input features, each node guides different subsequent nodes or makes classification decisions through a series of decision rules determined by training.

7. The AI ​​computer vision target detection method based on the adaptive compensation mechanism according to claim 1, characterized in that: Step six S6 uses the trained model to process the new sample, and the input data is a single frame image captured from the surveillance video.

Images