A Lightweight Aircraft Target Recognition Method Based on Improved DeepLabV3+

Abstract

This invention relates to the field of image recognition technology and discloses a lightweight aircraft target recognition method based on an improved DeepLabV3+ network. The method includes the following steps: constructing a teacher model based on the improved DeepLabV3+ network, the teacher model comprising a heterogeneous hollow residual module, an adaptive scale-aware attention pyramid module, and a decoder; performing multi-objective evolutionary pruning on the heterogeneous hollow residual module in the teacher model based on a dynamic topological mask search algorithm oriented towards encoding / decoding association, thereby obtaining a Pareto optimal candidate student model; automatically searching from the candidate student models using a multi-objective genetic algorithm to find a Pareto optimal lightweight student model that balances accuracy and low latency; and performing knowledge distillation on the teacher model and the Pareto optimal lightweight student model. This invention solves the problems of large differences in aircraft target scale, strong background interference, and the difficulty of deploying high-precision models in real time at the edge in airport surface surveillance.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a lightweight aircraft target recognition method based on an improved DeepLabV3+. Background Technology

[0002] Runway incursion is one of the core hidden dangers threatening the operational safety of civil aviation airports. With the increase in the number of flights, traditional monitoring methods relying on surface surveillance radar (ASDE-X) and controller visual inspection have limitations in terms of coverage of blind spots and all-weather perception capabilities. Surface surveillance technology based on computer vision has become a research hotspot in the construction of smart airports due to its intuitiveness and low cost.

[0003] Currently, deep convolutional neural networks (CNNs) have been widely used in aircraft detection tasks. The DeepLabV3+ network proposed by Chen et al. effectively expands the receptive field through spatial pyramid pooling with holes (ASPP) and performs excellently in semantic segmentation tasks. However, directly applying it to airport surface monitoring faces two major challenges: First, insufficient environmental robustness. In large scenarios like airports, aircraft target pixels account for a small proportion and are easily affected by tarmac markings, shadows, and lighting, leading to false detections and missed detections. Second, high model redundancy. For example, the DeepLabV3+ network proposed by Wang et al. has a large number of parameters and computational cost, making it difficult to adapt to the computing power-constrained airport edge terminals and unable to meet the millisecond-level real-time early warning requirements.

[0004] Attention mechanisms have become the mainstream research approach for feature enhancement and robustness improvement in complex backgrounds. However, existing general attention modules often struggle to balance global context (such as runway layout) and local details (such as wing edges) when dealing with multi-scale aircraft targets unique to airport environments, leading to decreased segmentation accuracy under conditions of strong light or occlusion.

[0005] Current methods for lightweighting models mainly include network pruning and knowledge distillation. Channel pruning, proposed by He et al., effectively compresses the model but often leads to the loss of fine-grained features; knowledge distillation, proposed by Hinton et al., while able to transfer knowledge, relies excessively on the structural design of the teacher network. In recent years, Neural Architecture Search (NAS) has shown potential in finding optimal lightweight structures, but the balance between multiple objectives in semantic segmentation tasks remains to be explored. Summary of the Invention

[0006] The purpose of this invention is to address the challenges of large differences in aircraft target scale, strong background interference, and the difficulty in deploying high-precision models in real time at the edge during airport surface surveillance. It provides a lightweight aircraft target recognition method based on an improved DeepLabV3+. This method constructs a high-precision teacher network by introducing an Adaptive Scale-Aware Attention Pyramid (ASAP) module and an improved Heterogeneous Hollow Residual Module (HARM). A multi-objective evolutionary algorithm is used to automatically search for the optimal pruning strategy for the student model, and knowledge distillation is combined to compensate for the student model's performance, aiming to achieve a synergistic optimization of high precision and latency.

[0007] To achieve the above-mentioned objectives, the embodiments of the present invention provide the following technical solutions:

[0008] The lightweight aircraft target recognition method based on the improved DeepLabV3+ includes the following steps:

[0009] Step 1: Construct a teacher model based on the improved DeepLabV3+ network. The teacher model includes an encoder and a decoder. The encoder includes a heterogeneous hollow residual module and an adaptive scale-aware attention pyramid module.

[0010] The acquired raw image X in Input heterogeneous hole residual module, main path captures the original image X in Main feature U; capturing the original image X from the path in The spatial Laplace gradient is obtained by calculating the Laplace gradient of the principal feature U from the feature V; the positive cross-sectional fusion unit is used to calculate the Laplace gradient of the principal feature U. After processing the main feature U, secondary features V, and the tanh activation function... Element-wise multiplication yields the first fusion feature B. fuse First fusion feature B fuse The second fused feature C is obtained by adding it element-wise to the main feature U. fuse Second fusion feature C fuse Compared with the original image X in Element-wise addition is performed to obtain the low-level detail features F output by the heterogeneous void residual module. backbone Low-level detail features F backbone The high-level semantic features F are obtained by inputting the adaptive scale-aware attention pyramid module. final ;

[0011] Step 2: Based on the dynamic topological mask search algorithm for encoding and decoding association, perform multi-objective evolutionary pruning on the heterogeneous hollow residual modules in the teacher model, which is constrained by the asymmetric topological penalty mechanism, so as to obtain the candidate student model that reaches Pareto optimality.

[0012] Step 3: Use a multi-objective genetic algorithm to automatically search for the Pareto optimal lightweight student model that balances accuracy and latency from the candidate student models.

[0013] Step 4: Perform knowledge distillation on the teacher model and the Pareto optimal lightweight student model.

[0014] In the above scheme, by introducing the Adaptive Scale-Aware Attention Pyramid (ASAP) module to reconstruct the feature response, improving the Heterogeneous Hollow Residual Module (HARM), using a multi-objective genetic algorithm to search for the optimal pruning strategy, and combining knowledge distillation technology, the problems of large differences in aircraft target scale, strong background interference, and difficulty in deploying high-precision models in real time at the edge are solved in airport surface surveillance.

[0015] Furthermore, in step 1, the adaptive scale-aware attention pyramid module includes a feature refinement unit based on a coupling and penalty mechanism;

[0016] The feature refinement unit based on coupling and penalty mechanisms includes a channel attention unit, a spatial attention unit, a gradient entropy calculator, and an exponential decay unit; for low-level detailed features F backbone The processing is as follows:

[0017] ;

[0018] Among them, F refined A represents the refined features output by the feature refinement unit based on coupling and penalty mechanisms; c Channel characteristics; A s Spatial features; is the sigmoid activation function; e is the base of the natural logarithm; H(x,y) is the background gradient entropy penalty coefficient; H(x,y) is the spatial gradient entropy of the local region (x,y) in the image. This is a gradient entropy penalty term; For element-wise multiplication; This is an element-wise addition.

[0019] Furthermore, in step 1, the adaptive scale-aware attention pyramid module also includes a multi-branch structure;

[0020] Refining Feature F refined The input multi-branch structure first passes through parallel 1×1 convolutional layers, 3×3 convolutional layers with sampling rate d=6, 3×3 convolutional layers with sampling rate d=12, 3×3 convolutional layers with sampling rate d=18, and an image pooling layer. These parallel convolutional layers and image pooling layers are then concatenated along the channel dimension to obtain the multi-scale feature F output by the adaptive scale-aware attention pyramid module. scale ;

[0021] Finally, in the encoder stage, the multi-scale features are reduced in dimensionality by a 1×1 convolutional layer to obtain the high-level semantic features F output by the encoder. final .

[0022] Furthermore, in step 2, the step of performing multi-objective evolutionary pruning on the heterogeneous hollow residual modules in the teacher model based on the dynamic topological mask search algorithm oriented towards codec association, constrained by an asymmetric topological penalty mechanism, includes:

[0023] The heterogeneous hollow residual modules in the teacher model are mapped to N residual blocks, with the set of residual blocks being R={Block1,Block2,...,BlockN}, where Blocki represents the i-th residual block, and i is the network depth index, i=1,2,...,N; using the binary mask vector B i This indicates the retention state of the corresponding residual block Blocki, B i =1 is reserved, B i =0 means not to be retained;

[0024] Calculate the fitness evaluation function:

[0025] ;

[0026] in, This is the accuracy balance coefficient; This is the delay balance coefficient; is the topology penalty balance coefficient; IoU is the precision function; Latency is the delay function; It is a volume function; is the asymmetric penalty weight for the i-th residual block.

[0027] Furthermore, step 3 specifically includes the following steps:

[0028] Step 3-1, Population initialization;

[0029] Step 3-2, fitness assessment;

[0030] Step 3-3, Evolutionary Iteration;

[0031] Steps 3-4 terminate and output the Pareto optimal solution set.

[0032] Furthermore, step 3-1 specifically includes:

[0033] Randomly generate the initial population P g ;

[0034] Calculate the importance score of the b-th residual block (Blockb):

[0035] ;

[0036] Among them, I b Assign an importance score to the b-th residual block, where b = 1, 2, ..., N, and N is the total number of residual blocks; p is any parameter. , Let |.| be the set of parameters for the residual block Blockb; |.| is the L1 norm.

[0037] Furthermore, step 3-2 specifically includes:

[0038] Calculate the fitness evaluation function for each residual block, construct candidate student models, and transplant the corresponding asymmetric penalty weights from the trained teacher network. To locate the Pareto front solution.

[0039] Furthermore, step 3-3 specifically includes:

[0040] After calculating the fitness evaluation function (Fitness) for the entire population, the multi-objective genetic algorithm enters an iterative evolutionary loop until a preset termination condition is met. If the termination condition is not met, a new generation population P is generated. g+1 ;

[0041] The iterative evolutionary cycle includes:

[0042] Selection: The best individual is retained, and the remaining individuals compete in a tournament to determine the parent individual;

[0043] Crossover: Pairs of parent individuals perform single-point crossover with a preset probability;

[0044] Mutation: Heuristic mutation is introduced. If the binary mask vector of a residual block with an importance score higher than the first threshold is mutated to 1, the residual block with an importance score lower than the second threshold is mutated to 0 if the binary mask vector is 1.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] (1) This invention introduces an adaptive scale-aware attention pyramid module (ASAP) based on the traditional DeepLabV3+ network, and improves the traditional ResNet-101 backbone network in the encoder to form a heterogeneous hollow residual module (HARM). The teacher model obtained after the improvement has improved performance.

[0047] (2) This invention proposes an adaptive network structure search strategy based on block importance pruning, which defines the heterogeneous hollow residual modules in the teacher network as "blocks" that can be pruned. With the precision function, delay function and volume function as joint optimization objectives, a multi-objective genetic algorithm is used to automatically search for the Pareto optimal lightweight student model.

[0048] (3) The present invention designs a composite total loss function for intrusion and constructs a composite total loss function with dual constraints of “region-target”. The runway intrusion risk is quantified as a penalty term and embedded in the training process. The penalty term is calculated based on Dice Loss and BCELoss. The activation condition is strictly limited to when false alarms and false alarms occur, thereby improving the model’s sensitivity to intrusion events. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a schematic diagram of the improved DeepLabV3+ network structure of this invention.

[0051] Figure 2 This is a schematic diagram of the heterogeneous void residual module structure of the present invention.

[0052] Figure 3 This is a schematic diagram of the feature refinement unit structure based on the coupling and penalty mechanism of the present invention.

[0053] Figure 4 This is a schematic diagram of the dynamic topology mask search algorithm of the present invention.

[0054] Figure 5 This is a flowchart of the multi-objective genetic algorithm of the present invention.

[0055] Figure 6 This is a flowchart of the method of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0057] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0058] Example 1:

[0059] This invention is achieved through the following technical solutions, such as... Figure 6 As shown, a lightweight aircraft target recognition method based on an improved DeepLabV3+ includes the following steps:

[0060] Step 1: Construct a teacher model based on the improved DeepLabV3+ network. The teacher model includes an encoder and a decoder. The encoder includes a heterogeneous hollow residual module (HARM) and an adaptive scale-aware attention pyramid module (ASAP).

[0061] While the traditional DeepLabV3+ network, used as a teacher model, utilizes Spatial Pyramid Pooling with Hollows (ASPP) to extract multi-scale features, it suffers from a lack of differentiation between foreground (aircraft) and background (runway) features when processing airport panoramic images, as ASPP indiscriminately processes all channels of the input features. Therefore, this proposal improves the teacher model based on the DeepLabV3+ network by introducing an Adaptive Scale-Aware Attention Pyramid (ASAP) module into the encoder of the traditional DeepLabV3+ network and modifying the traditional ResNet-101 backbone network in the encoder to form a Heterogeneous Hollow Residual Module (HARM).

[0062] For details, please see Figure 1 The teacher model includes an encoder and a decoder. The encoder aims to repair high-frequency geometric features while suppressing background noise, extracting multi-scale semantic information from the input image. The encoder includes a heterogeneous dilated residual module and an adaptive scale-aware attention pyramid module. The decoder is responsible for fusing the high-level semantic features extracted by the encoder with the low-level detail features output from the heterogeneous dilated residual module to generate a high-resolution, well-defined segmentation map. The decoder includes a 1×1 convolutional layer, a 3×3 convolutional layer, and an upsampling layer.

[0063] Please see Figure 2 The heterogeneous dilated residual module includes a main path, a secondary path, and a positive AC-shaped fusion unit. The main path is a dilated convolutional layer with a dilation rate r>1, the secondary path is a standard convolutional layer with a dilation rate r=1, and the positive AC-shaped fusion unit includes processing the input features by performing Laplacian gradient calculation, tanh activation function, and feature fusion.

[0064] The acquired raw image X inThe input is a heterogeneous dilated residual module, and the main path uses a dilated convolutional layer with a dilation rate r>1 to capture the original image X. in From the wide-area environmental context, the main feature U is obtained; the original image X is captured from the path using a standard convolutional layer with dilation rate r=1. in Fine-grained geometric pixels are used to obtain feature V. To avoid representational confusion caused by simply adding features from two paths, a positive AC fusion unit is designed to calculate the spatial Laplacian gradient of the principal feature U. Used to control fusion, when the main path experiences a characteristic break at a slender edge such as a wing (i.e., During periods of drastic fluctuations, the tanh activation function opens the geometric valve, injecting from feature V to precisely fill the gaps, thereby eliminating low-level topological breaks. Specifically, this involves processing the principal feature U, the secondary feature V, and the tanh activation function. Element-wise multiplication yields the first fusion feature B. fuse Then the first fusion feature B fuse Then, the second fused feature C is obtained by adding it element-wise to the main feature U. fuse Finally, the second fusion feature C fuse Compared with the original image X in Element-wise summation yields the low-level detail features F, which contain rich and continuous context, output by the heterogeneous void residual module. backbone .

[0065] To address the "mesh effect" caused by the expanded receptive field of deep networks, this solution upgrades the traditional ResNet-101 backbone network to a master-slave heterogeneous dilated residual module designed for continuous geometric feature fidelity. The improved heterogeneous dilated residual module decouples the conflicting processes of "extracting semantics from a large receptive field" and "preserving local pixel continuity" within a single residual structure, establishing a self-healing mechanism within the deep network. This effectively prevents pixel-level fragmentation and feature loss in high-frequency, slender structures such as aircraft wings and tails after skip sampling through multiple dilated convolutions without reducing the network's global receptive field, thus providing low-level detailed feature inputs with complete physical structure for subsequent decoding and reconstruction. The heterogeneous dilated residual module does not rely on a single convolution method but combines two convolutions with different fields of view and designs a gradient-based fusion mechanism (i.e., a positive AC fusion unit) to adaptively enhance key details (such as edges) in image features, thereby improving the accuracy of subsequent tasks (such as object detection and image segmentation).

[0066] Please see Figure 1 The adaptive scale-aware attention pyramid module includes feature refinement units based on coupling and penalty mechanisms and a multi-branch structure. For example... Figure 3As shown, the feature refinement unit based on the coupling and penalty mechanism includes a channel attention unit, a spatial attention unit, a gradient entropy calculator, and an exponential decay unit. Figure 1 As shown, the multi-branch structure includes parallel 1×1 convolutional layers, 3×3 convolutional layers with sampling rate d=6, 3×3 convolutional layers with sampling rate d=12, 3×3 convolutional layers with sampling rate d=18, and image pooling layers.

[0067] like Figure 3 As shown, the low-level detail features F backbone The input is a feature refinement unit based on a coupling and penalty mechanism. Channel features are obtained through a channel attention unit, and spatial features are obtained through a spatial attention unit. The channel features and spatial features are then added element-wise, and a dynamic penalty mask is obtained using a sigmoid activation function. A gradient entropy calculator generates spatial gradient entropy for local image regions. An exponential decayer generates a gradient entropy penalty term based on the spatial gradient entropy. The dynamic penalty mask and the gradient entropy penalty term are then multiplied element-wise, and finally multiplied by the low-level detail features F. backbone Element-wise multiplication yields the refined feature F output by the feature refinement unit based on coupling and penalty mechanisms. refined .

[0068] Feature refinement units based on coupling and penalty mechanisms refine low-level detail features F backbone The processing procedure can be described as follows:

[0069] ;

[0070] Among them, F refined F represents the refined features output by the feature refinement unit based on the coupling and penalty mechanism. backbone A represents the low-level detail features output by the heterogeneous void residual module; c Channel characteristics; A s Spatial features; is the sigmoid activation function; e is the base of the natural logarithm; The background gradient entropy penalty coefficient is used. H(x,y) is the spatial gradient entropy of the local region (x,y) of the image. For element-wise multiplication; This is an element-wise addition.

[0071] The feature refinement unit based on the coupling and penalty mechanism not only achieves spatial and channel weighting, but also introduces a gradient entropy penalty term. Adaptively, before features enter the multi-branch structure, the background weights of the smooth runway with low gradients exhibit an exponential, cliff-like decay.

[0072] like Figure 1 As shown, the refining feature F will then be applied.refined The input multi-branch structure first passes through parallel 1×1 convolutional layers, 3×3 convolutional layers with sampling rate d=6, 3×3 convolutional layers with sampling rate d=12, 3×3 convolutional layers with sampling rate d=18, and an image pooling layer. These parallel convolutional layers and image pooling layers are then concatenated along the channel dimension to obtain the multi-scale feature F output by the adaptive scale-aware attention pyramid module. scale Finally, in the encoder stage, the multi-scale features F scale After dimensionality reduction using a 1×1 convolutional layer, the high-level semantic features F of the encoder's final output are obtained. final The parallel convolutional layers extract multi-scale features based on the high-response structure of the aircraft, and then concatenate and reduce the dimensionality in the channel dimension to generate high-level semantic features of the encoder path.

[0073] Obtain low-level detail features F backbone Subsequently, attention-guided feature refinement is the core improvement in the encoder of this scheme. Unlike the traditional DeepLabV3+ network that directly inputs features into the ASPP, the adaptive scale-aware attention pyramid module proposed in this scheme is used to prevent interference from large areas of irrelevant background (such as large smooth pavement areas, grass, and marking textures) in airport panoramic monitoring on the multi-scale feature pooling process. Its effect is that by constructing a dynamic gating barrier with gradient awareness, the network can first perform high signal-to-noise ratio purification on spatial and channel responses, avoiding the amplification of global noise as a target feature as in traditional modules. This significantly improves the interference purity and focusing accuracy of the multi-branch structure when extracting multi-scale features of aircraft.

[0074] Please continue reading Figure 1 The decoder first processes the high-level semantic features F final We perform a 4x upsampling to obtain feature F1; simultaneously, we use a 1×1 convolutional layer to refine the low-level detail features F1. backbone After processing, the feature is fused with feature F1 to obtain feature F2; then, a 3×3 convolutional layer is used to refine the boundaries of feature F2 to obtain feature F3; finally, feature F3 is upsampled by 4 times to restore the original image X. in The resolution is determined, and the final pixel-level segmentation mask for the aircraft is output.

[0075] By using the improved DeepLabV3+ network, the teacher model architecture integrates the attention mechanism into the semantic segmentation process, ensuring that the teacher model's computational resources are prioritized for the most important feature regions before multi-scale analysis, thereby significantly improving the segmentation accuracy and robustness of the teacher model.

[0076] Step 2: Based on the dynamic topological mask search algorithm for encoding and decoding association, perform multi-objective evolutionary pruning on the heterogeneous void residual modules in the teacher model, which is constrained by the asymmetric topological penalty mechanism, so as to obtain the candidate student model that reaches Pareto optimality.

[0077] When constructing the teacher model, an attention mechanism is introduced to improve the model's robustness. In order to alleviate the problem of redundant teacher model parameters, reduce computational overhead, and adapt to edge devices with limited computing power, this solution proposes a two-stage lightweight strategy of "evolutionary pruning-knowledge distillation" to construct a lightweight student model.

[0078] like Figure 4 As shown, the heterogeneous hollow residual modules in the teacher model are mapped to N residual blocks (N=33), and the set of residual blocks is R={Block1,Block2,...,BlockN}, where Blocki represents the i-th residual block, and i is the network depth index, i=1,2,...,N. The binary mask vector B... i This indicates the retention state of the corresponding residual block Blocki (B i =1 is reserved, B i =0 indicates no retention). Because the decoder relies heavily on the unique architectural properties of the high-frequency geometric features extracted from the shallow layers of the heterogeneous hole residual module to recover the aircraft profile, this step introduces an asymmetric topology protection mechanism, establishes a fitness evaluation function, and designs an asymmetric penalty weight that varies with network depth i. (That is, to enforce protection of shallow modules and relax pruning restrictions on deep modules).

[0079] The fitness evaluation function is:

[0080] ;

[0081] in, This is the accuracy balance coefficient; This is the delay balance coefficient; is the topology penalty balance coefficient; IoU is the precision function; Latency is the delay function; It is a volume function; B is the asymmetric penalty weight for the i-th residual block; i This represents the preserved state of the i-th residual block.

[0082] The core of the fitness evaluation function lies in constructing an asymmetric penalty weight that varies with network depth i. (The parameter is set to 0.1 here based on the depth N of the residual block and the nonlinear damping constant at the strict mapping of the characteristic abstract curvature).

[0083] Step 3: Use a multi-objective genetic algorithm to automatically search for the Pareto optimal lightweight student model that balances accuracy and latency from the candidate student models.

[0084] The decoder relies on high-frequency continuous geometric features extracted by the shallow network (smaller i, smaller receptive field) of the heterogeneous hollow residual module to recover the target boundary, and on abstract semantic features extracted by the deep network (larger i, larger receptive field) of the heterogeneous hollow residual module for category determination. In airport monitoring scenarios, the slender edges of aircraft, such as wings, are extremely sensitive to shallow high-frequency geometric features, while the background (runway) is relatively uniform, resulting in significant redundancy in deep abstract semantic features. When the multi-objective genetic algorithm performs pruning operations on the shallow network, it incurs a large penalty value, forcibly protecting the shallow geometric features; while when performing pruning operations on the deep network, the penalty value decreases exponentially, allowing the algorithm to prune redundant semantic branches over a large area.

[0085] The fitness evaluation function constructed in step 2 simultaneously optimizes three conflict objectives: accuracy (IoU), latency (Latency), and mass (Volume). Mathematically, it's impossible for three conflicting objectives to simultaneously achieve absolute optimality (for example, the model with the highest accuracy often cannot have the lowest latency). Therefore, step 3, using a multi-objective genetic algorithm, outputs a Pareto-optimal solution set upon convergence of the last generation. Each optimal lightweight student model in this set achieves an optimal balance between accuracy and efficiency (i.e., latency cannot be further reduced without sacrificing accuracy, and vice versa). The multi-objective genetic algorithm serves as the solution domain, by applying the binary mask vector B... i Multiple candidate student model retention schemes are iteratively cross-crossed and mutated, with the fitness evaluation function serving as the evaluation constraint for each generation. Finally, in 2... 33 The Pareto optimal solution set that balances accuracy and low latency is automatically searched from possible topological combinations.

[0086] Next, in actual airport deployment, the most suitable Pareto-optimal lightweight student model can be selected from the Pareto-optimal solution set based on specific hardware computing power limitations. If the hardware facility has extremely weak computing power (such as an edge camera), the student model with the lowest latency in the set is selected; if the hardware facility has moderate computing power (such as a control tower server), the student model with the highest accuracy in the set is selected. This fully demonstrates the greatest advantage of the multi-objective evolutionary search method of this scheme compared to the traditional single pruning.

[0087] Specifically, to solve the multi-objective optimization problem of the fitness evaluation function, step 3 employs a multi-objective genetic algorithm for automated network search. The multi-objective genetic algorithm is a stochastic search and optimization algorithm that simulates the process of biological evolution. Through iterative selection, crossover, and mutation operations on a population composed of candidate solutions, it eventually converges to a Pareto optimal solution set. Figure 5 As shown, the specific steps include:

[0088] Step 3-1, Population initialization.

[0089] A multi-objective genetic algorithm begins with a randomly generated initial population P. g Each individual is a binary chromosome representing a candidate student model. To improve search efficiency, a population initialization strategy based on prior knowledge is introduced, namely block importance evaluation. Let the parameter set of the b-th residual block, Blockb, be... The L1 norm is used to quantify the importance score of Blocks:

[0090] ;

[0091] Among them, I b Assign an importance score to the b-th residual block, where b = 1, 2, ..., N, and N is the total number of residual blocks; p is any parameter. |.| represents the L1 norm.

[0092] Initialize population P g At that time, according to I b A probability distribution is set so that residual blocks with high importance scores are more likely to be initialized as binary mask vector B. b =1 (i.e., reserved).

[0093] Step 3-2, fitness assessment.

[0094] Calculate the fitness evaluation function for each residual block, construct candidate student models, and transplant the corresponding asymmetric penalty weights from the trained teacher network. This avoids time-consuming training. It quickly calculates IoU, Latency, and [other parameters] on the validation set. To locate the Pareto front solution.

[0095] Step 3-3: Evolution and Iteration.

[0096] After calculating the fitness evaluation function (Fitness) for the entire population, the multi-objective genetic algorithm enters an iterative evolutionary loop until a preset termination condition is met. If the termination condition is not met, a new generation population P is generated through a series of evolutionary operators. g+1 .

[0097] Selection: The best individual is retained, and the remaining individuals compete in a tournament to determine the parent individual;

[0098] Crossover: Pairs of parent individuals perform a single-point crossover with a certain probability;

[0099] Mutation: Heuristic mutation is introduced, that is, if the binary mask vector of a residual block with a high importance score is 0, it will be mutated to 1 (restored); if the binary mask vector of a residual block with a low importance score is 1, it will be mutated to 0 (pruned), in order to protect the key feature paths.

[0100] Steps 3-4: Termination and Output.

[0101] After reaching the predicted iteration number, output the Pareto optimal solution set.

[0102] This invention deploys a student model in airports with limited computing power. The core advantages of deploying a student model are twofold: First, through structured pruning using a multi-objective genetic algorithm, the student model accurately removes redundant feature branches in the deep layers of the network, significantly reducing model complexity and computational overhead, making it suitable for edge devices in airports with limited computing power. Second, through knowledge distillation technology, the student model inherits the "soft label" prior knowledge containing rich inter-class relationships from the teacher network, compensating for the performance loss due to the shallower network structure. Ultimately, the lightweight student model maintains aircraft target recognition accuracy close to that of the teacher model while ensuring extremely low inference time, achieving a synergistic optimization of high accuracy and high real-time performance.

[0103] Step 4: Perform knowledge distillation on the teacher model and the Pareto optimal lightweight student model.

[0104] Although the Pareto-optimal lightweight student model obtained through multi-objective genetic algorithm search has the best structure, its performance is severely degraded due to the removal of a large number of parameters. To restore its performance, this step employs knowledge distillation based on the teacher-student paradigm.

[0105] The total loss function L of the Pareto optimal lightweight student model is... total It consists of two weighted parts: one part is the distillation loss function L calculated between the predicted mask and the true mask of the Pareto optimal lightweight student model. Soft The other part is the supervised loss function L calculated between the prediction mask of the Pareto optimal lightweight student model and the soft label output by the teacher network. Hard .

[0106] Distillation loss function L Soft This allows the Pareto-optimal lightweight student model to inherit prior knowledge, improving its ability to understand the characteristics of complex scenes. Distillation loss function L Soft The calculation formula is:

[0107] ;

[0108] Where c is the category index; p c q represents the soft label probability value for the c-th category output by the teacher model; c This represents the probability value predicted for the c-th class by the Pareto-optimal lightweight student model.

[0109] After introducing the temperature coefficient T, p i and q i It can be represented as:

[0110] ;

[0111] ;

[0112] Among them, v c This represents the teacher model's original predicted logistic value for the c-th class before the classification layer (Softmax activation layer); v j z represents the teacher model's initial predictive logistic value for the j-th category before the classification layer; c z represents the original predicted logistic values ​​for the c-th class of the Pareto-optimal lightweight Student model before the classification layer (Softmax activation layer); j This is the original prediction logistic value for the j-th class of the Pareto-optimal lightweight student model before the classification layer; the temperature coefficient T is empirically set to 10.

[0113] Airport scenes have a small proportion of aircraft, resulting in a serious foreground-background category imbalance. This solution combines a supervised loss function that integrates BCE Loss and Dice Loss. The former constrains false positives, while the latter suppresses false negatives, thus ensuring both pixel-level prediction accuracy and the integrity of segmented region contours.

[0114] BCE Loss is a pixel-level binary classification task loss function that transforms aircraft identification into pixel-by-pixel binary classification, thereby suppressing the risk of false alarms from background pixels. The calculation formula is as follows:

[0115] ;

[0116] Among them, L BCE The loss function is binary classification; K is the total number of pixels, k=1,2,...,K; y k The true label (0 or 1) for pixel k; Let k be the probability that pixel k is predicted to be a foreground element.

[0117] Dice Loss originates from the Dice similarity coefficient. Given that aircraft target pixels account for a small proportion in top-view images and are easily missed, Dice Loss is introduced to evaluate the overlap between the predicted and the actual aircraft mask. This loss compensates for the difficulty in identifying small targets and enhances the model's ability to extract the complete outline of the aircraft. The calculation formula is as follows:

[0118] ;

[0119] Among them, L Dice The Dice loss function; For smoothing coefficients, .

[0120] Based on this, L BCE and L Dice Weighted fusion to construct the supervised loss function L Hard The calculation formula is:

[0121]

[0122] in, For L BCE The weight, ; For L Dice The weight, .

[0123] After introducing a supervised loss function, a distillation loss function is introduced into the total loss function for retraining to restore the performance of the Pareto-optimal lightweight student model. The total loss function L... total The calculation formula is:

[0124] ;

[0125] in, As the total loss weight, .

[0126] The Pareto-optimal lightweight student model of this invention outperforms the teacher model in terms of clear contour segmentation and accurate details in interference-free and motion-blurred scenes. In particular, under Gaussian noise interference, the teacher model suffers from large-area missed detections due to excessive focus on high-frequency texture details caused by parameter redundancy, while the Pareto-optimal lightweight student model eliminates redundant channels through evolutionary pruning and focuses more on extracting the overall structural features of the aircraft, thus showing stronger generalization ability under harsh conditions.

[0127] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A lightweight aircraft target recognition method based on improved DeepLabV3+, characterized in that, Includes the following steps: Step 1: Construct a teacher model based on the improved DeepLabV3+ network. The teacher model includes an encoder and a decoder. The encoder includes a heterogeneous hollow residual module and an adaptive scale-aware attention pyramid module. The collected original image X in The input isomorphic hole residual module, the main path captures the main features U of the original image X in The slave path captures the slave features V of the original image X in ​ The spatial Laplace gradient is obtained by calculating the Laplace gradient of the principal feature U using the positive AC-shaped fusion unit. After processing the main feature U, secondary features V, and the tanh activation function... Element-wise multiplication yields the first fusion feature B. fuse First fusion feature B fuse The second fused feature C is obtained by adding it element-wise to the main feature U. fuse Second fusion feature C fuse Compared with the original image X in Element-wise addition is performed to obtain the low-level detail features F output by the heterogeneous void residual module. backbone Low-level detail features F backbone The high-level semantic features F are obtained by inputting the adaptive scale-aware attention pyramid module. final ; Step 2: Based on the dynamic topological mask search algorithm for encoding and decoding association, perform multi-objective evolutionary pruning on the heterogeneous hollow residual modules in the teacher model, which is constrained by the asymmetric topological penalty mechanism, so as to obtain the candidate student model that reaches Pareto optimality. Step 3: Use a multi-objective genetic algorithm to automatically search for the Pareto optimal lightweight student model that balances accuracy and latency from the candidate student models. Step 4: Perform knowledge distillation on the teacher model and the Pareto optimal lightweight student model.

2. The lightweight aircraft target recognition method based on the improved DeepLabV3+ according to claim 1, characterized in that, In step 1, the adaptive scale-aware attention pyramid module includes a feature refinement unit based on a coupling and penalty mechanism; The feature refinement unit based on the coupling and punishment mechanism comprises a channel attention unit, a spatial attention unit, a gradient entropy calculator and an exponential attenuator. backbone The processing of the low-level detail feature F ； Among them, F refined A represents the refined features output by the feature refinement unit based on coupling and penalty mechanisms; c Channel characteristics; A s Spatial features; is the sigmoid activation function; e is the base of the natural logarithm; H(x,y) is the background gradient entropy penalty coefficient; H(x,y) is the spatial gradient entropy of the local region (x,y) in the image. This is a gradient entropy penalty term; For element-wise multiplication; This is an element-wise addition.

3. The lightweight aircraft target recognition method based on the improved DeepLabV3+ according to claim 2, characterized in that, In step 1, the adaptive scale-aware attention pyramid module also includes a multi-branch structure; Refining features F refined The input multi-branch structure first passes through a parallel 1x1 convolutional layer, a 3x3 convolutional layer with a sampling rate d=6, a 3x3 convolutional layer with a sampling rate d=12, a 3x3 convolutional layer with a sampling rate d=18, an image pooling layer, and a plurality of parallel convolutional layers and image pooling layers are spliced in the channel dimension to obtain a multi-scale feature F output by the adaptive scale perception attention pyramid module scale ; Finally, in the encoder stage, the multi-scale features are reduced in dimensionality by a 1×1 convolutional layer to obtain the high-level semantic features F output by the encoder. final .

4. The lightweight aircraft target recognition method based on the improved DeepLabV3+ according to claim 1, characterized in that, Step 2, based on the dynamic topological mask search algorithm for encoding / decoding association, involves performing multi-objective evolutionary pruning on the heterogeneous hollow residual modules in the teacher model, constrained by an asymmetric topological penalty mechanism. This includes: The heterogeneous hollow residual modules in the teacher model are mapped to N residual blocks, with the set of residual blocks being R={Block1,Block2,...,BlockN}, where Blocki represents the i-th residual block, and i is the network depth index, i=1,2,...,N; using the binary mask vector B i This indicates the retention state of the corresponding residual block Blocki, B i =1 is reserved, B i =0 means not to be retained; Calculate the fitness evaluation function: ； in, This is the accuracy balance coefficient; This is the delay balance coefficient; is the topology penalty balance coefficient; IoU is the precision function; Latency is the delay function; It is a volume function; is the asymmetric penalty weight for the i-th residual block.

5. The lightweight aircraft target recognition method based on the improved DeepLabV3+ according to claim 4, characterized in that, Step 3 specifically includes the following steps: Step 3-1, Population initialization; Step 3-2, fitness assessment; Step 3-3, Evolutionary Iteration; Steps 3-4 terminate and output the Pareto optimal solution set.

6. The lightweight aircraft target recognition method based on the improved DeepLabV3+ according to claim 5, characterized in that, Step 3-1 specifically involves: Randomly generate the initial population P g ; Calculate the importance score of the b-th residual block (Blockb): ； Among them, I b Assign an importance score to the b-th residual block, where b = 1, 2, ..., N, and N is the total number of residual blocks; p is any parameter. , Let |.| be the set of parameters for the residual block Blockb; |.| is the L1 norm.

7. The lightweight aircraft target recognition method based on the improved DeepLabV3+ according to claim 6, characterized in that, Step 3-2 specifically involves: Calculate the fitness evaluation function for each residual block, construct candidate student models, and transplant the corresponding asymmetric penalty weights from the trained teacher network. To locate the Pareto front solution.

8. The lightweight aircraft target recognition method based on the improved DeepLabV3+ according to claim 7, characterized in that, Step 3-3 specifically involves: After calculating the fitness evaluation function (Fitness) for the entire population, the multi-objective genetic algorithm enters an iterative evolutionary loop until a preset termination condition is met. If the termination condition is not met, a new generation population P is generated. g+1 ; The iterative evolutionary cycle includes: Selection: The best individual is retained, and the remaining individuals compete in a tournament to determine the parent individual; Crossover: Pairs of parent individuals perform single-point crossover with a preset probability; Mutation: Heuristic mutation is introduced. If the binary mask vector of a residual block with an importance score higher than the first threshold is mutated to 1, the residual block with an importance score lower than the second threshold is mutated to 0 if the binary mask vector is 1.

Images