Remote sensing target detection method and system based on dynamic adaptive query

Abstract

The invention discloses a remote sensing target detection method based on dynamic adaptive query, and the method comprises the following steps: obtaining a target detection data set, and carrying out the preprocessing of the target detection data set; the method comprises the following steps of: constructing an initial network structure which is formed by combining a ResNet50 convolutional neural network and a Transform self-attention mechanism; meanwhile, dynamic self-adaptive position query is adopted to replace original position query; training the initial network structure by using the preprocessed target detection data set to obtain a trained target detection network model; and identifying a target in the remote sensing image by using the target detection network model. The method not only improves the target detection precision, but also can improve the model convergence speed.

Description

Technical Field

[0001] The present invention relates to the field of computer vision detection, and in particular to a remote sensing target detection method and system based on dynamic adaptive query. Background Art

[0002] Object detection is one of the basic tasks in computer vision. Object detection is the superposition of classification and regression problems, and is one of the most challenging problems in the field of computer vision. Among them, remote sensing image object detection has been a research hotspot in recent years and is widely used in vehicle detection, ship detection, urban planning, environmental monitoring, aviation and other fields. Unlike ordinary RGB images, remote sensing images mostly come from various remote sensing platforms such as satellite drones. Remote sensing images contain geometric and physical information of the earth's surface. Because aerial remote sensing images cover a wide range, the images taken from overhead contain many small targets with uncertain directions, and the background complexity of the images is high. One picture contains a variety of backgrounds, and the cluttered background will seriously interfere with the results of target detection.

[0003] The DETR model combines CNN with Transformer for the first time, treating object detection as a set prediction problem to achieve end-to-end object detection. By leveraging the inherent self-attention mechanism in Transformer, DETR is able to model global context in images, thereby enhancing its ability to recognize objects in complex backgrounds. More importantly, compared with the previous CNN-based two-stage object detection framework, DETR's end-to-end training paradigm abandons manually designed components such as anchor boxes and NMS, simplifying the detection process and potentially reducing the errors introduced by this training strategy.

[0004] Although DETR has a novel design, the computational complexity is high because the DETR model needs to encode and decode the entire image, and the query is not clearly defined. These factors lead to a relatively slow convergence speed of the DETR model. Specifically, first, DETR requires a long period of training to converge, with up to 500 training rounds. Second, since the original attention mechanism in DETR has almost uniform weights for all pixels on the feature map, each query needs to be calculated with the keys at all positions, resulting in many invalid calculations and increasing the computational complexity of the model.

[0005] In order to speed up the convergence of the model and improve performance. As a variant of DETR, Deformable DETR proposes multi-scale deformable attention, which can handle features of four different scales. For each query, the key is only sampled in the local position, and the value is also the value of the partial position. It only focuses on a small number of key sampling points around the reference point, which greatly improves the convergence speed. Deformable DETR has made great breakthroughs in the detection performance of small objects, but since each query in Deformable DETR is responsible for a relatively large area, it will inevitably lead to internal conflicts between multiple queries, which brings fuzzy spatial priors to the model. In recent years, scholars have begun to improve the design of queries. DABDETR (dynamic anchor box DETR) redefines the object query in DETR, uses a 4-dimensional anchor box, adds width and height to the definition of object queries for size modulation, and introduces dynamic query design. Object queries will be updated layer by layer.

[0006] Although DETR has shown great potential in remote sensing target detection, it also faces some challenges. DETR variants are not suitable for remote sensing images with cluttered backgrounds and dense targets, because DETR variants do not consider query settings and use the same set of initialization queries for all input samples. The query is single and unchanged for different inputs, so the performance of the DETR model will degrade when processing remote sensing images, especially for small targets and overlapping targets. Summary of the invention

[0007] Purpose of the invention: The purpose of the present invention is to address the deficiencies in the prior art and provide a remote sensing target detection method based on dynamic adaptive query, which is improved based on DETR and its variants to solve the problems of single query content and inability to adapt to input.

[0008] Another aspect of the present invention is to provide a remote sensing target detection system based on dynamic adaptive query.

[0009] In order to achieve the above object, the present invention first provides a remote sensing target detection method based on dynamic adaptive query, comprising the following steps:

[0010] S1. Obtain a target detection data set and preprocess the target detection data set;

[0011] S2. Construct an initial network structure composed of a ResNet50 convolutional neural network and a Transformer self-attention mechanism; and use dynamic adaptive location query to replace the original location query;

[0012] S3, using the preprocessed target detection data set to train the initial network structure to obtain a trained target detection network model;

[0013] S4. Use the target detection network model to identify targets in remote sensing images.

[0014] A further preferred technical solution of the present invention is that the target detection dataset obtained in step S1 includes a COCO dataset and an RSOD remote sensing dataset.

[0015] Preferably, the target detection data set is preprocessed in step S1, and the specific steps are:

[0016] S11, resize the images in the target detection dataset to 1024×1024;

[0017] S12, the image first passes through the feature extraction network to obtain multi-scale image features, and then the multi-scale image features are transformed to form a feature pyramid network;

[0018] S13, passing the transformed multi-scale features and the corresponding position codes together through an encoder to obtain features enhanced by the encoder.

[0019] Preferably, in step S2, an initial network structure composed of a ResNet50 convolutional neural network and a Transformer self-attention mechanism is constructed; and a dynamic adaptive location query is used to replace the original location query. The specific method is:

[0020] The network structure includes a feature extraction network, a Transformer encoder and decoder, and a dynamic adaptive weight generation module;

[0021] S21, select ResNet50 as the feature extraction network, take the top layer features of the feature pyramid network as the input of the dynamic adaptive weight generation module, and obtain a set of dynamic weights; initialize a set of queries, and linearly combine the set of queries with the dynamic weights obtained by the dynamic adaptive weight generation module to obtain dynamic adaptive position queries;

[0022] S22. Send the encoder features and dynamic adaptive queries to the decoder for attention operation; finally, the decoder obtains the prediction set by stacking multiple attention operations and then passing through the classification head and regression head.

[0023] Preferably, the dynamic weight is obtained by a dynamic adaptive weight generation module, and the specific method is as follows:

[0024] S211. Use the global adaptive average pooling layer to obtain global features, and then obtain a set of unnormalized weights through two fully connected layers and RELU activation function. Apply the softmax function to the generated weights, map the input values ​​to probability real numbers between 0 and 1, and scale them by the temperature coefficient to obtain the corresponding weight values ​​of each group. The weight generation method is:

[0025]

[0026] where F " is the output of the second FC layer in the module, and τ is the set temperature coefficient. - represents the original weight value of the jth element in this set of weights. k is a specific index used to select a specific element from this set of weights for softmax calculation, and j is an index that traverses all elements in this set of weights and is used to calculate the normalization factor of the softmax function.

[0027] As a preferred method, the specific method of obtaining the dynamic adaptive location query is:

[0028] S212, first, redefine the object query, the object query is a set of predefined vectors, each object query represents a description of a target category and location, and n sets of learnable random vectors are set during initialization;

[0029] S213, the initial position query input to the first layer of the decoder is defined as:

[0030]

[0031] where q k Represents a set of random vectors for initialization, ω k Represents a set of weights, which are dynamic weights obtained by the dynamic adaptive weight generation module according to the input, thereby transforming the original location query into a dynamic adaptive location query.

[0032] Preferably, in step S3, the initial network structure is trained using the preprocessed target detection data set to obtain a trained target detection network model; the specific method is:

[0033] S31. Use the training set of the COCO dataset to train the model on RTX 4090 for 50 rounds, and set the number of queries to 300. After the COCO dataset training is completed, the model will generate the corresponding weight file, and the model trained on the COCO dataset will be used as the pre-trained model.

[0034] S32, importing the pre-trained model with trained weights, training again according to the training set of the RSOD remote sensing data set, fine-tuning the parameters of the model, and obtaining a trained target detection network model;

[0035] S33. Use the verification sets of the COCO dataset and the RSOD remote sensing dataset to verify the target detection network model.

[0036] Preferably, the hyperparameters set in the target detection network model training in step S3 are:

[0037] The size of the input image is set to 1024×1024, the number of categories of detected objects in the COCO dataset is 80, the number of categories of detected objects in the RSOD remote sensing dataset is 4, and the batch size is set to 2; the model is trained for a total of 50 epochs, and the learning rate is dropped to 0.1 in the 40th epoch; the initial learning rate is 0.0001, the weight decay is 0.0001, and the number of queries is set to 300.

[0038] In another aspect, the present invention provides a remote sensing target detection system based on dynamic adaptive query, comprising:

[0039] A data preprocessing module is used to obtain a target detection data set and preprocess the target detection data set;

[0040] The model building module is used to build an initial network structure composed of a ResNet50 convolutional neural network and a Transformer self-attention mechanism, using dynamic adaptive location query instead of the original location query;

[0041] The model training module is used to set model parameters, train the initial network structure using the preprocessed target detection data set, and obtain a trained target detection network model;

[0042] The target detection module is used to identify targets in remote sensing images using a target detection network model.

[0043] Preferably, when setting model parameters, the model training module configures appropriate model parameters according to the hardware device actually used and the expected detection results, including batch size, optimizer, learning rate, number of training rounds, number of training rounds for adjusting the learning rate, and feature extraction network;

[0044] The target detection module detects the image to be detected using the trained target detection network model and the corresponding weight file to obtain the target detection result, and compares the detection result with the inference FPS to see whether it meets the expectation.

[0045] Beneficial effects: (1) The basic framework of the present invention uses the DAB-Deformable-DETR network structure, redefines the Object query in DETR, and uses dynamic adaptive location query to replace the original location query. The use of dynamic adaptive query can provide better location priors, so that the DETR series model can dynamically adjust the location query for different images. The model is more accurate in predicting the target location and can capture the boundary of the target more accurately. It is more sensitive to the details in the image, can capture more subtle differences, and can make more accurate predictions for remote sensing images under complex backgrounds. This method not only improves the target detection accuracy, but also improves the model convergence speed.

[0046] (2) To address the problem of fixed and single query content, the present invention designs a dynamic adaptive weight generation module, which dynamically adjusts the weight according to the input, considers context information from a global perspective to adapt to various goals, and provides more accurate location priors for model training iterations.

[0047] (3) The present invention adopts dynamic adaptive query, and dynamically adjusts the query according to the input image features, so that the area responsible for each query is more adapted to the image and the prediction box is more suitable for the target. Especially for remote sensing images with complex backgrounds, the method proposed in the present invention can better identify dense or overlapping target instances and achieve more accurate query optimization. In addition, the method proposed in the present invention can be easily integrated into other existing DETR variants to further improve the detection performance of remote sensing image targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 Schematic diagram of a target detection network model in an embodiment of the present invention;

[0049] Figure 2 It is a display diagram of the detection effect in the embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments, and they should not be understood as limitations on the present invention. Based on the embodiments of 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. In the description of the present invention, it should be understood that the terms used are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0051] Combine the following Figure 1-Figure 2 The invention describes a remote sensing target detection method and system based on dynamic adaptive query.

[0052] Embodiment 1: This embodiment provides a remote sensing target detection method based on dynamic adaptive query, comprising the following steps:

[0053] S1. Obtain a target detection data set and preprocess the target detection data set;

[0054] S2. Construct an initial network structure composed of a ResNet50 convolutional neural network and a Transformer self-attention mechanism; and use dynamic adaptive location query to replace the original location query;

[0055] S3, using the preprocessed target detection data set to train the initial network structure to obtain a trained target detection network model;

[0056] S4. Use the target detection network model to identify targets in remote sensing images.

[0057] The present invention designs a remote sensing target detection method of improved DETR to improve the detection accuracy of remote sensing image targets. The present invention further improves DAB-Deformable-DETR. Specifically:

[0058] The target detection dataset obtained in step S1 includes the COCO dataset and the RSOD remote sensing dataset.

[0059] The COCO dataset is a large and rich dataset for object detection, segmentation and captioning. This dataset aims at scene understanding and is mainly captured from complex daily scenes. The objects in the images are calibrated by precise segmentation. The images include 91 categories of objects, 328,000 images and 2,500,000 labels. It is the largest dataset for semantic segmentation so far, providing 80 categories, more than 330,000 images, of which 200,000 are annotated, and the number of individuals in the entire dataset exceeds 1.5 million.

[0060] The RSOD remote sensing dataset is an open dataset for object detection in remote sensing images. The dataset was released by Wuhan University in 2015 and contains four types of targets: airplanes, playgrounds, overpasses, and oil drums. The numbers are: 4993 airplanes in 446 images, 191 playgrounds in 189 images, 180 overpasses in 176 images, and 1586 oil drums in 165 images.

[0061] First, the target detection data set needs to be preprocessed. The specific steps are:

[0062] S11, resize the images in the target detection dataset to 1024×1024;

[0063] S12, the image first passes through the feature extraction network Backbone to obtain multi-scale image features C3, C4, C5, and then transforms the multi-scale image features to obtain P3, P4, P5, P6 to form a feature pyramid network;

[0064] S13, passing the transformed multi-scale features and the corresponding position codes together through an encoder to obtain features enhanced by the encoder.

[0065] Then, an initial network structure consisting of a ResNet50 convolutional neural network and a Transformer self-attention mechanism is constructed; at the same time, dynamic adaptive location query is used to replace the original location query.

[0066] Therefore, the network structure includes a feature extraction network, a Transformer encoder and decoder, and a dynamic adaptive weight generation module. The specific construction method is:

[0067] S21. Select ResNet50 as the feature extraction network, take the features of the P3 layer of the feature pyramid network as the input of the dynamic adaptive weight generation module, and obtain a set of dynamic weights; initialize a set of queries, and linearly combine this set of queries with the dynamic weights obtained by the dynamic adaptive weight generation module to obtain dynamic adaptive position queries.

[0068] In this step, the dynamic weight is obtained by the dynamic adaptive weight generation module, and the specific method is as follows:

[0069] S211, use the global adaptive average pooling layer to obtain global features, and then use two fully connected layers and RELU activation function to obtain a set of unnormalized weights. RELU activation function can increase the nonlinearity of the network. Apply the softmax function to the generated weights, map the input values ​​to probability real numbers between 0 and 1, and scale them by the temperature coefficient to obtain the corresponding weight values ​​for each group. The temperature coefficient is used to control the smoothness of softmax to adjust the distribution of weights so as to better dynamically generate weights and improve the expressiveness of the model. The weight generation method is:

[0070]

[0071] where F " is the output of the second FC layer in the module, and τ is the set temperature coefficient. - represents the original weight value of the jth element in this set of weights. k is a specific index used to select a specific element from this set of weights for softmax calculation, and j is an index that traverses all elements in this set of weights and is used to calculate the normalization factor of the softmax function.

[0072] The specific method to obtain dynamic adaptive query is:

[0073] S212. First, redefine the object query. Object query is an innovative concept proposed by DETR. Object query is a set of predefined vectors. Each object query represents a description of a target category and location. In order to make the query better adapt to the complex background of remote sensing images and more accurately locate the target under complex backgrounds, dynamic adaptive query is used to replace the original query. n sets of learnable random vectors are set during initialization.

[0074] S213, the initial position query input to the first layer of the decoder is defined as:

[0075]

[0076] where q k Represents a set of random vectors for initialization, ω k Represents a set of weights, which are the dynamic weights obtained by the dynamic adaptive weight generation module according to the input, ω k It is not a fixed value but is dynamically adjusted according to the input, thus turning the original query into a dynamic adaptive query.

[0077] S22, the encoder features and dynamic adaptive queries are sent to the decoder for attention operation; finally, the decoder obtains the prediction set by stacking multiple attention operations and then passing through the classification head and regression head. The predicted box is matched with the true value box through the Hungarian matching algorithm. The Hungarian algorithm performs binary matching so that each object query has a unique matching target in theory, and there is no overlap, so DETR does not need nms for post-processing. Then the category loss is calculated, using the cross entropy loss, followed by the regression loss calculation, using the L1 loss and GIOU loss.

[0078] After the initial network structure is constructed, the preprocessed target detection data set is used to train the initial network structure to obtain a trained target detection network model; the specific method is:

[0079] S31. Use the training set of the COCO dataset to train the model on RTX 4090 for 50 rounds, and set the number of queries to 300. After the COCO dataset training is completed, the model will generate the corresponding weight file, and the model trained on the COCO dataset will be used as the pre-trained model.

[0080] S32, importing the pre-trained model with trained weights, training again according to the training set of the RSOD remote sensing data set, fine-tuning the parameters of the model, and obtaining a trained target detection network model;

[0081] The hyperparameters set during model training are:

[0082] The size of the input image is set to 1024×1024, the number of categories of detected objects in the COCO dataset is 80, the number of categories of detected objects in the RSOD remote sensing dataset is 4, and the batch size is set to 2; the model is trained for a total of 50 epochs, and the learning rate is dropped to 0.1 in the 40th epoch; the initial learning rate is 0.0001, the weight decay is 0.0001, and the number of queries is set to 300.

[0083] S33, the target detection network model is verified by using the validation set of COCO dataset and RSOD remote sensing dataset. In order to verify the effectiveness of the method of the present invention, the results of the COCO2017 validation set are first compared. Among them, the mAP of the original DAB-Def-DETR model is 46.9%, and it converges in 50 rounds. The mAP of the method of the present invention is 48.5%, and it achieves the best effect and converges in 46 rounds. The results show that the method proposed in the present invention improves the mAP by 1.6% and accelerates the convergence of the model. The experimental results obtained by the method of the present invention on the RSOD remote sensing image dataset are shown in Table 1. Compared with DAB-Def-DETR, the method of the present invention brings a mAP gain of 1.1%. In the RSOD remote sensing dataset, it is compared with the classic Faster-RCNN, FCOS and advanced OYOLO, SOLO-Net and LAR-YOLOv8 remote sensing target detection algorithms. The mAP of the method of the present invention reaches 95.8%. The results show that the method proposed in the present invention can improve the performance of the model, and the detection effect in remote sensing images is better than most traditional remote sensing target detection methods based on CNN.

[0084] Figure 2 The experimental results of the embodiments of the present invention are shown. Figure 2 (a) is the original image. Figure 2 (b) is the detection effect diagram of the unimproved DAB-Def-DETR. Figure 2 (c) is a detection effect diagram of the method proposed in the present invention. It can be observed that the method proposed in the present invention can detect smaller targets that are easily missed, and the positioning of the target is also more accurate.

[0085] Embodiment 2: This embodiment provides a remote sensing target detection system based on dynamic adaptive query, including:

[0086] The data preprocessing module is used to obtain the target detection data set and preprocess the target detection data set.

[0087] The model building module is used to build an initial network structure composed of a ResNet50 convolutional neural network and a Transformer self-attention mechanism, using dynamic adaptive location query to replace the original location query.

[0088] The model training module is used to set model parameters, train the initial network structure using the preprocessed target detection data set, and obtain a trained target detection network model;

[0089] When setting model parameters, the model training module configures appropriate model parameters according to the actual hardware equipment used and the expected detection results, including batch size, optimizer, learning rate, number of training rounds, number of training rounds for adjusting the learning rate, and feature extraction network.

[0090] The target detection module is used to identify targets in remote sensing images using a target detection network model;

[0091] For the image to be detected, the trained target detection network model and the corresponding weight file are used to detect the image to be detected to obtain the target detection result, and the detection result is compared with the inference FPS to see whether it meets the expectations.

[0092] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes in form and details may be made without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A remote sensing target detection method based on dynamic adaptive query, characterized in that: The steps include: S1. Obtain a target detection data set and preprocess the target detection data set; S2. Construct an initial network structure composed of a ResNet50 convolutional neural network and a Transformer self-attention mechanism; and use dynamic adaptive location query to replace the original location query; S3, using the preprocessed target detection data set to train the initial network structure to obtain a trained target detection network model; S4. Use the target detection network model to identify targets in remote sensing images.

2. The remote sensing target detection method based on dynamic adaptive query according to claim 1, characterized in that: The target detection dataset obtained in step S1 includes the COCO dataset and the RSOD remote sensing dataset.

3. The remote sensing target detection method based on dynamic adaptive query according to claim 2, characterized in that: Step S1 is to preprocess the target detection data set, and the specific steps are as follows: S11, resize the images in the target detection dataset to 1024×1024; S12, the image first passes through the feature extraction network to obtain multi-scale image features, and then the multi-scale image features are transformed to form a feature pyramid network; S13, passing the transformed multi-scale features and the corresponding position codes together through an encoder to obtain features enhanced by the encoder.

4. The remote sensing target detection method based on dynamic adaptive query according to claim 3 is characterized in that: In step S2, an initial network structure composed of a ResNet50 convolutional neural network and a Transformer self-attention mechanism is constructed; at the same time, a dynamic adaptive location query is used to replace the original location query; the specific method is: The network structure includes a feature extraction network, a Transformer encoder and decoder, and a dynamic adaptive weight generation module; S21, select ResNet50 as the feature extraction network, take the top layer features of the feature pyramid network as the input of the dynamic adaptive weight generation module, and obtain a set of dynamic weights; initialize a set of queries, and linearly combine the set of queries with the dynamic weights obtained by the dynamic adaptive weight generation module to obtain dynamic adaptive position queries; S22, sending encoder features and dynamic adaptive queries to the decoder for attention operation; Finally, the decoder obtains a set of predictions by stacking multiple attention operations and then passing them through the classification head and regression head.

5. The remote sensing target detection method based on dynamic adaptive query according to claim 4 is characterized in that: The dynamic weight is obtained through the dynamic adaptive weight generation module. The specific method is as follows: S211. Use the global adaptive average pooling layer to obtain global features, and then obtain a set of unnormalized weights through two fully connected layers and RELU activation function. Apply the softmax function to the generated weights, map the input values ​​to probability real numbers between 0 and 1, and scale them by the temperature coefficient to obtain the corresponding weight values ​​of each group. The weight generation method is: where F " is the output of the second FC layer in the module; τ is the set temperature coefficient; F - represents the original weight value of the jth element in this set of weights; k is a specific index used to select a specific element from this set of weights for softmax calculation, and j is an index that traverses all elements in this set of weights and is used to calculate the normalization factor of the softmax function.

6. The remote sensing target detection method based on dynamic adaptive query according to claim 5, characterized in that: The specific method to obtain dynamic adaptive location query is: S212, first, redefine the object query, the object query is a set of predefined vectors, each object query represents a description of a target category and location, and n sets of learnable random vectors are set during initialization; S213, the initial position query input to the first layer of the decoder is defined as: where q k Represents a set of random vectors for initialization, ω k Represents a set of weights, which are dynamic weights obtained by the dynamic adaptive weight generation module according to the input, thereby transforming the original location query into a dynamic adaptive location query.

7. The remote sensing target detection method based on dynamic adaptive query according to claim 2, characterized in that: Step S3 uses the preprocessed target detection data set to train the initial network structure to obtain a trained target detection network model; the specific method is: S31. Use the training set of the COCO dataset to train the model on RTX 4090 for 50 rounds, and set the number of queries to 300. After the COCO dataset training is completed, the model will generate the corresponding weight file, and the model trained on the COCO dataset will be used as the pre-trained model. S32, importing the pre-trained model with trained weights, training again according to the training set of the RSOD remote sensing data set, fine-tuning the parameters of the model, and obtaining a trained target detection network model; S33. Use the verification sets of the COCO dataset and the RSOD remote sensing dataset to verify the target detection network model.

8. The remote sensing target detection method based on dynamic adaptive query according to claim 7, characterized in that: The hyperparameters set in the target detection network model training in step S3 are: The size of the input image is set to 1024×1024, the number of categories of detected objects in the COCO dataset is 80, the number of categories of detected objects in the RSOD remote sensing dataset is 4, and the batch size is set to 2; the model is trained for a total of 50 epochs, and the learning rate is dropped to 0.1 in the 40th epoch; the initial learning rate is 0.0001, the weight decay is 0.0001, and the number of queries is set to 300.

9. A remote sensing target detection system based on dynamic adaptive query, characterized in that: include: A data preprocessing module is used to obtain a target detection data set and preprocess the target detection data set; The model building module is used to build an initial network structure composed of a ResNet50 convolutional neural network and a Transformer self-attention mechanism, using dynamic adaptive location query instead of the original location query; The model training module is used to set model parameters, train the initial network structure using the preprocessed target detection data set, and obtain a trained target detection network model; The target detection module is used to identify targets in remote sensing images using a target detection network model.

10. The remote sensing target detection system based on dynamic adaptive query according to claim 9, characterized in that: When setting model parameters, the model training module configures appropriate model parameters according to the actual hardware device used and the expected detection results, including batch size, optimizer, learning rate, number of training rounds, number of training rounds for adjusting the learning rate, and feature extraction network; The target detection module detects the image to be detected using the trained target detection network model and the corresponding weight file to obtain the target detection result, and compares the detection result with the inference FPS to see whether it meets the expectation.

Images