An airport video image classification method based on artificial intelligence

Abstract

The present invention discloses an airport video image classification method based on artificial intelligence, which relates to the technical field of video image classification. The method performs full video monitoring of the apron area, including the red line area, the stop line and the aircraft approach line; the images obtained by the video monitoring software are spliced ​​and fused in real time using a video image splicing method; the panoramic video image is separated frame by frame to obtain each image, that is, a panoramic image of the apron parking space; the panoramic image of the apron parking space is classified using an illumination-aware multi-scale feature pyramid network as the core architecture. The present invention adopts an adaptive normalization method to enhance the stability of feature expression by fusing local illumination statistics with global channel attention. The present invention adopts a deformable convolution kernel group and scale-aware attention on the basis of the feature pyramid network, which can adaptively adjust the receptive field, dynamically fuse multi-scale features, and improve the ability to capture cross-scale spatial correlations in the image.

Description

Technical Field

[0001] The present invention relates to the technical field of video image classification, and in particular to an airport video image classification method based on artificial intelligence. Background Art

[0002] In the apron environment, parking bay images usually have complex characteristics, including strong illumination changes, background noise interference, coexistence of multi-scale targets, and the existence of similar background areas. These factors make traditional image classification methods face many challenges when processing apron parking bay images.

[0003] Common illumination variations in ramp images can lead to significant brightness variations, which directly impact the stability of image features and, in turn, classification accuracy. Conventional normalization methods (such as Layer-Norm and Batch-Norm) often struggle to effectively address feature distribution shifts in dynamic lighting scenarios, making them ill-suited for classification under varying lighting conditions. Multi-scale objects, such as nearby vehicles and distant signs, often appear in ramp images. These scale variations hinder the ability of traditional convolutional neural networks to extract cross-scale spatial features. A single fixed receptive field cannot cover objects of all scales, resulting in incomplete and inaccurate feature extraction. Apron parking bay images often have similar background areas (such as runway textures). This background similarity makes it difficult to distinguish fine-grained features, and traditional spatial pooling operations can easily lead to loss of detail, which in turn reduces the accuracy and robustness of classification models. Occlusion is also common in ramp images, such as vehicles partially obscuring parking lines or signs. This occlusion not only affects the integrity of the target, but also makes it difficult for traditional methods such as Dropout to effectively deal with the problem of feature loss, thereby reducing the robustness of the model in complex scenarios. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide an airport video image classification method based on artificial intelligence.

[0005] The technical solution adopted to solve the above technical problems is: an airport video image classification method based on artificial intelligence, comprising the following steps:

[0006] S1, conducts full video surveillance of the apron area, including the red line area, stop line and aircraft approach line; S2, uses the video image stitching method to perform real-time stitching and fusion processing on the images obtained by the video surveillance software; S3, separates the panoramic video image frame by frame to obtain each image, that is, the panoramic image of the apron parking space; S4, classifies the panoramic image of the apron parking space, and adopts the illumination-aware multi-scale feature pyramid network as the core architecture. The training process is: S401, adopts the channel-space dual-domain adaptive normalization method to avoid the feature distribution offset caused by significant illumination changes in the panoramic image of the apron parking space; S402, adopts the dynamic multi-core feature pyramid, and adopts the feature pyramid network based on the feature pyramid network. Use deformable convolution kernel group and scale-aware attention, adaptively adjust the receptive field through deformable convolution, and dynamically fuse multi-scale features in combination with scale attention; S403, adopt the difference maximum pooling strategy to enhance feature differences in the maximum pooling process; S404, adopt spatial perception DropBlock, dynamically generate occlusion masks according to feature importance, and generate masks by combining feature amplitude and gradient significance; S405, adopt contrast center loss, and enhance the ability to capture fine-grained differences by combining intra-class aggregation and inter-class separation constraints; S406, repeat the above steps S401 to S405 until the preset stop iteration condition is met, which means that the model training is completed.

[0007] Furthermore, the S403 also includes a light perception activation function, which uses a light intensity normalization term and is expressed as:

[0008] ,

[0009] In the above formula, is the illumination perception activation function, is the input of the light-aware activation function, is the hyperbolic tangent function, is the element-wise multiplication symbol, is the adaptive adjustment coefficient, is the function for calculating the local light intensity, To prevent division by zero errors for small constant terms, For The Gaussian heat map centered on , characterizing the anchor point influence area, is the number of predefined topological anchor points, i.e. the preset number of key areas in the apron parking image. For the predefined Topological anchor point coordinates, For the The attention weights of dynamic anchor points, is the Sigmoid activation function, Anchor point position offset.

[0010] Furthermore, the They are:

[0011] ,

[0012] In the above formula, is a multi-layer perceptron with one hidden layer, GAP is a global average pooling operation; is a 5×5 Gaussian blur kernel, Represents the convolution operation; is the convolution kernel weight matrix;

[0013] is the tensor concatenation operation, For the predefined Topological anchor point coordinates;

[0014] Indicates that the Gaussian heat map is The horizontal coordinate index value corresponding to the eigenvalue of the position, Indicates that the Gaussian heat map is The vertical coordinate index value corresponding to the eigenvalue of the position, For Gaussian heat map The eigenvalues ​​of the position, For the The horizontal coordinates of the predefined topological anchor points are initialized by the analysis or statistical distribution of the apron CAD drawings. For the The horizontal coordinate offset of the anchor point is calculated by Convolution learns from input features, For the The vertical coordinates of the predefined topological anchor points are initialized in the same way , For the The vertical coordinate offset of each anchor point is calculated in the same way , is the standard deviation of the Gaussian kernel;

[0015] is the anchor point position offset, for Convolutional layer operation.

[0016] Furthermore, the Gaussian blur kernel is generated using a two-dimensional Gaussian distribution function The calculation method of the elements in the convolution kernel weight matrix is ​​expressed as:

[0017] ,

[0018] In the above formula, is the Gaussian kernel standard deviation, For Gaussian heat map The eigenvalues ​​of the position, is a positive integer, Is a positive integer.

[0019] Furthermore, the occlusion mask in S404 is expressed as:

[0020] ,

[0021] In the above formula, The feature map is at position The mask value of Sort by importance, is the function for calculating feature importance, The maximum pooling feature map of the difference is The eigenvalues ​​of the position, is the discard ratio, is the feature map height, W is the feature map width;

[0022] Feature location The neighborhood window of is the neighborhood of the difference maximum pooling feature map eigenvalues, L is the loss function, is the symbol of the partial derivative.

[0023] Furthermore, the S404 also includes using a slice adaptive compression module to perform image segmentation and compression, which is expressed as:

[0024] ,

[0025] In the above formula, for Convolutional layer operation, for Convolutional layer operation, is the Hadamard product symbol, is the feature map height, is the feature map width, is the compression feature of the sample, is the number of image slices, is the mask value of the feature map, its position The mask value is , is the function that generates the slice attention mask, expressed as the attention weight matrix of the rth slice, is the illumination-aware activation function.

[0026] Furthermore, the contrast center loss in S405 is expressed as:

[0027] ,

[0028] In the above formula, is the comparison center loss, N is the number of batch samples, For the The compressed features of samples, for norm, is the feature encoding function, For the The center vector of each category, To balance the hyperparameters, To obtain the maximum value function, is the nearest negative class center, and m is the interval threshold;

[0029] To find the nearest negative class among all center vectors of non-current categories, For the The center vector of each category is dynamically updated through sliding average;

[0030] For the The first iteration The center vector of each category, For the The first iteration The center vector of each category, is the momentum coefficient, The first The number of samples in each category, The first The compressed features of samples;

[0031] The loss function combines cross entropy loss and contrast center loss to achieve multi-objective optimization. The calculation method is expressed as:

[0032] ,

[0033] Where, is the cross entropy loss, and L is the loss function.

[0034] Furthermore, the feature encoding function in S405 is implemented by using a ResNet-50 backbone network, removing the global average pooling layer and then connecting an adaptive feature fusion module, which is expressed as:

[0035] ,

[0036] Where, is the Gaussian error linear unit activation function, for Convolutional layer operation.

[0037] The beneficial effects of the present invention are as follows: (1) The present invention adopts an adaptive normalization method to fuse local illumination statistics with global channel attention, which can effectively solve the problem that traditional normalization methods easily cause feature distribution deviation in dynamic illumination environments, thereby enhancing the stability of feature expression. (2) The present invention adopts a dynamic multi-core feature pyramid, and adopts a deformable convolution kernel group and scale-aware attention on the basis of the feature pyramid network. The deformable convolution is used to adaptively adjust the receptive field, and the scale attention is combined to dynamically fuse multi-scale features. The receptive field can be adaptively adjusted, and multi-scale features can be dynamically fused to improve the ability to capture cross-scale spatial correlations in the image. (3) The present invention adopts a differential maximum pooling strategy, and by combining maximum pooling and local standard deviation, the differential expression of subtle features is enhanced, thereby improving the ability to capture details. (4) The present invention adopts an illumination-aware activation function, and by normalizing the illumination intensity, it can dynamically adapt to illumination changes and improve the accuracy of image classification. (5) The present invention adopts a spatial-aware DropBlock method, and by dynamically generating an occlusion mask, it discards low-importance regional features, thereby enhancing the robustness of the model. (6) The present invention uses a tiled adaptive compression module to combine image segmentation and compression, improve computational efficiency and memory utilization, and optimize the computational efficiency of high-resolution images. (7) The present invention can effectively enhance the ability to capture fine-grained differences by combining intra-class aggregation and inter-class separation through contrastive center loss. (8) The loss function in the present invention combines cross-entropy loss and contrastive center loss, and adopts a multi-objective optimization strategy, which can balance classification accuracy and inter-class separation during training, thereby improving the overall performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Schematic diagram of illumination stability analysis of channel-space dual-domain adaptive normalization in the present invention.

[0039] Figure 2 It is a radar chart comparing the performance of multi-scale target detection.

[0040] Figure 3 It is a line chart of feature differentiation feature analysis of the differentiated pooling strategy.

[0041] Figure 4 It is a comparison chart of the distribution of fine-grained classification indicators. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0043] An artificial intelligence-based airport video image classification method of this embodiment includes the following steps:

[0044] S1, conducts full video surveillance of the apron area, including the red line area, stop line and aircraft entry line.

[0045] Utilizing the airport's existing aircraft stand security surveillance cameras, the video surveillance host is the data source for the entire system. Through distributed multi-point installation layout optimization, it collects multi-viewpoint and multi-angle original video images on the apron parking stands. The monitoring software is the display device of the panoramic monitoring system. It receives the panoramic video data stream output by the video processing unit through the network interface, and uses a graphical UI interface to display the 360° panoramic monitoring video screen of the apron parking stands.

[0046] S2, the images obtained by the video monitoring software are stitched and fused in real time using a video image stitching method.

[0047] The video processing unit is the core device of the panoramic monitoring system. It first pulls the video streams from multiple high-definition video surveillance cameras, then performs real-time splicing and fusion of multiple video images, and then outputs the panoramic monitoring video stream data.

[0048] The video image stitching method is:

[0049] S201: Perform feature analysis on the image transmitted by the video surveillance host to obtain multiple frame images corresponding to required features.

[0050] S202 , performing image preprocessing on the acquired image to obtain a target image, where the image preprocessing includes filtering, denoising, and color difference optimization.

[0051] S203. Perform feature extraction and feature matching on the target image obtained after preprocessing, determine the matching feature points of two adjacent images, align the target image through two-dimensional stitching or video projection, and finally draw it into a large scene image or project the target image onto a sphere or cylinder after matching and stitching for drawing.

[0052] S3, separating the panoramic video image frame by frame to obtain each image, namely, a panoramic image of the apron parking space.

[0053] S4, classify the panoramic image of the apron parking space.

[0054] Panoramic image classification uses an illumination-aware multi-scale feature pyramid network as its core architecture. The training process is as follows:

[0055] S401. Adopt a channel-space dual-domain adaptive normalization method to avoid feature distribution offset caused by significant illumination changes in the panoramic image of the apron parking space.

[0056] Apron parking stand images are subject to strong illumination variations and background noise. Conventional normalization methods (such as Layer-Norm, Batch-Norm, and RMS-Norm) are prone to feature distribution shift in dynamic lighting scenarios. This embodiment adopts a channel-space dual-domain adaptive normalization method. By fusing local illumination statistics with global channel attention, it addresses the feature distribution shift caused by significant illumination variations in panoramic apron parking stand images and enhances the stability of feature expression.

[0057] The channel-space dual-domain adaptive normalization method in S401 is:

[0058] ,

[0059] In the above formula, is the normalized output of the cth channel of the panoramic image of the apron parking space, is the cth channel of the input feature map, is the channel adaptive scaling factor, which is learned from the global average pooling through a multi-layer perceptron. It dynamically adjusts the light sensitivity of different channels, enhances the response of key channels, and weakens redundant background channels. is the local illumination mean, is the global standard deviation, is the spatial adaptive bias term, Convolution extracts spatial information from the local illumination mean and compensates for the loss of local features caused by illumination occlusion or noise.

[0060] GAP is a global average pooling operation.

[0061] for The local average pooling operation captures the local illumination changes of the apron image and avoids the sensitivity of global normalization to dynamic illumination.

[0062] It is a statistical function for calculating the global standard deviation, which characterizes the global distribution stability of features within the channel and suppresses the feature distribution offset caused by uneven illumination.

[0063] for Convolutional layer operation.

[0064] S402, using a dynamic multi-core feature pyramid, using a deformable convolution kernel group and scale-aware attention on the basis of the feature pyramid network, adaptively adjusting the receptive field through deformable convolution, and dynamically fusing multi-scale features in combination with scale attention, expressed as:

[0065] ,

[0066] In the above formula, is the multi-scale feature fusion output, s is an element in the multi-scale convolution kernel set, S is the multi-scale convolution kernel set, and the multi-kernel group combines deformable convolution to solve the problem of insufficient coverage of multi-scale targets by a single fixed receptive field. Calculate the scale attention weight function, characterize the importance weights of features at different scales, dynamically allocate the importance of features at different scales, and suppress irrelevant scale interference. is a feature map that has undergone channel-space dual-domain adaptive normalization. For the deformable convolution operation, perform the convolution calculation and adjust the position of the receptive field according to the offset. is the size of the sth convolution kernel, is the normalized output of the cth channel of the panoramic image of the apron parking space, The offset of the deformable convolution kernel is used to represent the spatial offset of the deformable convolution kernel. The offset is predicted by the multi-layer perceptron and the convolution kernel sampling position is adaptively adjusted.

[0067] MLP is a multi-layer perceptron with one hidden layer. is a global average pooling operation with a scale of s.

[0068] is the Sigmoid activation function, for Convolutional layer operation, is the maximum pooling operation with a step size of s.

[0069] The deformable convolution operation is expressed as:

[0070] ,

[0071] In the above formula, is the convolution kernel weight, For the The offset of the position.

[0072] Given the multi-scale characteristics of parking stand targets (such as nearby vehicles and distant signs), conventional single-scale convolution kernels find it difficult to effectively capture cross-scale spatial correlations. This embodiment can solve the feature extraction problem caused by the coexistence of multi-scale targets in panoramic images of apron parking stands.

[0073] S403: Adopting a maximum pooling strategy based on difference to enhance feature differences during the maximum pooling process.

[0074] The differential maximum pooling strategy is expressed as:

[0075] ,

[0076] In the above formula, is the feature map after the maximum difference pooling, is the multi-scale feature fusion output, is the maximum pooling operation, is a positive integer, is a positive integer, The adaptive weighting coefficient dynamically adjusts the standard deviation pooling weight through the light perception activation function, while avoiding the distortion of variance calculation in the strong light area. To calculate the local standard deviation pooling function, it strengthens the expression of subtle differences in similar background areas and enhances the capture of subtle differences in texture;

[0077] H is the feature map height, W is the feature map width, For the feature map The eigenvalues ​​of the position, is the function for calculating the mean;

[0078] LIFA is a light-aware activation function that normalizes the light intensity to avoid the gradient saturation problem of the simple ReLU activation function in strong light areas.

[0079] Traditional activation functions (such as the Sigmoid activation function and the ReLU activation function) are prone to gradient saturation in areas with strong illumination. This embodiment uses an illumination-aware activation function. By employing an illumination intensity normalization term, it reduces the nonlinear intensity in bright areas and enhances the feature response in dark areas. This dynamically adapts to the significant illumination changes in the panoramic image of the apron parking bays. This is expressed as:

[0080] ,

[0081] In the above formula, is the input of the illumination perception activation function, which is represented in the function of calculating the local standard deviation pooling. , is the hyperbolic tangent function, is the element-wise multiplication symbol, It is an adaptive adjustment coefficient that dynamically controls the nonlinear intensity, enhances the response in the dark area, and suppresses over-activation in the bright area. To calculate the function of local light intensity, Gaussian blur kernel calculation, smoothing illumination mutations, In order to prevent the small constant term from dividing by zero, it is preset manually, such as 0.0001. is the Sigmoid activation function, the input of the Sigmoid activation function right The attention weights of the anchor points are weighted summed with the corresponding Gaussian heat map to generate a topology-aware attention mask. The Sigmoid activation function maps the weighted sum to , as the spatial attention coefficient, strengthens the feature response of the key area and suppresses background noise, For The Gaussian heat map centered on ,characterizes the anchor point influence area, and the apron parking space has a strong spatial topological structure. is the number of predefined topological anchor points, i.e. the preset number of key areas in the apron parking image. For the The attention weights of dynamic anchor points, is the convolution kernel weight matrix;

[0082] MLP is a multi-layer perceptron with one hidden layer;

[0083] for Gaussian blur kernel, characterizing the spatial smoothing of local light intensity, Represents the convolution operation;

[0084] For tensor splicing operations, the global average pooling features and predefined anchor coordinates are spliced ​​as joint inputs, and global statistics and geometric priors are fused through splicing. For the predefined Topological anchor point coordinates;

[0085] Indicates that the Gaussian heat map is The horizontal coordinate index value corresponding to the eigenvalue of the position, Indicates that the Gaussian heat map is The vertical coordinate index value corresponding to the eigenvalue of the position, For Gaussian heat map The characteristic value of the position, the Gaussian heat map strengthens the characteristic response of the anchor point neighborhood and improves the sensitivity to local subtle differences. For the The horizontal coordinates of the predefined topological anchor points are initialized by the analysis or statistical distribution of the apron CAD drawings. For the The horizontal coordinate offset of the anchor point is calculated by Convolution learns from input features, For the The vertical coordinates of the predefined topological anchor points are initialized in the same way , For the The vertical coordinate offset of each anchor point is calculated in the same way , is the Gaussian kernel standard deviation, which is set to 1.0 by default;

[0086] The offset of the anchor point position is generated by a learnable parameter. The offset allows the anchor point to move according to the image content, avoiding the misalignment problem caused by the rigid prior.

[0087] Gaussian blur kernel is generated using a two-dimensional Gaussian distribution function The calculation method of the elements in the convolution kernel weight matrix is ​​expressed as:

[0088] ,

[0089] In the above formula, is the Gaussian kernel standard deviation, which is set to 1.0 by default. For Gaussian heat map The eigenvalue of the location.

[0090] There are a large number of similar background areas (such as runway textures) in parking stand images. Conventional spatial pooling operations easily lead to the loss of fine-grained features. This application fuses the maximum value and standard deviation statistics to retain significant features while enhancing the expression of local differences, thereby improving the ability to distinguish scenes with high background similarity in panoramic images of apron parking stands.

[0091] S404, using spatially aware DropBlock, dynamically generates an occlusion mask based on feature importance, and generates a mask by combining feature amplitude and gradient significance, expressed as:

[0092] ,

[0093] In the above formula, The feature map is at position If the importance of the feature at that position is low, its value is set to 0, otherwise it is set to 1, indicating the validity of the feature at that position, aiming to discard the low-importance area. Sort by importance, To calculate the function of feature importance, the feature amplitude and gradient significance are combined to identify the key areas. The maximum pooling feature map of the difference is The eigenvalues ​​of the position, To reduce the discard ratio, we specifically discard low-importance areas to improve the model's robustness to occlusion. Set to 0.3, is the feature map height, W is the feature map width;

[0094] Feature location The neighborhood window of is the neighborhood of the difference maximum pooling feature map eigenvalues, L is the loss function, is the symbol of partial derivative;

[0095] To address the high computational cost of high-resolution images, this embodiment uses a tiled adaptive compression module for image segmentation and compression, thereby achieving more efficient computation and memory optimization, as expressed as:

[0096] ,

[0097] In the above formula, for Convolutional layer operation, for Convolutional layer operation, is the Hadamard product symbol, is the feature map height, is the feature map width, is the compression feature of the sample, is the number of image slices, that is, the total number of 32×32 slices. If H and W are not integer multiples of 32, the edge slices are allowed to be reduced in size (for example, the last slice is ), or fill in 0 eigenvalues ​​to make up an integer multiple of 32, is the mask value of the feature map, its position The mask value is , dynamic masking avoids random discarding and destroying key features, is a function that generates a slice attention mask, expressed as the attention weight matrix of the rth slice.

[0098] To address the issue of parking stand occlusion (such as parking lines partially obscured by vehicles), the conventional Dropout method randomly discards features, which can easily lead to the loss of key spatial information. This embodiment specifically discards features in low-importance areas during training, improving the model's robustness to occlusion interference in panoramic images of apron parking stands.

[0099] S405. Using contrastive center loss, we enhance the ability to capture fine-grained differences by combining intra-class aggregation and inter-class separation constraints.

[0100] The contrastive center loss is expressed as:

[0101] ,

[0102] In the above formula, To compare the center loss, we enhance the ability to distinguish subtle differences by intra-class aggregation and inter-class separation. N is the number of batch samples. For the The compressed features of samples, for norm, is the feature encoding function, representing the deep feature map extracted by the backbone network, For the The center vector of each category, To balance the hyperparameters, To obtain the maximum value function, is the nearest negative class center, used to enforce inter-class separation, m is the interval threshold, representing the minimum inter-class distance, constraining the minimum spacing of fine-grained categories (such as distinguishing adjacent parking spaces); Set to 0.2, m to 0.5,

[0103] To find the nearest negative class among all center vectors of non-current categories, For the The center vector of each category is dynamically updated by sliding average.

[0104] For the The first iteration The center vector of each category, For the The first iteration The center vector of each category, is the momentum coefficient, which is set to 0.9 by default. The first The number of samples in each category, The first The compressed features of samples;

[0105] The loss function combines cross entropy loss and contrast center loss to achieve multi-objective optimization. The calculation method is expressed as:

[0106] ,

[0107] Where, is the cross entropy loss, L is the loss function;

[0108] The feature encoding function in S405 is implemented by using a ResNet-50 backbone network, removing the global average pooling layer and then connecting an adaptive feature fusion module, which is expressed as:

[0109] ,

[0110] Where, is the Gaussian error linear unit activation function, for Convolutional layer operation.

[0111] When calculating the loss function, traditional cross entropy loss is difficult to cope with the subtle differences between parking slot categories.

[0112] S406: Repeat the steps S401 to S405 until the preset stop iteration condition is met, indicating that the model training is completed. 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.

[0113] After model training is completed, the trained model is used to classify panoramic images of apron parking stands. If the goal is to identify specific parking stand numbers or logos, the classification category can be the unique identifier of each parking stand (for example, "parking stand 1", "parking stand 2"); if the goal is to monitor the status of the parking stand (such as whether it is free or occupied), the category can be different status labels, such as "free", "occupied", and "under maintenance"; if the goal is to distinguish the types of parking stands (for example, VIP parking stands and ordinary parking stands), the category label can be different parking stand types.

[0114] It should be noted that the classification category depends on the sample annotation category of the training model, that is, the sample annotation category of the training model is the same as the classification category of the model.

[0115] To verify the effectiveness of this technology, the following experiments were conducted:

[0116] like Figure 1 As shown in the figure, this experiment aims to verify the feature stability of the channel-space dual-domain adaptive normalization method in dynamic lighting scenes. By comparing traditional methods such as batch normalization, layer normalization, and instance normalization, the classification performance changes under different lighting intensities and feature distribution offsets are simulated. The experimental results show that with the drastic fluctuations in light intensity and the increase in feature offset, the accuracy of the traditional method shows a significant oscillation and decline. However, this technology can maintain a stable feature distribution under extreme lighting interference by fusing local lighting statistics with global channel attention. The accuracy attenuation is significantly smaller than that of other methods, especially in high-intensity lighting areas, it shows stronger anti-interference ability, which effectively solves the problem of feature offset caused by dynamic lighting.

[0117] like Figure 2 As shown in the figure, this experiment focuses on evaluating the feature extraction capability of dynamic multi-kernel feature pyramid for multi-scale targets. The comparison objects include feature pyramid network based on fixed convolution kernel, high-resolution network and current mainstream visual Transformer architecture. For cross-scale targets such as close vehicles and medium and long-distance signs in airport scenes, traditional methods are limited by fixed receptive field, resulting in insufficient feature extraction of small targets. However, this technology adaptively adjusts the sampling position through deformable convolution kernel group, combines scale attention to dynamically fuse multi-level features, and shows more balanced detection performance in complex scenes such as occluded targets and low-light targets. In particular, it has significant advantages in retaining the detailed features of small targets, verifying the effectiveness of the multi-scale dynamic perception architecture.

[0118] like Figure 3 As shown in the figure, this experiment focuses on the feature discrimination ability of the difference maximum pooling strategy in scenes with highly similar backgrounds, and compares it with conventional maximum pooling, average pooling and hybrid pooling methods. In images with high background similarity such as runway textures and repetitive signs, traditional pooling operations are prone to lose subtle difference features. This technology enhances the expression of subtle differences such as texture wear and color gradients by fusing the maximum response and standard deviation statistics of local areas. The experimental results show that as the background similarity increases, the feature discrimination index of this method decays significantly slower than other methods, proving that it can effectively capture discriminative details in similar backgrounds and improve the classification model's ability to identify similar scenes.

[0119] like Figure 4 As shown in the figure, this experiment explores the optimization effect of the contrast center loss function on fine-grained classification tasks. The comparison benchmarks include conventional cross entropy loss, triple loss, and sub-center loss. For fine-grained categories with small visual differences such as parking space numbers and signboard text, traditional loss functions are difficult to fully constrain the compactness of intra-class features and the degree of separation between classes. This technology jointly optimizes intra-class aggregation and inter-class exclusion constraints to make the distribution of similar samples in the feature space denser and the boundaries between heterogeneous classes clearer. It shows stronger discrimination ability in classification tasks where samples and edge cases are difficult to distinguish, and significantly improves the model's perception sensitivity to subtle differences and classification robustness.

[0120] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.

Claims

1. An airport video image classification method based on artificial intelligence, characterized in that: The following steps are involved: S1, conducts full video surveillance of the apron area, including the red line area, stop line and aircraft approach line; S2, using a video image stitching method to perform real-time stitching and fusion processing on the images obtained by the video surveillance software; S3, separating the panoramic video image frame by frame to obtain each image, i.e., a panoramic image of the apron parking space; S4, for panoramic image classification of apron parking spaces, uses an illumination-aware multi-scale feature pyramid network as its core architecture. The training process is as follows: S401. Adopting a channel-space dual-domain adaptive normalization method to avoid feature distribution shift caused by significant illumination changes in the panoramic image of the apron parking bay; The channel-space dual-domain adaptive normalization method is: ， In the above formula, is the normalized output of the cth channel of the panoramic image of the apron parking space, is the cth channel of the input feature map, is the channel adaptive scaling factor, which is learned from the global average pooling through a multi-layer perceptron. It dynamically adjusts the light sensitivity of different channels, enhances the response of key channels, and weakens redundant background channels. is the local illumination mean, is the global standard deviation, is the spatial adaptive bias term, Convolution extracts spatial information from the local illumination mean to compensate for the loss of local features caused by illumination occlusion or noise; GAP is the global average pooling operation; for Local average pooling operation captures local illumination changes in the apron image and avoids the sensitivity of global normalization to dynamic illumination; A statistical function for calculating the global standard deviation, which characterizes the global distribution stability of features within the channel and suppresses the feature distribution offset caused by uneven illumination; for Convolutional layer operation; S402: Adopting a dynamic multi-core feature pyramid, based on the feature pyramid network, adopting a deformable convolution kernel group and scale-aware attention, adaptively adjusting the receptive field through deformable convolution, and dynamically fusing multi-scale features in combination with scale attention; S403, adopting the maximum pooling strategy of difference to enhance the feature difference in the maximum pooling process; The differential maximum pooling strategy is expressed as: , In the above formula, is the feature map after the maximum difference pooling, is the multi-scale feature fusion output, is the maximum pooling operation, is a positive integer, is a positive integer, The adaptive weighting coefficient dynamically adjusts the standard deviation pooling weight through the light perception activation function, while avoiding the distortion of variance calculation in the strong light area. To calculate the local standard deviation pooling function, it strengthens the expression of subtle differences in similar background areas and enhances the capture of subtle differences in texture; H is the feature map height, W is the feature map width, For the feature map The eigenvalues ​​of the position, is the function for calculating the mean; It is a light-aware activation function that normalizes the light intensity to avoid the gradient saturation problem of the simple ReLU activation function in strong light areas. S404, using spatially aware DropBlock to dynamically generate an occlusion mask based on feature importance, and generating a mask by combining feature amplitude and gradient saliency; S405, using contrastive center loss, by combining intra-class aggregation and inter-class separation constraints to enhance the ability to capture fine-grained differences; S406: Repeat the above steps S401 to S405 until the preset stop iteration condition is met, which means that the model training is completed.

2. The method for classifying airport video images based on artificial intelligence according to claim 1, characterized in that: The S403 also includes a light perception activation function, which uses a light intensity normalization term and is expressed as: ， In the above formula, is the illumination perception activation function, is the input of the light-aware activation function, is the hyperbolic tangent function, is the element-wise multiplication symbol, is the adaptive adjustment coefficient, is the function for calculating the local light intensity, To prevent division by zero errors for small constant terms, For The Gaussian heat map centered on , characterizing the anchor point influence area, is the number of predefined topological anchor points, i.e. the preset number of key areas in the apron parking image. For the predefined Topological anchor point coordinates, For the The attention weights of dynamic anchor points, is the Sigmoid activation function, Anchor point position offset.

3. The method for classifying airport video images based on artificial intelligence according to claim 2, characterized in that: The They are: ， In the above formula, is a multi-layer perceptron with one hidden layer, GAP is a global average pooling operation; is a 5×5 Gaussian blur kernel, represents the convolution operation, is the convolution kernel weight matrix; is the tensor concatenation operation, For the predefined Topological anchor point coordinates; Indicates that the Gaussian heat map is The horizontal coordinate index value corresponding to the eigenvalue of the position, Indicates that the Gaussian heat map is The vertical coordinate index value corresponding to the eigenvalue of the position, For Gaussian heat map The eigenvalues ​​of the position, For the The horizontal coordinates of the predefined topological anchor points are initialized by the analysis or statistical distribution of the apron CAD drawings. For the The horizontal coordinate offset of the anchor point is calculated by Convolution learns from input features, For the The vertical coordinates of the predefined topological anchor points are initialized in the same way , For the The vertical coordinate offset of each anchor point is calculated in the same way , is the standard deviation of the Gaussian kernel; is the anchor point position offset, for Convolutional layer operation.

4. The method for classifying airport video images based on artificial intelligence according to claim 3 is characterized in that: The Gaussian blur kernel is generated using a two-dimensional Gaussian distribution function The calculation method of the elements in the convolution kernel weight matrix is ​​expressed as: ， In the above formula, is the Gaussian kernel standard deviation, For Gaussian heat map The eigenvalues ​​of the position, is a positive integer, Is a positive integer.

5. The method for classifying airport video images based on artificial intelligence according to claim 1, characterized in that: The occlusion mask in S404 is expressed as: ， In the above formula, The feature map is at position The mask value of Sort by importance, is the function for calculating feature importance, The maximum pooling feature map of the difference is The eigenvalues ​​of the position, is the discard ratio, is the feature map height, is the feature map width; Feature location The neighborhood window of is the neighborhood of the difference maximum pooling feature map eigenvalues, L is the loss function, is the symbol of the partial derivative.

6. The method for classifying airport video images based on artificial intelligence according to claim 1, characterized in that: The S404 also includes using a slice adaptive compression module to perform image segmentation and compression, which is expressed as: , In the above formula, for Convolutional layer operation, for Convolutional layer operation, is the Hadamard product symbol, is the feature map height, is the feature map width, is the compression feature of the sample, is the number of image slices, is the mask value of the feature map, its position The mask value is , is the function that generates the slice attention mask, expressed as the attention weight matrix of the rth slice, is the illumination-aware activation function.

7. The method for classifying airport video images based on artificial intelligence according to claim 1, characterized in that: The contrast center loss in S405 is expressed as: ， In the above formula, is the comparison center loss, N is the number of batch samples, For the The compressed features of samples, for norm, is the feature encoding function, For the The center vector of each category, To balance the hyperparameters, To obtain the maximum value function, is the nearest negative class center, and m is the interval threshold; To find the nearest negative class among all center vectors of non-current categories, For the The center vector of each category is dynamically updated through sliding average; For the The first iteration The center vector of each category, For the The first iteration The center vector of each category, is the momentum coefficient, The first The number of samples in each category, The first The compressed features of samples; The loss function combines cross entropy loss and contrast center loss to achieve multi-objective optimization. The calculation method is expressed as: ， Where, is the cross entropy loss, and L is the loss function.

8. The method for classifying airport video images based on artificial intelligence according to claim 7, characterized in that: The feature encoding function in S405 is implemented by using a ResNet-50 backbone network, removing the global average pooling layer and then connecting an adaptive feature fusion module, which is expressed as: ， Where, is the Gaussian error linear unit activation function, for Convolutional layer operation.

Images