Visual perception model construction and efficient evolution method for unmanned aerial vehicle aerial image

Abstract

The invention provides a visual perception model construction and efficient evolution method for an unmanned aerial vehicle aerial image, and the method comprises the steps: obtaining a walkable region segmentation data set, and completing the data preprocessing and data division; the method comprises the following steps: constructing a lightweight fine tuning network consisting of a ViT-Adapter and a VPT; an auxiliary task guiding framework is constructed, a plurality of lightweight auxiliary decoders are added in an original encoder-decoder framework, and a vehicle segmentation auxiliary task is executed; iteratively updating and optimizing network parameters by adopting a back propagation algorithm until model loss tends to converge; and inputting a to-be-segmented unmanned aerial vehicle aerial image into the trained model, wherein the model output is a segmentation result. The method makes up the defects of an existing efficient fine tuning method, assists a general image segmentation model to efficiently evolve the sight line in a drivable area segmentation task facing an unmanned aerial vehicle aerial image, and improves the performance of a final model in an image segmentation perception task.

Description

Technical Field

[0001] The present invention belongs to the technical field of basic perception large model construction and evolution, and in particular relates to a visual perception model construction and efficient evolution method for unmanned aerial vehicle aerial images. Background Art

[0002] The task of drivable area segmentation is one of the important tasks in drone aerial photography scenarios. It has important applications in intelligent transportation, military reconnaissance and other fields, and is also a current academic research hotspot. Drivable area segmentation belongs to the category of semantic segmentation. In the field of semantic segmentation, "pre-training + full parameter fine-tuning" has become the mainstream learning paradigm. However, the parameter scale of the currently popular basic perception models such as SAM (Segment Anything Model) is usually very large, and the computational complexity is very high when the full amount is fine-tuned, and the applicability is poor. To this end, the researchers proposed a parameter-efficient fine-tuning method (Parameter-EfficientFine-Tuning, PEFT) to help the model evolve quickly and efficiently.

[0003] The core idea of ​​the parameter efficient fine-tuning method is to fix most of the pre-trained parameters and only fine-tune a small number of the original parameters of the model, or fine-tune a small number of additional parameters, so as to minimize the computation and storage requirements. The parameter efficient fine-tuning methods currently proposed can be roughly divided into three categories: selective methods, incremental methods, and re-parameter methods. In addition, according to the location where the parameters are added, the incremental method can be further divided into two subcategories, namely the adapter method and the soft prompt method.

[0004] In simple terms, the selective method is to "select some of the original parameters of the pre-trained model for update". Typical methods include BitFit, DiffPruning, and SPT. Among them, BitFit is the most popular. This method only selects the bias parameters in the model for fine-tuning. The method is simple and effective.

[0005] The adapter method fine-tunes the frozen pre-trained model by adding a new trainable network layer. Its representative Adapter method adds two adapter network modules to each Transformer layer to fine-tune the pre-trained model. This method has been widely used in the fields of natural language processing and computer vision, and a series of variants have also emerged, such as AdaMix, Scaled Parallel Adapter, AdapterFormer, etc.

[0006] The soft prompt method is to add trainable tokens to the input of the model. These tokens can indirectly adjust the attention mechanism and help the model focus on the most relevant features for downstream tasks, thereby achieving fine-tuning. Typical soft prompt methods in natural language processing include Prompt Tuning, Prefix Tuning, P-Tuning, P-Tuning v2, etc. Jia et al. transferred the idea of ​​soft prompts to the field of computer vision and proposed Visual Prompt Tuning, referred to as VPT.

[0007] The core idea of ​​the reparameterization method is to first add additional parameters, which can be merged into the model in some way after fine-tuning, which is equivalent to directly optimizing the model. The most famous reparameterization method is the LoRA method proposed by Hu et al., which is based on the principle of expressing the increment of the weight matrix as the multiplication of two low-rank matrices. There have been many subsequent works to improve LoRA, including KronA, AdaLoRA, etc. The SSF proposed in recent years also belongs to the category of reparameterization methods. This method uses trainable scaling parameters and offset parameters to adjust the weights, and these parameters can be merged into the original parameters after training.

[0008] Existing efficient fine-tuning methods pursue universality in design, aiming to be able to fine-tune models for multiple tasks such as image classification, detection, segmentation, and generation. Few are specifically designed and optimized for a specific fine-tuning task. Existing efficient fine-tuning methods face the following two main problems when aligning general pre-trained models with semantic segmentation tasks of drone aerial photography scenes.

[0009] On the one hand, existing studies have shown that multi-scale feature maps have a positive impact on semantic segmentation tasks. However, general pre-training models based on the Transformer architecture, such as SAM, are mostly "straight-tube" structures that can only generate single-scale feature maps, and the image feature information they contain is relatively scarce. Although multi-scale feature maps can be indirectly obtained through upsampling, they are still based on the feature transformation of a single scale, and the image information obtained is still not rich enough. In order to ensure universality, the currently popular parameter efficient fine-tuning method will only perform fine-tuning within the model by "unfreezing some parameters" or "adding some parameters", which makes it difficult to change the output scale of the model, limiting the alignment effect.

[0010] On the other hand, in the drivable area segmentation task for drone aerial images, traffic-related scene information such as road vehicles can usually enhance the accuracy of drivable area segmentation. However, the general pre-trained model is trained on general task datasets without mining traffic scene-related information. In addition, the existing parameter efficient fine-tuning methods do not consider how to help the model understand traffic scenes during the fine-tuning process, resulting in poor segmentation performance on related tasks.

[0011] Therefore, making full use of multi-scale feature maps and introducing relevant information of traffic scenes are effective means to improve the evolution effect of general image segmentation models such as SAM in drivable area segmentation tasks. Summary of the invention

[0012] The present invention aims to solve at least one of the technical problems existing in the prior art.

[0013] The present invention provides a method for constructing and efficiently evolving a visual perception model for unmanned aerial vehicle aerial images, characterized in that the method for constructing and efficiently evolving a visual perception model for unmanned aerial vehicle aerial images comprises:

[0014] S1: Obtain the segmentation dataset of the drivable area from the perspective of drone aerial photography, complete data preprocessing, and perform data division;

[0015] S2: Construct a lightweight fine-tuning network: The lightweight fine-tuning network consists of two parts: the Adapter network and the VPT network. The Adapter network includes the SPM module, the Injector module, and the Extractor module. The SPM module is used to generate multi-scale feature maps, and the Injector module and the Extractor module are used to interact the generated multi-scale feature maps with the features in the SAM image encoder. The VPT network is used to add trainable tokens at the input of each layer of the Transformer model to indirectly adjust the attention mechanism.

[0016] S3: Construct an auxiliary task guidance framework: Add multiple lightweight auxiliary decoders to the original "encoder-decoder" architecture to perform the "vehicle segmentation" auxiliary task; during training, multiple decoders share one encoder and perform decoding separately; finally, the losses of these decoders are weighted summed to obtain the final loss;

[0017] S4: Train the segmentation network: Use the loss function set in S3 and the back propagation algorithm to iteratively update and optimize the network parameters until the model loss converges;

[0018] S5: Segment the drivable area: preprocess the drone aerial image to be segmented according to the preprocessing method in S1, and then input the preprocessed image into the trained segmentation network. The output result of the segmentation network model is the drivable area segmentation result.

[0019] The technical solution of the present invention provides a method for constructing and efficiently evolving a visual perception model for drone aerial images. The method preprocesses and divides the data, constructs a lightweight fine-tuning network, constructs an auxiliary task guidance framework, trains a segmentation network, and finally performs drivable area segmentation. By fully mining multi-scale feature maps and using traffic scene related information such as vehicle targets, the general image segmentation model can be efficiently evolved to the task of drivable area segmentation of drone aerial images, thereby improving the performance of the model on this task. Compared with the prior art, the technical solution of the present invention can solve the technical problems in the prior art that the straight-tube general image segmentation model is difficult to directly obtain multi-scale feature maps, and the existing fine-tuning method is difficult to effectively introduce traffic scene information. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The included drawings are used to provide a further understanding of the embodiments of the present invention, which constitute a part of the specification, are used to illustrate the embodiments of the present invention, and together with the text description, explain the principles of the present invention. Obviously, the drawings in the following description are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0021] Figure 1 A schematic diagram of a process flow of an efficient evolution method of a general image segmentation model for drivable area segmentation provided according to a specific embodiment of the present invention is shown;

[0022] Figure 2 A schematic diagram of a lightweight fine-tuning network structure for multi-scale feature extraction provided according to a specific embodiment of the present invention is shown;

[0023] Figure 3 A schematic diagram of the auxiliary task guidance framework structure provided according to a specific embodiment of the present invention is shown. DETAILED DESCRIPTION

[0024] It should be noted that, in the absence of conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present invention and its application or use. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0026] Unless otherwise specifically stated, the relative arrangement, numerical expressions and numerical values ​​of the parts and steps set forth in these embodiments do not limit the scope of the present invention. At the same time, it should be understood that, for ease of description, the sizes of the various parts shown in the accompanying drawings are not drawn according to actual proportional relationships. The technology, methods and equipment known to those of ordinary skill in the relevant art may not be discussed in detail, but in appropriate cases, the technology, methods and equipment should be considered as a part of the specification. In all examples shown and discussed here, any specific value should be interpreted as being merely exemplary, rather than as a limitation. Therefore, other examples of exemplary embodiments may have different values.

[0027] like Figure 1 As shown, according to a specific embodiment of the present invention, a method for constructing and efficiently evolving a visual perception model for drone aerial images is provided, and the method specifically includes:

[0028] S1: Obtain the drivable area segmentation dataset from the perspective of drone aerial photography, complete data preprocessing, and perform data division.

[0029] S2: Construct a lightweight fine-tuning network. The lightweight fine-tuning network consists of two parts: the Adapter network and the VPT network. The Adapter network includes the SPM module, the Injector module, and the Extractor module; the SPM module is used to generate multi-scale feature maps, and the Injector module and the Extractor module are used to interact the generated multi-scale feature maps with the features in the SAM image encoder. The VPT network is used to add trainable tokens at the input of each layer of the Transformer model to indirectly adjust the attention mechanism.

[0030] S3: Build an auxiliary task guidance framework. Add multiple lightweight auxiliary decoders to the original "encoder-decoder" architecture to perform the "vehicle segmentation" auxiliary task. During training, multiple decoders share one encoder and perform decoding separately. Finally, the losses of these decoders are weighted and summed to obtain the final loss.

[0031] S4: Train the segmentation network. Using the loss function set in S3 and the back propagation algorithm, iteratively update and optimize the network parameters until the model loss converges.

[0032] S5: Segment the drivable area. The drone aerial image to be segmented is preprocessed according to the preprocessing method in S1, and then the preprocessed image is input into the trained segmentation network. The output result of the segmentation network model is the drivable area segmentation result.

[0033] By applying this configuration, a method for constructing and efficiently evolving a visual perception model for drone aerial images is provided. This method preprocesses and divides the data, constructs a lightweight fine-tuning network, constructs an auxiliary task guidance framework, trains a segmentation network, and finally performs drivable area segmentation. By fully mining multi-scale feature maps and using traffic scene related information such as vehicle targets, the general image segmentation model is efficiently evolved to the drivable area segmentation task of drone aerial images, thereby improving the performance of the model on this task. Compared with the prior art, the technical solution of the present invention can solve the technical problems in the prior art that the straight-tube general image segmentation model is difficult to directly obtain multi-scale feature maps, and the existing fine-tuning method is difficult to effectively introduce traffic scene information.

[0034] Furthermore, in the present invention, S1 is first executed: a drivable area segmentation data set under the aerial photography perspective of a drone is obtained, data preprocessing is completed, and data division is performed.

[0035] In the present invention, data preprocessing includes label processing and image enhancement. Label processing means retaining only the Road label in UAVid and classifying the rest of the labels as background labels. Image enhancement includes resizing, mosaic enhancement, random cropping, random flipping, etc.

[0036] The data is divided based on the criteria given by the selected dataset.

[0037] Furthermore, in the present invention, S2 is executed to construct a lightweight fine-tuning network. The lightweight fine-tuning network includes two parts: an Adapter network and a VPT network. Among them, the Adapter network includes an SPM (Spatial Prior Module) module, an Injector module, and an Extractor module; the SPM module is used to generate a multi-scale feature map, and the Injector module and the Extractor module are used to interact the generated multi-scale feature map with the features in the SAM image encoder. The VPT network is used to add trainable words at the input of each layer of the Transformer model to indirectly adjust the attention mechanism. The lightweight fine-tuning network structure is as follows Figure 2 shown.

[0038] In the present invention, the lightweight fine-tuning network is built on top of the SAM image encoder. Since the original SAM task head cannot predict semantic labels and cannot be directly used for semantic segmentation tasks, the decoder in the present invention uses the UPerNet semantic segmentation task head instead.

[0039] In the present invention, the specific structures and functions of the three components of the Adapter network are as follows:

[0040] The SPM module adopts a convolutional neural network structure, which includes a Stem network and three convolutional layers, and will obtain four downsampling scales, namely Then, These three multi-scale feature maps are flattened and concatenated to obtain the feature Input to the following series of Injector modules and Extractor modules. represents the i-th multi-scale feature containing spatial prior information, It represents a real-valued feature space of a specific dimension, D represents the number of feature channels, and HW represents the length-width product of the image.

[0041] The Injector module uses a cross-attention mechanism to integrate multi-scale features The spatial prior information contained in is injected into the Transformer Block corresponding to ViT. Specifically, for the i-th Injector module, the input of the i-th Block As a query, multi-scale features As the key (Key) and value (Value), perform attention calculation, and compare the result with the original Add them together and get This enables the injection of spatial prior information. The corresponding formula is shown below.

[0042]

[0043] in, is the input of the ith Block, is the input corresponding to the updated fusion spatial prior information, γ i is a learnable vector used to balance and attention calculation results.

[0044] The Extractor module consists of a cross-attention (Attention) and a feed-forward neural network (FFN), which utilizes the previous multi-scale features And the output of the corresponding Transformer Block in ViT Generating new multi-scale features Specifically, for the i-th Extractor module, multi-scale features As a query, take the output of the i-th Block As keys and values, the attention is calculated and the result is compared with the original Add together to get Then, Projected through a feedforward neural network to obtain new multi-scale features

[0045]

[0046] in, To fuse multi-scale features and the output of the i-th Transformer Block of ViT The new features obtained, It is the multi-scale feature output by the i-th Extractor.

[0047] The 12 layers in the SAM image encoder are divided into 4 groups, each of which interacts with a pair of Injector-Extractor. After multiple iterations, high-quality multi-scale features (with Then, it is combined with the spatial prior module to obtain The scale feature maps are combined to form feature maps with four scales. Finally, the multi-scale features obtained in the Adapter network and the features in the SAM image encoder are fused to obtain a multi-scale feature with richer information, which is then passed to the decoder for decoding.

[0048] Furthermore, in the present invention, S3 is executed to construct an auxiliary task guidance framework. Multiple lightweight auxiliary decoders are added to the original "encoder-decoder" architecture to perform the "vehicle segmentation" auxiliary task. During training, multiple decoders share one encoder and perform decoding separately. Finally, the losses of these decoders are weighted and summed to obtain the final loss. The auxiliary task guidance framework is as follows: Figure 3 shown.

[0049] In the present invention, the auxiliary task head is constructed on the basis of the main task head UPerNet network: the fused features in the UPerNet network are directly used as the input of the auxiliary task decoder. At this time, the auxiliary task decoder only needs to use a simple FCN task head. In this way, there is no need to require the auxiliary task head to also use UPerNet, which reduces the amount of trainable parameters and helps the model evolve more efficiently.

[0050] The labels corresponding to the auxiliary tasks can be obtained directly from the selected dataset (if the dataset has corresponding annotations), or they can be generated using existing tools (if the dataset does not have corresponding annotations). Among them, the "vehicle segmentation" auxiliary task can use the InternImage model to generate auxiliary segmentation maps.

[0051] As a specific embodiment of the present invention, the loss functions used by the main task and the auxiliary task are both cross entropy loss functions, and the final loss function is the weighted sum of the loss functions of each task. For example, the loss weight of the main task is 1.0, and the loss weight of the vehicle segmentation auxiliary task is 0.1.

[0052] In view of the technical defects of the existing efficient fine-tuning methods, the present invention proposes a method for constructing and efficiently evolving a visual perception model for UAV aerial images. This method can solve the problem that the straight-tube general image segmentation model is difficult to directly obtain multi-scale feature maps, and the existing fine-tuning methods are difficult to effectively introduce traffic scene information. By fully mining the multi-scale feature maps and utilizing traffic scene related information such as vehicle targets, the general image segmentation model can be efficiently evolved to the task of segmenting the drivable area of ​​UAV aerial images, thereby improving the performance of the model on this task.

[0053] The beneficial effects of the present invention include:

[0054] (1) A “lightweight fine-tuning network for multi-scale feature extraction” is proposed. This network is built on ViT-Adapter and VPT, which enables it to extract multi-scale features with rich information while fine-tuning the general model, thereby improving the model evolution effect of the general pre-trained model.

[0055] (2) An “auxiliary task guidance framework” was proposed, which introduced three auxiliary tasks: “lane line segmentation”, “road marking segmentation”, and “vehicle segmentation”. A multi-task training framework was designed to train these auxiliary tasks together with the target task, thereby helping the model learn the traffic scene-related knowledge required for the target task.

[0056] In order to further understand the present invention, the following Figures 1 to 3 The present invention describes in detail the construction and efficient evolution method of the visual perception model for drone aerial images.

[0057] A visual perception model construction and efficient evolution method for UAV aerial images, the steps are as follows:

[0058] S1. Dataset preparation: complete data set selection, data preprocessing and data partitioning.

[0059] 1.1 This example selects the UAVid dataset as a drivable area segmentation dataset for UAV aerial images to verify the invention.

[0060] 1.2 Data preprocessing mainly includes label processing and image enhancement. Label processing means retaining only the Road label in UAVid and attributing the rest of the labels to background labels. Image enhancement includes resizing, Mosaic enhancement, random cropping, random flipping, etc. This example selects two enhancement methods: resizing the image to 384×640 pixels and Mosaic enhancement with a probability of 50%.

[0061] 1.3 Data division is based on the standard division given by the selected dataset. For the UAVid dataset, 200 images are used for training, 70 images are used for validation, and 150 images are used for testing. Since the test set labels are not published, this example evaluates the model from the validation set.

[0062] S2. Build a lightweight fine-tuning network.

[0063] The 2.1SAM model is a smaller-scale SAM-B: it contains 12 Transformer layers and the dimension size of the image block embedding is 768.

[0064] 2.2 The hyperparameter settings of the Adapter network are consistent with those of ViT-Adapter-B.

[0065] 2.3 When the VPT network is introduced, 50 trainable words are inserted into each layer of the SAM image encoder, and the dimension of the trainable word is the same as the embedding dimension of SAM.

[0066] S3. Build an auxiliary task guidance framework. Figure 3As shown in Figure 1, the auxiliary task head is built on the basis of the main task head UPerNet. The specific framework design is as described above and will not be repeated here.

[0067] S4. Train the segmentation network. Use cross entropy loss to constrain model training. The specific constraint design is as described above and will not be repeated here.

[0068] The back-propagation algorithm is used to update and optimize the network parameter weights until the model loss area converges.

[0069] In this example, image segmentation model training and evaluation are implemented based on Pytorch and MMSegmentation framework. All experiments are performed on an NVIDIA GeForce RTX 3090. Batch Size is set to 2 and the number of iterations is 80k. The optimizer uses AdamW and the initial learning rate is 6×10 -5 , the weight decay is 0.01, and the weight parameters β1 and β2 are set to 0.9 and 0.999 respectively. The learning rate is warmed up in the first 1500 iterations, and the warming method is linear, with an increase rate of 1×10 -6 .

[0070] S5. After the model training is completed, the drivable area segmentation detection is performed. The drone aerial image to be segmented is preprocessed according to the preprocessing method in S1, and then the preprocessed image is input into the trained segmentation network. The model output result is the drivable area segmentation result.

[0071] After testing, the aAcc of the UAVid dataset is 97.41%, the mAcc is 92.73%, and the mIoU is 87.24%. The results show that the present invention can effectively complete the task of drivable area segmentation.

[0072] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A visual perception model construction and efficient evolution method for drone aerial images, characterized in that: The visual perception model construction and efficient evolution method for drone aerial images includes: S1: Obtain the segmentation dataset of the drivable area from the perspective of drone aerial photography, complete data preprocessing, and perform data division; S2: Construct a lightweight fine-tuning network: The lightweight fine-tuning network consists of two parts: the Adapter network and the VPT network. The Adapter network includes the SPM module, the Injector module, and the Extractor module. The SPM module is used to generate multi-scale feature maps, and the Injector module and the Extractor module are used to interact the generated multi-scale feature maps with the features in the SAM image encoder. The VPT network is used to add trainable tokens at the input of each layer of the Transformer model to indirectly adjust the attention mechanism. S3: Construct an auxiliary task guidance framework: Add multiple lightweight auxiliary decoders to the original "encoder-decoder" architecture to perform the "vehicle segmentation" auxiliary task; during training, multiple decoders share one encoder and perform decoding separately; finally, the losses of these decoders are weighted and summed to obtain the final loss; S4: Train the segmentation network: Use the loss function set in S3 and the back propagation algorithm to iteratively update and optimize the network parameters until the model loss converges; S5: Segment the drivable area: preprocess the drone aerial image to be segmented according to the preprocessing method in S1, and then input the preprocessed image into the trained segmentation network. The output result of the segmentation network model is the drivable area segmentation result.

2. The method for constructing and efficiently evolving a visual perception model for drone aerial images according to claim 1 is characterized in that: Data preprocessing includes label processing and image enhancement; label processing refers to retaining only the Road label in UAVid and attributing the remaining labels to background labels; image enhancement includes resizing, mosaic enhancement, random cropping and random flipping.

3. The method for constructing and efficiently evolving a visual perception model for drone aerial images according to claim 1 is characterized in that: The lightweight fine-tuning network is built on top of the SAM image encoder, and the decoder is replaced with the UPerNet semantic segmentation task head.

4. The method for constructing and efficiently evolving a visual perception model for drone aerial images according to claim 1 is characterized in that: The SPM module adopts a convolutional neural network structure, which includes a Stem network and three convolutional layers to obtain four downsampling scales: The following feature map; The three multi-scale feature maps are flattened and concatenated to obtain the feature Input into the Injector module and the Extractor module; among them, represents the i-th multi-scale feature containing spatial prior information, It represents a real-valued feature space of a specific dimension, D represents the number of feature channels, and HW represents the length-width product of the image.

5. The method for constructing and efficiently evolving a visual perception model for drone aerial images according to claim 4 is characterized in that: The Injector module uses a cross-attention mechanism to integrate multi-scale features The spatial prior information contained in is injected into the Transformer Block corresponding to ViT.

6. The method for constructing and efficiently evolving a visual perception model for drone aerial images according to claim 5 is characterized in that: For the i-th Injector module, take the input of the i-th Block As a query, multi-scale features As keys and values, attention is calculated according to To achieve the injection of spatial prior information, is the input of the ith Block, is the input corresponding to the updated fusion spatial prior information, γ i is a learnable vector used to balance and attention calculation results.

7. The method for constructing and efficiently evolving a visual perception model for drone aerial images according to claim 6 is characterized in that: The Extractor module consists of a cross-attention and feed-forward neural network, utilizing the previous multi-scale features. And the output of the corresponding Transformer Block in ViT Generating new multi-scale features 8. The method for constructing and efficiently evolving a visual perception model for drone aerial images according to claim 7 is characterized in that: according to generate in, To fuse multi-scale features and the output of the i-th Transformer Block of ViT The new features obtained are It is the multi-scale feature output by the i-th Extractor.

9. The method for constructing and efficiently evolving a visual perception model for drone aerial images according to claim 1, characterized in that: The 12 layers in the SAM image encoder are divided into 4 groups, each of which interacts with a pair of Injector-Extractor.

10. The method for constructing and efficiently evolving a visual perception model for drone aerial images according to claim 1, characterized in that: The auxiliary task head is built on the basis of the main task head UPerNet network.

Images