Small target detection method, system and device in high-altitude overlook scene and medium
By improving the structure and training method of the RT-DETR model, the computational complexity and accuracy problems of small target detection in high-altitude top-down scenarios are solved, achieving efficient end-to-end target detection, which is suitable for application scenarios such as smart construction sites.
Patent Information
- Application Number
- CN202410818773.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-24
- Publication Date
- 2024-11-15
AI Technical Summary
In high-altitude, top-down scenarios, small target detection faces challenges such as high computational complexity, limited detection capability for small targets, and poor detection results. Existing technologies struggle to meet the demands for high accuracy and real-time performance.
An improved RT-DETR model is adopted. By reorganizing the model structure and updating the Block module, and combining convolutional and batch normalization layers, a new FasterNet-Block block is constructed. The model is trained and validated using the COCO2017 dataset to achieve end-to-end small object detection.
The performance of the RT-DETR model in small target detection in high-altitude overhead scenarios has been improved, the model design and training process has been simplified, and high-precision and real-time target detection has been achieved.
Smart Images

Figure CN118968011A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of target detection technology. Specifically, the present invention is applied to the field of smart mine and smart city construction, and in particular to a small target detection method, system, equipment and medium in a high-altitude bird's-eye view scene. Background Art
[0002] In recent years, with the rapid development of deep learning and computer vision technology, safety monitoring systems in smart construction sites have become crucial; especially in high-altitude work scenarios, ensuring the safety of construction site personnel is a key challenge; computer vision technology has been widely used in smart construction sites, fault diagnosis, motion recognition and other fields; intelligent monitoring systems based on image acquisition equipment and target detection algorithms have been successful in many tasks; however, small target detection in high-altitude bird's-eye view scenarios still faces many challenges.
[0003] For example, in high-altitude work scenarios, using cameras to detect people on the ground is a critical task; compared with traditional large and medium target detection, small target detection in high-altitude bird's-eye view scenarios faces greater challenges, including small-sized objects, low resolution, small amounts of information, unbalanced samples, high requirements for position accuracy, low recognition efficiency, and complex detection backgrounds; traditional target detection models are difficult to achieve satisfactory detection results in such scenarios.
[0004] The target detection algorithms used in the prior art are mainly divided into one-stage methods and two-stage methods.
[0005] Two-stage methods, such as R-CNN and Fast R-CNN, first generate candidate boxes and then identify the objects in the boxes. These methods have high accuracy and flexibility, but they have high computational complexity and are not suitable for small target detection, especially in high-altitude bird's-eye view scenes. Due to the large image scale and high computational complexity, local features of small targets are easily ignored, resulting in poor detection results.
[0006] One-stage methods, such as the YOLO model, directly obtain prediction results, are simple and efficient, and are suitable for scenarios such as real-time detection; however, these methods have relatively low accuracy and limited detection capabilities for small objects.
[0007] In summary, both the one-stage method and the two-stage method are not suitable for small target detection in high-altitude bird's-eye view scenes; the main reason is that the computational complexity of large-scale images is high, and the local features of small targets are easily ignored, resulting in poor detection results. Therefore, for this special scene, further research and improvement of target detection algorithms are needed to improve their applicability and accuracy. Summary of the invention
[0008] The object of the present invention is to provide a small target detection method, system, device and medium in a high-altitude bird's-eye view scene, thereby solving all or one of the above-mentioned problems existing in the prior art.
[0009] In order to solve the above technical problems, the specific technical solutions of the present invention are as follows: On the one hand, the present invention provides a small target detection method in a high-altitude overlooking scene, comprising the following steps: Dataset collection: Use COCO2017 as the dataset for model training and validation; Model modification and training: Modify and train the model based on the model structure reorganization strategy and Block module update strategy; Model Validation: Using the data set to verify the modified and trained model; Model call: Call the verified model to perform the target detection task.
[0010] As an improved solution, the model structure reorganization strategy includes: Utilize multi-way structure during training; In the inference phase, multiple computing modules are fused into a single path model.
[0011] As an improved solution, the model structure reorganization strategy further includes: The convolution and batch normalization layers are combined to merge the Conv layer with the subsequent BN layer. Transform Step 1 to Step 2 by converting Conv1x1 and identity to equivalent Conv3x3 form; From Step 2 to Step 3, multiple two-dimensional convolution kernels of compatible sizes are operated on the same input with the same step amplitude to produce outputs of the same resolution. The outputs are summed to obtain equivalent kernels.
[0012] As an improved solution, the Block module update strategy includes: The FasterNet-Block module is combined with some convolution modules, batch normalization and rectified linear units to form a new FasterNet-Block block.
[0013] As an improved solution, the model adopts the RT-DETR model.
[0014] As an improved solution, the Block module update strategy further includes: The new FasterNet-Block block is used to replace the BasicBlock in the Backbone of the RT-DETR model.
[0015] As an improved solution, the use of the data set to verify the modified and trained model includes: The COCO2017 is used to verify the performance of the modified and trained RT-DETR model from the perspectives of precision and recall.
[0016] On the other hand, the present invention also provides a small target detection system in a high-altitude overlooking scene, comprising: The dataset collection module is used to: use COCO2017 as a dataset for model training and verification; Model modification and training module, used to: modify and train the model based on the model structure reorganization strategy and Block module update strategy; A model verification module, used to: verify the modified and trained model using the data set; The model calling module is used to call the verified model to perform the target detection task.
[0017] On the other hand, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the small target detection method in the high-altitude bird's-eye view scene are implemented.
[0018] On the other hand, the present invention further provides a computer device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; wherein: The memory is used to store computer programs; The processor is used to execute the steps of the small target detection method in the high-altitude bird's-eye view scene by running the program stored in the memory.
[0019] The beneficial effects of the technical solution of the present invention are: 1. The small target detection method in the high-altitude bird's-eye view scene described in the present invention can improve the small target detection performance of RT-DERT, enabling RT-DETR to achieve end-to-end target detection without the need to use the manually designed feature extractors, anchor box generators, and non-maximum suppression steps required by traditional target detection methods, thereby simplifying the design and training process of the target detection model.
[0020] 2. The small target detection system in the high-altitude bird's-eye view scene described in the present invention can improve the small target detection performance of RT-DERT through the mutual cooperation of system modules, so that RT-DETR can achieve end-to-end target detection without the need to use the manually designed feature extractors, anchor box generators, and non-maximum suppression steps required by traditional target detection methods, thereby simplifying the design and training process of the target detection model.
[0021] 3. The computer-readable storage medium described in the present invention can realize the cooperation of the guidance system module, thereby realizing the small target detection method in the high-altitude bird's-eye view scene described in the present invention, and the computer-readable storage medium described in the present invention also effectively improves the operability of the small target detection method in the high-altitude bird's-eye view scene.
[0022] 4. The computer device described in the present invention can store and execute the computer-readable storage medium, thereby realizing the small target detection method in the high-altitude bird's-eye view scene described in the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0024] Figure 1 1 is a flow chart of a small target detection method in a high-altitude overlooking scene according to Embodiment 1 of the present invention; Figure 2 1 is a schematic diagram of the structure of the RT-DETR model in the small target detection method in the high-altitude overlooking scene described in Example 1 of the present invention; Figure 3 1 is a schematic diagram of a single-path structure of the RT-DETR model of the small target detection method in the high-altitude overlooking scene described in Example 1 of the present invention after structural reorganization; Figure 4 1. It is a schematic diagram of the principle of conversion between a multi-path model and a single-path model of the RT-DETR model of the small target detection method in the high-altitude overlooking scene described in Example 1 of the present invention; Figure 5 1 is a schematic diagram of the structure of the FasterNet-Block module in the small target detection method in the high-altitude overlooking scene described in Example 1 of the present invention; Figure 6 1 is a schematic diagram of the structure of the improved BasicBlock module in the small target detection method in the high-altitude overlooking scene described in Example 1 of the present invention; Figure 7 Schematic diagram of the architecture of a small target detection system in a high-altitude overlooking scene according to Embodiment 2 of the present invention; Figure 8 is a schematic diagram of the structure of the computer device described in Example 4 of the present invention; The symbols in the accompanying drawings are explained as follows: 1501, processor; 1502, communication interface; 1503, memory; 1504, communication bus. DETAILED DESCRIPTION
[0025] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.
[0026] In the description of the present invention, it should be noted that the embodiments described in the present invention are only part of the embodiments of the present invention, rather than all of the embodiments; based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present invention.
[0027] The terms "first", "second", etc. in the specification and claims of this article and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of this article described here can be implemented in an order other than those illustrated or described here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, device, product or equipment that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or equipment.
[0028] In the description of the present invention, it should be noted that RT-DETR (Real-Time DEtection TRansformer) is an end-to-end real-time target detection model proposed by the Baidu PaddlePaddle team in 2023; its original design intention is to achieve real-time target detection while maintaining high precision to meet the actual needs of the industry. RT-DETR introduces a new decoder variant, called HybridEncoder, to decouple the simultaneous interactions within and between scales. This variant consists of two parts: AIFI (Attention-based Intra-scale Feature Interaction) and CCFM (CNN-based Cross-scale Feature-fusion Module). With this design, the model can effectively reduce the amount of computation while maintaining high accuracy and achieve real-time target detection. The AIFI part is an intra-scale feature interaction module based on the attention mechanism, which is used to capture the association between different features within the same scale. CCFM is an inter-scale feature fusion module based on convolutional neural networks, which is used to effectively fuse features of different scales to improve the performance of target detection. RT-DETR was compared on the COCOval2017 dataset, and the results showed that its performance was better than most mainstream models such as YOLOv5, YOLOv6, YOLOv7, and YOLOv8. This shows that RT-DETR meets the needs of real-time target detection while maintaining high accuracy, providing a highly practical solution for the industry.
[0029] In the description of the present invention, it should be noted that RT-DETR is a variant of a target detection model based on DETR (Detection Transformer); DETR is an end-to-end target detection model that transforms the target detection problem into a set prediction problem through the Transformer architecture, thereby achieving the prediction of the target position and category directly from the image without using traditional anchor boxes, NMS (Non-Maximum Suppression) and other means.
[0030] In the description of the present invention, it should be noted that RT-DETR is improved on the basis of DETR, and its network structure is shown as follows Figure 2 As shown in the figure, the RoI Transformer module is mainly introduced to improve its detection performance for small targets; RoI Transformer is a regional attention mechanism that can better handle small targets during the detection process and improve detection accuracy.
[0031] In the description of the present invention, it should be noted that the network structure of RT-DETR mainly includes the following parts: Backbone network: Transformer or CNN is usually used as the Backbone network to extract image features. Positional Encoding: Encode the pixel position information in the image into the feature vector so that the Transformer model can process the sequence information; Encoder: It is composed of multiple layers of Transformer encoders and is used to encode image features; Decoder: It is composed of multiple layers of Transformer decoders and is used to generate predictions of target categories and positions; RoITransformer module: It is used to process the detected target area and improve the detection accuracy of small targets; Post-processing module: Post-process the model output, such as decoding the position of the prediction box, applying NMS and other operations. Example 1
[0032] This embodiment provides a small target detection method in a high-altitude overlooking scene, such as Figure 1 to Figure 6 As shown, the following steps are included: The core function of this method is to improve the RT-DETR model, increase the small target detection speed of RT-DETR, and solve the small target missed detection and accuracy problems of the target detection model in complex scenes. The details are as follows: S100, Dataset Collection: Specifically, we use the COCO2017 dataset released by Microsoft to train and verify the model; S200, model modification and training: Specifically, the RT-DETR model is modified and a perfect RT-DETR model is trained; Specifically, since the performance of the multi-branch network model is significantly improved, it can utilize the receptive fields of different convolution kernels; however, the multi-path structure leads to a sharp increase in the demand for video memory, which places high demands on the device; in order to balance the model performance and resource consumption, in this implementation, a model structure reorganization strategy is adopted for model training; during training, the multi-path structure is used to give full play to the parallel computing capability of the GPU; in the inference stage, multiple computing modules are fused into a single path model to increase the training speed; Among them, two methods are designed in the model structure reorganization design: one method is to merge the convolution layer and the subsequent batch normalization (BN) layer into a new operation, which can reduce the number of convolution operations in the model while keeping the input and output sizes unchanged, thereby simplifying the model structure and improving performance; the other method is to add the results of multiple parallel convolution branches of the same specification for the same input and then perform a convolution operation, so that multiple convolution kernels can be merged into one, and then a single convolution operation is performed on the input to obtain the same result as the branch operation; the same specifications here include the size, padding and stride of the convolution kernel, because the batch normalization layer is usually connected after the convolution operation; these methods effectively improve the performance of the detection model without affecting the accuracy; in summary, the single-path structure after structural reorganization is shown as follows Figure 3 ; Convert the multi-channel model into a single-channel model such as Figure 4 As shown in the figure; in this process, the fusion of convolution and batch normalization (BN) layers is first used to merge the Conv layer with the subsequent BN layer, and the transition from Step1 to Step2 is realized by converting Conv1x1 and identity into equivalent Conv3x3 forms; then, in Step2 to Step3, multiple two-dimensional convolution kernels of compatible sizes are operated on the same input with the same amplitude to produce outputs of the same resolution, and these outputs are summed to obtain an equivalent kernel, and then the convolution layers of the three branches are added to form the final Conv3x3 form; from this transformation process, it can be seen that the core of structural reparameterization is to determine a set of parameters in the training stage and use an equivalent set of parameters in the reasoning stage; as long as the parameters in the training stage can be equivalently converted into the parameters in the reasoning stage, the multi-way structure can be converted into a single-way structure, thereby realizing the structural conversion.
[0033] Specifically, in this implementation, the FasterNet-Block module is redesigned, combining the new design of multi-channel and single-channel network structures; this improvement is combined with the partial convolution module (Pconv), batch normalization (BN) and rectified linear unit (ReLU) to form a new FasterNet-Block block, and replaces the BasicBlock in the Backbone of RT-DETR; Through the above improvements, we hope to further improve the speed and accuracy of the RT-DETR model in small target detection, making it more suitable for practical application scenarios.
[0034] S300, Model Verification: Specifically, the aforementioned COCO dataset was used to verify the performance of the model from the perspectives of precision and recall; S400, model call: Specifically, the verified model is used for recognition applications of specific tasks.
[0035] It should be noted that the above examples are only for explaining the present invention and cannot limit the protection scope of the present invention. Example 2
[0036] This embodiment is based on the same inventive concept as the small target detection method in a high-altitude overlooking scene described in Example 1, and provides a small target detection system in a high-altitude overlooking scene, such as Figure 7 As shown, including: Dataset collection module, used to: Specifically, we use the COCO2017 dataset released by Microsoft to train and verify the model; Model modification and training modules for: Specifically, the RT-DETR model is modified and a perfect RT-DETR model is trained; Specifically, since the performance of the multi-branch network model is significantly improved, it can utilize the receptive fields of different convolution kernels; however, the multi-path structure leads to a sharp increase in the demand for video memory, which places high demands on the device; in order to balance the model performance and resource consumption, in this implementation, a model structure reorganization strategy is adopted for model training; during training, the multi-path structure is used to give full play to the parallel computing capability of the GPU; in the inference stage, multiple computing modules are fused into a single path model to increase the training speed; Among them, two methods are designed in the model structure reorganization design: one method is to merge the convolution layer and the subsequent batch normalization (BN) layer into a new operation, which can reduce the number of convolution operations in the model while keeping the input and output sizes unchanged, thereby simplifying the model structure and improving performance; the other method is to add the results of multiple parallel convolution branches of the same specification for the same input and then perform a convolution operation, so that multiple convolution kernels can be merged into one, and then a single convolution operation is performed on the input to obtain the same result as the branch operation; the same specifications here include the size, padding and stride of the convolution kernel, because the batch normalization layer is usually connected after the convolution operation; these methods effectively improve the performance of the detection model without affecting the accuracy; in summary, the single-path structure after structural reorganization is shown as follows Figure 3 ; Convert the multi-channel model into a single-channel model such as Figure 4As shown in the figure; in this process, the fusion of convolution and batch normalization (BN) layers is first used to merge the Conv layer with the subsequent BN layer, and the transition from Step1 to Step2 is realized by converting Conv1x1 and identity into equivalent Conv3x3 forms; then, in Step2 to Step3, multiple two-dimensional convolution kernels of compatible sizes are operated on the same input with the same amplitude to produce outputs of the same resolution, and these outputs are summed to obtain an equivalent kernel, and then the convolution layers of the three branches are added to form the final Conv3x3 form; from this transformation process, it can be seen that the core of structural reparameterization is to determine a set of parameters in the training stage and use an equivalent set of parameters in the reasoning stage; as long as the parameters in the training stage can be equivalently converted into the parameters in the reasoning stage, the multi-way structure can be converted into a single-way structure, thereby realizing the structural conversion.
[0037] Specifically, in this implementation, the FasterNet-Block module is redesigned, combining the new design of multi-channel and single-channel network structures; this improvement is combined with the partial convolution module (Pconv), batch normalization (BN) and rectified linear unit (ReLU) to form a new FasterNet-Block block, and replaces the BasicBlock in the Backbone of RT-DETR; Through the above improvements, we hope to further improve the speed and accuracy of the RT-DETR model in small target detection, making it more suitable for practical application scenarios.
[0038] Model validation module for: Specifically, the aforementioned COCO dataset was used to verify the performance of the model from the perspectives of precision and recall; Model calling module, used to: Specifically, the verified model is used for recognition applications of specific tasks.
[0039] It should be noted that the above examples are only for explaining the present invention and cannot limit the protection scope of the present invention. Example 3
[0040] This embodiment provides a computer-readable storage medium, including: The storage medium is used to store computer software instructions used to implement the small target detection method in the high-altitude overhead view scene described in the above-mentioned Example 1, which includes a program for executing the above-mentioned small target detection method in the high-altitude overhead view scene; specifically, the executable program can be built into the small target detection system in the high-altitude overhead view scene described in Example 2, so that the small target detection system in the high-altitude overhead view scene can implement the small target detection method in the high-altitude overhead view scene described in Example 1 by executing the built-in executable program.
[0041] In addition, the computer-readable storage medium of this embodiment may adopt any combination of one or more computer-readable storage media, wherein the computer-readable storage medium includes electrical, optical, electromagnetic, infrared or semiconductor systems, devices or components, or any combination thereof. Example 4
[0042] This embodiment provides an electronic device, such as Figure 8 As shown, the electronic device may include: a processor 1501 , a communication interface 1502 , a memory 1503 and a communication bus 1504 , wherein the processor 1501 , the communication interface 1502 , and the memory 1503 communicate with each other via the communication bus 1504 .
[0043] Memory 1503, used for storing computer programs; The processor 1501 is used to implement the steps of the small target detection method in the high-altitude bird's-eye view scene described in the above-mentioned embodiment 1 when executing the computer program stored in the memory 1503.
[0044] As an embodiment of the present invention, the communication bus mentioned in the above terminal can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus. The communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 8 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.
[0045] As an implementation mode of the present invention, the communication interface is used for communication between the above-mentioned terminal and other devices.
[0046] As an embodiment of the present invention, the memory may include a random access memory (RAM) or a non-volatile memory, such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.
[0047] As an embodiment of the present invention, the above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components.
[0048] Different from the prior art, the small target detection method, system, device and medium in a high-altitude bird's-eye view scenario of the present application can improve the small target detection performance of RT-DERT, enabling RT-DETR to achieve end-to-end target detection without the need for manually designed feature extractors, anchor frame generators and non-maximum suppression required by traditional target detection methods, thereby simplifying the design and training process of the target detection model.
[0049] It should be understood that in the various embodiments of this document, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this document.
[0050] It should also be understood that in the embodiments of this article, the term "and / or" is only a description of the association relationship of the associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.
[0051] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this article.
[0052] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.
[0053] In the several embodiments provided herein, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, or can be electrical, mechanical or other forms of connection.
[0054] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the embodiments of this article.
[0055] In addition, each functional unit in each embodiment of this invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above integrated unit may be implemented in the form of hardware or in the form of software functional unit.
[0056] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this article is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of this article. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.
[0057] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A small target detection method in a high-altitude overlooking scene, characterized in that: The following steps are involved: Dataset collection: Use COCO2017 as the dataset for model training and validation; Model modification and training: Modify and train the model based on the model structure reorganization strategy and Block module update strategy; Model Validation: Using the data set to verify the modified and trained model; Model call: Call the verified model to perform the target detection task.
2. The small target detection method in a high-altitude overlooking scene according to claim 1 is characterized in that: The model structure reorganization strategy includes: Utilize multi-way structure during training; In the inference phase, multiple computing modules are fused into a single path model.
3. The small target detection method in a high-altitude overlooking scene according to claim 2 is characterized in that: The model structure reorganization strategy further includes: The convolution and batch normalization layers are combined to merge the Conv layer with the subsequent BN layer. Transform Step 1 to Step 2 by converting Conv1x1 and identity to equivalent Conv3x3 form; From Step 2 to Step 3, multiple two-dimensional convolution kernels of compatible sizes are operated on the same input with the same step amplitude to produce outputs of the same resolution. The outputs are summed to obtain equivalent kernels.
4. The small target detection method in a high-altitude overlooking scene according to claim 1 is characterized in that: The Block module update strategy includes: The FasterNet-Block module is combined with some convolution modules, batch normalization and rectified linear units to form a new FasterNet-Block block.
5. The small target detection method in a high-altitude overlooking scene according to claim 4 is characterized in that: The model adopts the RT-DETR model.
6. The small target detection method in a high-altitude overlooking scene according to claim 5 is characterized in that: The Block module update strategy further includes: The new FasterNet-Block block is used to replace the BasicBlock in the Backbone of the RT-DETR model.
7. The small target detection method in a high-altitude overlooking scene according to claim 4 is characterized in that: The step of using the data set to verify the modified and trained model includes: The COCO2017 is used to verify the performance of the modified and trained RT-DETR model from the perspectives of precision and recall.
8. A small target detection system in a high-altitude overlooking scene, characterized in that: include: The dataset collection module is used to: use COCO2017 as a dataset for model training and verification; Model modification and training module, used to: modify and train the model based on the model structure reorganization strategy and Block module update strategy; A model verification module, used to: verify the modified and trained model using the data set; The model calling module is used to call the verified model to perform the target detection task.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the steps of the small target detection method in the high-altitude bird's-eye view scene described in any one of claims 1 to 7 are implemented.
10. A computer device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; wherein: The memory is used to store computer programs; The processor is used to execute the steps of the small target detection method in a high-altitude bird's-eye view scene described in any one of claims 1 to 7 by running the program stored in the memory.
Citation Information
Cited By
Bridge multi-mode multi-target disease intelligent identification method and device under complex background and medium
CN120071007A
