Airport video image classification method based on artificial intelligence

Abstract

The invention discloses an airport video image classification method based on artificial intelligence, and relates to the technical field of video image classification, and the method comprises the steps: carrying out the whole-course video monitoring of an airport apron region, including a red line region, a stop line and an airplane entry line; carrying out real-time splicing fusion processing on pictures obtained by the video monitoring software by using a video image splicing method; performing frame-by-frame separation on the panoramic video image to obtain each image, namely an airport apron gate panoramic image; a panoramic image of an airport apron parking space is classified, and an illumination perception multi-scale feature pyramid network is used as a core architecture. According to the method, the stability of feature expression is enhanced by fusing local illumination statistics and global channel attention by adopting an adaptive normalization method. According to the method, the deformable convolution kernel group and the scale are adopted to perceive attention on the basis of the feature pyramid network, the receptive field can be adaptively adjusted, the multi-scale features are dynamically fused, and the cross-scale space correlation capturing capability in the image is improved.

Description

Technical Field

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

[0002] In the apron environment, apron images usually have complex characteristics, including strong illumination changes, background noise interference, coexistence of multi-scale targets, and the presence of similar background regions. These factors pose many challenges to traditional image classification methods when dealing with apron images.

[0003] Common illumination changes in apron images can lead to large differences in image brightness. This change directly affects the stability of image features and thus affects classification accuracy. Conventional normalization methods (such as Layer-Norm, Batch-Norm, etc.) are often difficult to effectively solve the problem of feature distribution shift in dynamic illumination scenarios, and thus cannot well adapt to the classification requirements in different illumination environments; there are often multi-scale targets in apron images, such as vehicles nearby and signs in the distance. The scale differences of these targets make it difficult for traditional convolutional neural networks to extract cross-scale spatial features. A single fixed receptive field cannot cover all scales of targets, resulting in incomplete and inaccurate feature extraction; apron images often have similar background regions (such as runway textures). This background similarity makes it difficult to distinguish fine-grained features, and traditional spatial pooling operations are prone to detail loss, thereby reducing the accuracy and robustness of the classification model; occlusion often occurs in apron images, for example, a vehicle occludes part of the parking line or sign. This occlusion not only affects the integrity of the target but also makes it difficult for traditional methods such as Dropout to effectively handle the problem of feature loss, thus 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 above-mentioned disadvantages of the prior art and provide an airport video image classification method based on artificial intelligence.

[0005] The technical solution adopted to solve the above technical problem is: an airport video image classification method based on artificial intelligence, including the following steps: S1. Conduct full - process video monitoring on the apron area, including the red - line area, stop line, and aircraft entry line; S2. Use the video image stitching method to perform real - time stitching and fusion processing on the images obtained by the video monitoring software; S3. Separate each frame of the panoramic video image to obtain each image, that is, the panoramic image of the apron parking position; S4. Classify the panoramic images of the apron parking positions. Use the light - aware multi - scale feature pyramid network as the core architecture. The training process is as follows: S401. Adopt the channel - space dual - domain adaptive normalization method to avoid the feature distribution deviation caused by the significant change in illumination in the panoramic images of the apron parking positions; S402. Adopt a dynamic multi - core feature pyramid. On the basis of the feature pyramid network, use a deformable convolution kernel group and scale - aware attention. Adaptively adjust the receptive field through deformable convolution, and dynamically fuse multi - scale features by combining scale attention; S403. Adopt the maximum difference pooling strategy to strengthen the feature difference during the maximum pooling process; S404. Adopt spatial - aware DropBlock, dynamically generate occlusion masks according to feature importance, and generate masks by combining feature amplitude and gradient saliency; S405. Adopt contrastive center loss, and enhance the ability to capture fine - grained differences by jointly constraining intra - class aggregation and inter - class separation; S406. Repeat the above steps until the preset stop - iteration condition is met, indicating that the model training is completed.

[0006] Further, the S403 also includes a light - aware activation function, which adopts a light - intensity normalization term, expressed as: , In the above formula, is the light - aware 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 local light intensity, is a small constant term to prevent division - by - zero errors, is the Gaussian heatmap centered at representing the anchor - point influence area, is the predefined number of topological anchor points, that is, the preset number of key areas in the apron parking - position image, is the predefined th topological anchor - point coordinate, is the th dynamic anchor - point attention weight, is the Sigmoid activation function, is the anchor - point position offset.

[0007] Further, the are respectively: , In the above formula, is a multi-layer perceptron with one hidden layer, and GAP is a global average pooling operation; is a 5×5 Gaussian blur kernel, represents a convolution operation; is the convolution kernel weight matrix; is a tensor concatenation operation, is the predefined th topological anchor coordinate; represents the abscissa index value corresponding to the eigenvalue of the Gaussian heat map at the position, represents the ordinate index value corresponding to the eigenvalue of the Gaussian heat map at the position, is the eigenvalue of the Gaussian heat map at the position, is the th predefined topological anchor's abscissa, initialized by parsing the apron CAD drawing or statistical distribution, is the th anchor's abscissa offset, learned from the input features through convolution, is the th predefined topological anchor's ordinate, initialized in the same way as , is the th anchor's ordinate offset, calculated in the same way as , is the Gaussian kernel standard deviation; is the anchor position offset, is the convolution layer operation.

[0008] Furthermore, the above-mentioned Gaussian blur kernel uses a two-dimensional Gaussian distribution function to generate the convolution kernel, and the calculation method of the elements in its weight matrix is expressed as: , In the above formula, is the Gaussian kernel standard deviation, is the eigenvalue of the Gaussian heat map at the position, is a positive integer, is a positive integer.

[0009] Furthermore, the occlusion mask in S404 is expressed as: , In the above formula, is the mask value of the feature map at position , is the importance ranking, is the function for calculating feature importance, is the feature value of the maximum-difference pooling feature map at position, is the dropout ratio, is the height of the feature map, and W is the width of the feature map; is the neighborhood window of the feature position , is the -th feature value within the neighborhood of the maximum-difference pooling feature map, L is the loss function, is the partial derivative symbol.

[0010] Furthermore, the S404 further includes performing image segmentation and compression using a piecewise adaptive compression module, expressed as: , In the above formula, is the convolutional layer operation, is the convolutional layer operation, is the Hadamard product symbol, is the height of the feature map, is the width of the feature map, is the compressed feature of the sample, is the number of image pieces, is the mask value of the feature map, and its mask value at position is , is the function for generating a piecewise attention mask, expressed as the attention weight matrix of the r-th piece, is the light perception activation function.

[0011] Furthermore, the contrastive center loss in the S405 is expressed as: , In the above formula, is the contrastive center loss, N is the number of batch samples, is the compressed feature of the -th sample, is the norm, is the feature encoding function, is the -th class center vector, For balancing hyperparameters, is the maximum value function, is the nearest negative class center, and m is the margin threshold; is to find the nearest negative class among all center vectors of non-current classes, is the center vector of the th class, which is dynamically updated by moving average; is the th class's center vector in the th iteration, is the th class's center vector in the th iteration, is the momentum coefficient, is the number of samples of the th class in the current batch, is the compressed feature of the th sample in the current batch; , In the formula, is the cross-entropy loss, and L is the loss function.

[0012] Furthermore, the implementation method of the feature encoding function in S405 is: using a ResNet-50 backbone network, removing the global average pooling layer and then connecting an adaptive feature fusion module, which is expressed as: , In the formula, is the Gaussian error linear unit activation function, is the convolutional layer operation.

[0013] The beneficial effects of the present invention are as follows: (1) By fusing local illumination statistics and global channel attention, the adaptive normalization method of the present invention can effectively solve the problem that traditional normalization methods are prone to cause feature distribution deviation in dynamic illumination environments, and enhance the stability of feature expression. (2) The present invention adopts a dynamic multi-core feature pyramid. Based on the feature pyramid network, a deformable convolution kernel group and scale-aware attention are used. The receptive field is adaptively adjusted through deformable convolution, and multi-scale features are dynamically fused by combining scale attention, which can adaptively adjust the receptive field, dynamically fuse multi-scale features, and improve the ability to capture cross-scale spatial associations in images. (3) The present invention adopts a maximum difference pooling strategy. By combining maximum pooling and local standard deviation, the differential expression of subtle features is strengthened, thereby enhancing the ability to capture details. (4) The present invention adopts an illumination-aware activation function. Through illumination intensity normalization, it can dynamically adapt to illumination changes and improve the accuracy of image classification. (5) The present invention adopts a spatial-aware DropBlock method. By dynamically generating occlusion masks and discarding features in low-importance regions, the robustness of the model is enhanced. (6) The present invention adopts a piecewise adaptive compression module, which combines image segmentation and compression to improve computational efficiency and memory utilization rate, and optimize the computational efficiency of high-resolution images. (7) By comparing the center loss and combining intra-class aggregation and inter-class separation, the present invention can effectively enhance the ability to capture fine-grained differences. (8) The loss function in the present invention combines cross-entropy loss and 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

[0014] Figure 1 FIG. is a schematic diagram of the illumination stability analysis of channel-space dual-domain adaptive normalization in the present invention.

[0015] Figure 2 FIG. is a radar chart for comparing the performance of multi-scale object detection.

[0016] Figure 3 FIG. is a line chart for analyzing the feature differentiation features of the differential pooling strategy.

[0017] Figure 4 FIG. is a comparison chart of the distribution of fine-grained classification metrics. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] In order to make the objectives, technical solutions and advantages of the present invention clearer, 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 used to limit the present invention.

[0019] A method for classifying airport video images based on artificial intelligence in this embodiment includes the following steps: S1. Conduct full-process video monitoring on the apron area, which includes the red line area, the stop line, and the aircraft entry line.

[0020] Utilize the existing apron security monitoring cameras at the airport. The video monitoring host is the data source of the entire system. Through the optimization of the distributed multi-point installation layout, it collects multi-viewpoint and multi-perspective original video images on the apron parking positions. 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 displays the 360° panoramic monitoring video of the apron parking positions using a graphical UI interface.

[0021] S2. Perform real-time stitching and fusion processing on the images obtained by the video monitoring software using the video image stitching method.

[0022] The video processing unit is the core device of the panoramic monitoring system. First, it pulls the video streams of multiple high-definition video monitoring cameras. Secondly, it performs real-time stitching and fusion of multiple video images. Then, it outputs the panoramic monitoring video stream data.

[0023] The video image stitching method is as follows: S201. Analyze the features of the images transmitted by the video monitoring host to obtain multiple frame images corresponding to the required features.

[0024] S202. Perform image preprocessing on the obtained images to obtain the target images. The image preprocessing includes filtering for noise reduction and color difference optimization.

[0025] S203. Extract and match the features of the target images obtained after preprocessing, determine the matching feature points of two adjacent images, align the target images through two-dimensional stitching or video projection-based methods, and finally draw a large-scale scene image or project the target images after matching and stitching onto a spherical or cylindrical surface for drawing.

[0026] S3. Separate each frame of the panoramic video image to obtain each image, that is, the panoramic image of the apron parking position.

[0027] S4. Classify the panoramic images of the apron parking positions.

[0028] The panoramic image classification uses the light perception multi-scale feature pyramid network as the core architecture. The training process is as follows: S401. Adopt the channel-spatial dual-domain adaptive normalization method to avoid the feature distribution deviation caused by the significant change in illumination in the panoramic images of the apron parking positions.

[0029] The apron parking position image has strong illumination changes and background noise interference. Conventional normalization methods (such as Layer-Norm normalization, Batch-Norm normalization, and RMS-Norm normalization, etc.) are prone to cause feature distribution deviation in dynamic illumination scenarios. In this embodiment, a channel-spatial dual-domain adaptive normalization method is adopted. By fusing local illumination statistics and global channel attention, the problem of feature distribution deviation caused by significant illumination changes in the panoramic image of the apron parking position is solved, and the stability of feature expression is enhanced.

[0030] The channel-spatial dual-domain adaptive normalization method in S401 is as follows: , In the above formula, is the normalized output of the c-th channel of the panoramic image of the apron parking position, is the c-th channel of the input feature map, is the channel adaptive scaling coefficient. The channel adaptive scaling coefficient is learned from global average pooling through a multi-layer perceptron, dynamically adjusted for the illumination sensitivity of different channels, enhancing the response of key channels and weakening redundant background channels. is the local illumination mean, is the global standard deviation, is the spatial adaptive bias term, through convolution extracts spatial information from the local illumination mean to compensate for local feature loss caused by illumination occlusion or noise.

[0031] GAP is the global average pooling operation.

[0032] is the local average pooling operation, capturing the local illumination changes in the apron image and avoiding the sensitivity of global normalization to dynamic illumination.

[0033] is the statistical function for calculating the global standard deviation, characterizing the global distribution stability of features within the channel and suppressing the feature distribution deviation caused by uneven illumination.

[0034] is the convolutional layer operation.

[0035] S402. Adopt a dynamic multi-core feature pyramid. Based on the feature pyramid network, a deformable convolution kernel group and scale-aware attention are adopted. The receptive field is adaptively adjusted through deformable convolution, and multi-scale features are dynamically fused in combination with scale attention, expressed as: , In the above formula, It is the output of multi-scale feature fusion. s is an element in the set of multi-scale convolution kernels, and S is the set of multi-scale convolution kernels. The multi-kernel group joint deformable convolution solves the problem of insufficient coverage of multi-scale targets by a single fixed receptive field. A function for calculating scale attention weights, which represents the importance weights of features at different scales, dynamically assigns the importance of features at different scales, and suppresses interference from irrelevant scales. It is the feature map after channel-space double-domain adaptive normalization. It is a deformable convolution operation that performs convolution calculation and adjusts the position of the receptive field according to the offset. It is the size of the s-th convolution kernel. It is the normalized output of the c-th channel of the apron parking position panoramic image. It is the deformable convolution offset, which represents the spatial offset of the deformable convolution kernel. The offset is predicted by a multi-layer perceptron to adaptively adjust the sampling position of the convolution kernel.

[0036] MLP is a multi-layer perceptron with one hidden layer. It is the global average pooling operation with scale s.

[0037] It is the Sigmoid activation function. It is Convolution layer operation. It is the max pooling operation with stride s.

[0038] The deformable convolution operation is expressed as: , In the above formula, It is the convolution kernel weight. It is the Offset at the i-th position.

[0039] For the multi-scale characteristics of parking position targets (such as nearby vehicles and distant signs), conventional single-scale convolution kernels are difficult to effectively capture cross-scale spatial correlations. This embodiment can solve the feature extraction problem brought about by the coexistence of multi-scale targets in the apron parking position panoramic image.

[0040] S403. Adopt the maximum difference pooling strategy to strengthen feature differences during the max pooling process.

[0041] The maximum difference pooling strategy is expressed as: , In the above formula, It is the feature map after maximum difference pooling. It is the output of multi-scale feature fusion. It is the max pooling operation. It is a positive integer. is a positive integer, is an adaptive weighting coefficient, which dynamically adjusts the standard deviation pooling weights through a light perception activation function, while avoiding the distortion of variance calculation in strong light areas. is a function for calculating local standard deviation pooling, which strengthens the expression of subtle differences in background similar areas and enhances the capture of subtle texture differences; H is the height of the feature map, and W is the width of the feature map. at the position is the feature value, is a function for calculating the mean value; LIFA is a light perception activation function, which avoids the gradient saturation problem of the simple ReLU activation function in strong light areas through light intensity normalization.

[0042] Traditional activation functions (such as the Sigmoid activation function and the ReLU activation function) are prone to gradient saturation in strong light areas. In this embodiment, a light perception activation function is adopted. By using a light intensity normalization term, the nonlinear intensity is reduced in bright areas, and the feature response is enhanced in dark areas, dynamically adapting to the significant characteristics of the illumination change in the panoramic image of the apron parking position, which is expressed as: , In the above formula, is the input of the light perception activation function, which is expressed as in the function for calculating local standard deviation pooling. is the hyperbolic tangent function, is the element multiplication symbol, is an adaptive adjustment coefficient, which dynamically controls the nonlinear intensity, enhances the response in dark areas, and suppresses over-activation in bright areas. is a function for calculating local light intensity, which is calculated through a Gaussian blur kernel to smooth the illumination mutation. is a small constant term to prevent division by zero error, which is preset artificially, such as set to 0.0001. is the Sigmoid activation function, and the input of the Sigmoid activation function performs weighted summation on the attention weights of anchor points and 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 key areas and suppresses background noise. is the Gaussian heat map centered on , which represents the influence area of the anchor point. The apron parking position has a strong spatial topology structure. is the predefined number of topology anchor points, that is, the preset number of key areas in the apron parking position image. is the attention weight of the th dynamic anchor point, is the convolutional kernel weight matrix; MLP is a multi-layer perceptron with one hidden layer; is the Gaussian blur kernel, representing the spatial smoothing process of local illumination intensity, represents the convolution operation; is the tensor concatenation operation, which concatenates the global average pooling feature and the predefined anchor point coordinates into a joint input, and fuses the global statistics and geometric priors through concatenation, is the predefined th topological anchor point coordinate; represents the abscissa index value corresponding to the eigenvalue of the Gaussian heatmap at the position, represents the ordinate index value corresponding to the eigenvalue of the Gaussian heatmap at the position, is the eigenvalue of the Gaussian heatmap at the position. The Gaussian heatmap enhances the feature response in the anchor neighborhood and improves the sensitivity to local subtle differences, is the th abscissa of the predefined topological anchor point, initialized by parsing the apron CAD drawing or statistical distribution, is the th abscissa offset of the anchor point, learned from the input features through convolution, is the th ordinate of the predefined topological anchor point, initialized in the same way as , is the th ordinate offset of the anchor point, calculated in the same way as , is the standard deviation of the Gaussian kernel, default set to 1.0; is the anchor point position offset, generated by learnable parameters. The offset allows the anchor point to move according to the image content, avoiding the misalignment problem caused by the rigid prior; The Gaussian blur kernel is generated using the two-dimensional Gaussian distribution function The convolutional kernel, the calculation method of the elements in its weight matrix is expressed as: , In the above formula, is the standard deviation of the Gaussian kernel, default set to 1.0, is the eigenvalue of the Gaussian heatmap at the position.

[0043] There are a large number of similar background regions in the apron image (such as runway texture). Conventional spatial pooling operations are prone to losing fine-grained features. In this application, by fusing the maximum value and the standard deviation statistic, while retaining significant features, the local differential expression is enhanced, and the discrimination ability for scenes with high background similarity in the panoramic image of the apron parking position is improved.

[0044] S404. Adopt spatial-aware DropBlock, dynamically generate an occlusion mask according to the feature importance, and generate a mask by jointly using the feature amplitude and gradient saliency, which is expressed as: , In the above formula, is the mask value of the feature map at the position . If the importance of the feature at this position is low, its value is set to 0, otherwise it is set to 1, indicating the validity of the feature at this position, aiming to discard low-importance regions. is the importance ranking, is a function for calculating feature importance, which jointly uses the feature amplitude and gradient saliency to identify key regions. is the feature value of the maximum-difference pooling feature map at the position , is the discard ratio, which specifically discards low-importance regions to improve the robustness of the model to occlusion. is set to 0.3, is the height of the feature map, and W is the width of the feature map; is the neighborhood window of the feature position , is the th feature value in the neighborhood of the maximum-difference pooling feature map, L is the loss function, is the partial derivative symbol; Aiming at the problem of high computational cost for high-resolution images, this embodiment adopts a piecewise adaptive compression module for image segmentation and compression, and then realizes more efficient calculation and memory optimization, which is expressed as: , In the above formula, is the convolutional layer operation, is the convolutional layer operation, is the Hadamard product symbol, is the height of the feature map, is the width of the feature map, is the compressed feature of the sample, is the number of image slices, that is, the total number of slices divided into 32×32 slices. If H and W are not integer multiples of 32, the size of the edge slices is allowed to be reduced (such as the last slice is ), or padding with 0 eigenvalues to make up 32 integer multiples, is the mask value of the feature map, and its position has a mask value of , and the dynamic mask avoids randomly discarding and destroying key features, is a function for generating slice attention masks, represented as the attention weight matrix of the r-th slice.

[0045] Regarding the problem of apron position occlusion (such as the apron line partially occluded by a vehicle), the conventional Dropout method randomly discards features and is likely to lose key spatial information. In this embodiment, features in low-importance regions are specifically discarded during training to improve the robustness of the model to occlusion interference in the panoramic image of the apron parking position.

[0046] S405. Adopt contrastive center loss. By jointly constraining intra-class aggregation and inter-class separation, the ability to capture fine-grained differences is enhanced.

[0047] The contrastive center loss is expressed as: , In the above formula, is the contrastive center loss. By intra-class aggregation and inter-class separation, the ability to distinguish subtle differences is enhanced. N is the number of batch samples, is the compressed feature of the -th sample, is norm, is the feature encoding function, representing the depth feature map extracted by the backbone network, is the center vector of the -th category, is the balancing hyperparameter, is the maximum function, is the nearest negative class center, used to enforce inter-class separation. m is the margin threshold, representing the minimum inter-class distance, and constraining the minimum distance between fine-grained categories (such as distinguishing adjacent apron positions); is set to 0.2 and m is set to 0.5, is to find the nearest negative class among all center vectors of non-current categories, is the center vector of the -th category, dynamically updated by exponential moving average, is the center vector of the -th iteration of the -th category, is the The center vector of the th category in the current iteration, is the momentum coefficient, and the default setting is 0.9. is the number of samples of the th category in the current batch, is the compressed feature of the th sample in the current batch; The loss function combines the cross-entropy loss and the contrastive center loss to achieve multi-objective optimization. The calculation method is expressed as: , In the formula, is the cross-entropy loss, and L is the loss function; The implementation method of the feature encoding function in S405 is as follows: Use the ResNet-50 backbone network, and connect an adaptive feature fusion module after removing the global average pooling layer, which is expressed as: , In the formula, is the Gaussian error linear unit activation function, is the convolutional layer operation.

[0048] When calculating the loss function, the traditional cross-entropy loss is difficult to handle the subtle differences between the parking position categories.

[0049] S406. Repeat the above steps iteratively until the preset stop iteration condition is met, which means the model training is completed. The preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0050] After the model training is completed, use the trained model to implement the classification of the apron parking position panoramic image. If the goal is to identify the specific parking position number or identifier, the classification categories can be the unique identifier of each parking position (such as "Parking Position 1", "Parking Position 2"); if the goal is to monitor the status of the parking position (such as whether it is idle, occupied, etc.), the categories can be different status labels, such as "idle", "occupied", "under maintenance"; if the goal is to distinguish the types of parking positions (for example, VIP parking positions and ordinary parking positions), the category labels can be different parking position types.

[0051] It should be noted that the classification categories depend on the sample annotation categories of the training model, that is, the sample annotation categories of the training model are the same as the categories that the model can classify.

[0052] To verify the effectiveness of this technology, the following experiments are carried out: As Figure 1As shown, this experiment aims to verify the feature stability of the channel-spatial dual-domain adaptive normalization method in dynamic lighting scenarios. 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 as the lighting intensity fluctuates violently and the feature offset increases, the accuracy of traditional methods shows a significant oscillatory decline. However, by integrating local lighting statistics and global channel attention, this technology can still maintain a stable feature distribution under extreme lighting interference, and the attenuation amplitude of the accuracy is significantly smaller than that of other methods. In particular, it shows stronger anti-interference ability in high-intensity lighting areas, effectively solving the feature offset problem caused by dynamic lighting.

[0053] As Figure 2 shown, this experiment focuses on evaluating the feature extraction ability of the dynamic multi-core feature pyramid for multi-scale targets. The comparison objects include the feature pyramid network based on fixed convolutional kernels, the high-resolution network, and the current mainstream vision Transformer architecture. For cross-scale targets such as nearby vehicles and medium- and long-distance signs in airport scenes, traditional methods are limited by the fixed receptive field and result in insufficient feature extraction for small targets. However, this technology adaptively adjusts the sampling positions through a deformable convolutional 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.

[0054] As Figure 3 shown, this experiment focuses on the feature discrimination ability of the maximum difference pooling strategy in scenarios with highly similar backgrounds. It is compared 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 losing subtle difference features. However, by integrating the maximum response and standard deviation statistics of local regions, this technology strengthens the expression of subtle differences such as texture wear and color gradient. The experimental results show that as the background similarity increases, the attenuation trend of the feature discrimination index of this method is significantly slower than that of other methods, proving that it can effectively capture discriminative details in similar backgrounds and improve the discrimination ability of the classification model for similar scenes.

[0055] As Figure 4As shown, this experiment explores and compares the optimization effect of the center loss function on the fine-grained classification task. The comparison benchmarks include methods such as conventional cross-entropy loss, triplet loss, and sub-center loss. For fine-grained categories with tiny visual differences such as parking space numbers and sign texts, traditional loss functions are difficult to fully constrain the intra-class feature compactness and inter-class separation. Through the joint optimization of intra-class aggregation and inter-class repulsion constraints, this technology makes the distribution of samples of the same class in the feature space denser and the boundaries between different classes clearer, showing stronger discriminative ability in the classification tasks of difficult-to-distinguish samples and edge cases, and significantly improving the model's sensitivity to subtle differences and classification robustness.

[0056] The above is only a preferred embodiment of the present invention and is not intended to limit the protection scope 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 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 monitoring 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 the panoramic image classification of the apron parking space, uses the illumination-aware multi-scale feature pyramid network as the core architecture, and the training process is as follows: S401, using 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 space; S402, using dynamic multi-core feature pyramid, using deformable convolution kernel group and scale-aware attention on the basis of feature pyramid network, adaptively adjusting receptive field through deformable convolution, and dynamically fusing multi-scale features in combination with scale attention; S403, adopting the difference maximum pooling strategy to strengthen the feature difference in the maximum pooling process; S404, using spatially aware DropBlock, dynamically generating an occlusion mask based on feature importance, and generating a mask by combining feature amplitude and gradient significance; 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 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 light-aware 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 represents 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 stand 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 is 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 ordinate index value corresponding to the eigenvalue of the position, For Gaussian heat map The eigenvalues ​​of the positions, 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 Convolution learns from input features, For the The ordinate of the predefined topological anchor point is initialized in the same way , For the The vertical coordinate offset of the anchor point is calculated in the same way , is the Gaussian kernel standard deviation; 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 positions, 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 In order of importance, is the function for calculating feature importance, The maximum pooling feature map is the difference in The eigenvalues ​​of the positions, is the discard ratio, is the feature map height, is the feature map width; Characteristic 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.

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 for generating the slice attention mask, expressed as the attention weight matrix of the rth slice, is the lighting-aware activation function.

7. The method for classifying airport video images based on artificial intelligence according to claim 1, characterized in that: The comparison 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 closest 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 iteration The center vector of each category, For the The iteration The center vector of each category, is the momentum coefficient, The current batch The number of samples in each category, The current batch The compression 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: ， In the formula, 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 is 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: ， In the formula, is the Gaussian error linear unit activation function, for Convolutional layer operation.

Images