An image small target detection method based on improved YOLOV11 algorithm

Abstract

The application provides an image small target detection method based on an improved YOLOV11 algorithm, obtains a pretreated image by acquiring an image, and forms a training data set; an image small target detection model based on the improved YOLOV11 algorithm is constructed and includes a backbone network, a neck network and a head network, the backbone network integrates the input image to extract multi-scale features through parallel branches and a frequency domain self-attention mechanism, and convolution and self-attention are fused to enhance the features, three feature maps are output to the neck network, the neck network outputs four fusion features to the head network, and the third feature map output by the backbone network is directly output to the head network; the head network outputs a detection result; a trained YOLO-EX model is obtained; a target detection result is obtained; the method can realize high-precision and high-efficiency image target detection, the small target detection effect is remarkable through experimental verification, and the method has high industrial application value.

Description

Technical Field

[0001] This invention relates to a method for detecting small objects in images based on an improved YOLOv11 algorithm, belonging to the fields of computer vision and image processing technology. Background Technology

[0002] Image object detection is a key technology in computer vision, widely used in military reconnaissance, agricultural monitoring, and urban planning. However, images are characterized by high resolution, complex backgrounds, large target scale variations, and dense small targets, posing significant challenges to detection algorithms. Traditional object detection algorithms, such as Faster R-CNN, rely on manually designed rules, making it difficult to handle complex semantic information. While single-stage detectors based on deep learning, such as the YOLO series, offer real-time advantages, they still have shortcomings in multi-scale and small target detection. For example, YOLOv11 improves flexibility through the C3K2 module and PSA attention mechanism, but its feature extraction capabilities are limited in images, easily missing small targets or falsely detecting background noise. Existing improved models, such as YOLO-HR, enhance multi-scale detection through multiple detector heads, but suffer from low parameter efficiency and reliance on pre-trained weights.

[0003] Furthermore, loss functions such as CIoU overemphasize geometric properties, leading to training instability when there are large differences in target scale. Therefore, there is an urgent need for a lightweight model that can efficiently handle multi-scale and small target detection in images. Summary of the Invention

[0004] The purpose of this invention is to provide an image small target detection method based on the improved YOLOv11 algorithm to solve the problems of low accuracy and low efficiency in the existing small target detection technology.

[0005] The technical solution of this invention is: A method for detecting small targets in images based on an improved YOLOv11 algorithm includes the following steps: Step 1: Acquire images, perform preprocessing, and obtain preprocessed images to form the training dataset; Step 2: Construct an image small target detection model based on the improved YOLOV11 algorithm, namely the YOLO-EX model. The YOLO-EX model includes a backbone network, a neck network, and a head network. The backbone network integrates a parallel feature extraction module to extract multi-scale features from the input image through parallel branching and frequency domain self-attention mechanism, and performs feature enhancement by fusing convolution and self-attention, outputting three feature maps to the neck network. The neck network fuses the three input feature maps and outputs four fused features to the head network. At the same time, the third feature map output by the backbone network is directly output to the head network, and the head network outputs the detection result.

[0006] Step 3: Train the YOLO-EX model using the training dataset to obtain the trained YOLO-EX model; Step 4: Use the trained YOLO-EX model to obtain the target detection results for the image to be detected.

[0007] Furthermore, in the backbone network, the input image passes through the following layers in sequence: first convolutional layer Conv1, second convolutional layer Conv2, first feature extraction module PFS(F)1, third convolutional layer Conv3, second feature extraction module PFS(F)2, fourth convolutional layer Conv4, first deep feature extraction module PFS(T)1, fifth convolutional layer Conv5, second deep feature extraction module PFS(T)2, spatial pyramid pooling and convolutional self-attention fusion module SPPFC, and C2PSA module. Among them, the second feature extraction module PFS(F)2 extracts multi-scale features through parallel branching and frequency domain self-attention mechanism and outputs the first feature map to the neck network. The first deep feature extraction module PFS(T)1 outputs the second feature map to the neck network, and the C2PSA module outputs the third feature map to the neck network.

[0008] Furthermore, the first feature extraction module PFS(F)1 and the second feature extraction module PFS(F)2 adopt the same structure. The second feature extraction module PFS(F)2 includes a sixth convolutional layer Conv6, a first segmentation layer Split1, a first bottleneck module FBottleneck1 based on the frequency domain main attention mechanism, a second bottleneck module FBottleneck2 based on the frequency domain main attention mechanism, a first concatenation layer Concat1, a seventh convolutional layer Conv7, and a first parallel branch module ParnetBlock1. The sixth convolutional layer Conv6 convolves the input feature map and outputs the convolved features. Figure 1 The first segmentation layer Split1 and the first concatenation layer Concat1 are used to segment the convolutional feature map and output the first feature map. The first feature map passes through the first bottleneck module FBottleneck1 based on the frequency domain main attention mechanism and outputs the second feature map. The second feature map passes through the second bottleneck module FBottleneck2 based on the frequency domain main attention mechanism and outputs the third feature map, which is then fed to the first concatenation layer Concat1. The first concatenation layer Concat1 concatenates the feature maps and then passes them through the seventh convolutional layer Conv7 and the first parallel branch module ParnetBlock1 to output the first feature map.

[0009] Furthermore, the first bottleneck module FBottleneck1 based on the frequency domain main attention mechanism and the second bottleneck module FBottleneck2 based on the frequency domain main attention mechanism adopt the same structure. The first bottleneck module FBottleneck1 based on the frequency domain main attention mechanism includes an eighth convolutional layer Conv8 and a frequency domain self-attention solver FSAS. After the first feature map passes through the eighth convolutional layer Conv8 and the frequency domain self-attention solver FSAS, intermediate features are output. The intermediate features are concatenated with the first feature map to output the second feature map.

[0010] Furthermore, the first deep feature extraction module PFS(T)1 and the second deep feature extraction module PFS(T)2 adopt the same structure. The first deep feature extraction module PFS(T)1 includes a ninth convolutional layer Conv9, a second segmentation layer Split2, a first residual feature extraction module C3k_F1, a second residual feature extraction module C3k_F12, a second concatenation layer Concat2, a tenth convolutional layer Conv10, and a second parallel branch module ParnetBlock2. The ninth convolutional layer Conv9 convolves the input feature map and outputs the convolved feature map. Figure 2 The second segmentation layer Split2 and the second concatenation layer Concat2 are given. After the second segmentation layer Split2 performs segmentation, it outputs a fourth feature map, which is then given to the first residual feature extraction module C3k_F1 and the second concatenation layer Concat2. The first residual feature extraction module C3k_F1 outputs a fifth feature map, which is then given to the second residual feature extraction module C3k_F12. The second residual feature extraction module C3k_F12 outputs a sixth feature map, which is then given to the second concatenation layer Concat2. The concatenated features output by the second concatenation layer Concat2 are then passed through the tenth convolutional layer Conv10 and the second parallel branch module ParnetBlock2 to output a second feature map.

[0011] Furthermore, the first residual feature extraction module C3k_F1 and the second residual feature extraction module C3k_F12 have the same structure. The first residual feature extraction module C3k_F1 includes an eleventh convolutional layer Conv11, a third FBottleneck module, a fourth FBottleneck module, a third concatenation layer Concat3, a twelfth convolutional layer Conv12, and a thirteenth convolutional layer Conv13. The fourth feature map is output as a second intermediate feature after passing through the eleventh convolutional layer Conv11. The second intermediate feature is then output as a third intermediate feature after passing through the third FBottleneck module and the fourth FBottleneck module, and then passed to the third concatenation layer Concat3. The second intermediate feature is then output as a fourth intermediate feature after passing through the twelfth convolutional layer Conv12, and then passed to the third concatenation layer Concat3. The output feature of the third concatenation layer Concat3 is then output as a fifth feature map after passing through the thirteenth convolutional layer Conv13.

[0012] Furthermore, the Spatial Pyramid Pooling and Convolutional Self-Attention Fusion Module (SPPFC) includes a first switchable dilated convolutional layer (SAConv1), a second 2D max pooling layer (MaxPool2d1), a third 2D max pooling layer (MaxPool2d2), a fourth concatenation layer (Concat4), a CAFM module, and a second switchable dilated convolutional layer (SAConv2). The first switchable dilated convolutional layer (SAConv1) performs switchable dilated convolution on the input feature map and outputs the dilated convolutional features to the second 2D max pooling layer (MaxPool2d1) and the fourth concatenation layer (Concat4). The second 2D max pooling layer (MaxPool2d1) outputs... The features output after the first pooling are fed into the second two-dimensional max pooling layer MaxPool2d2 and the fourth concatenation layer Concat4. The features output after the second pooling from the second two-dimensional max pooling layer MaxPool2d2 are fed into the third two-dimensional max pooling layer MaxPool2d3 and the fourth concatenation layer Concat4. The features output after the third pooling from the third two-dimensional max pooling layer MaxPool2d3 are fed into the fourth concatenation layer Concat4. The features output from the fourth concatenation layer Concat4 are fed into the second switchable dilated convolutional layer SAConv2 through the CAFM module. The features output from the second switchable dilated convolutional layer SAConv2 are used as the output of the spatial pyramid pooling and convolutional self-attention fusion module SPPFC.

[0013] Furthermore, the head network employs five detection heads to process targets of different scales, outputting prediction results for targets of different scales.

[0014] Furthermore, in step S3, the YOLO-EX model is trained using the loss function Inner-WIoU to obtain the trained YOLO-EX model.

[0015] The beneficial effects of this invention are as follows: This image small target detection method based on the improved YOLOv11 algorithm, by constructing a YOLO-EX model and combining a parallel feature extraction module and a spatial pyramid pooling and convolutional self-attention fusion module (SPPFC), effectively handles complex backgrounds and noise, achieving high-precision and high-efficiency image target detection. The Inner-WIoU loss function is used to improve model training stability and exhibits strong robustness. Experimental results show that the method achieves an AP50 of 84.4% on the SIMD dataset, a 3.0% improvement over YOLOv11n; and an mAP of 65.6% on the DIOR dataset, a 10.7% improvement, demonstrating significant performance in small target detection. Furthermore, the YOLO-EX model in this invention achieves only 5.4M parameters, 14.2G FLOPs, and an inference speed of 6.0ms, outperforming models such as YOLO-HR-n, and possessing high industrial application value. Attached Figure Description

[0016] Figure 1 This is a schematic flowchart of an image small target detection method based on an improved YOLOv11 algorithm according to an embodiment of the present invention; Figure 2 A schematic diagram illustrating the YOLO-EX model in the embodiments; Figure 3 A schematic diagram illustrating the second feature extraction module PFS(F)2 in the embodiment; Figure 4 A schematic diagram illustrating the first bottleneck module FBottleneck1 based on the frequency domain main attention mechanism in the embodiment; Figure 5 A schematic diagram illustrating the first PFS(T) module in the embodiment; Figure 6 A schematic diagram illustrating the first C3k_F layer in the embodiment; Figure 7 A schematic diagram illustrating the Spatial Pyramid Pooling and Convolutional Self-Attention Fusion Module (SPPFC) in the embodiment; Figure 8 This is a schematic diagram illustrating the detection visualization results of the YOLO-EX model on the SIMD dataset in this embodiment; Figure 9 This is a schematic diagram comparing the target detection results of the YOLO-EX model and the YOLOv11n model in the embodiment. Detailed Implementation

[0017] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] This embodiment provides a method for small object detection in images based on an improved YOLOv11 algorithm, such as... Figure 1 This includes the following steps: Step 1: Acquire images, perform preprocessing, and obtain preprocessed images to form the training dataset.

[0019] Step 2: Construct an image small target detection model based on the improved YOLOV11 algorithm, namely the YOLO-EX model. The YOLO-EX model includes a backbone network, a neck network, and a head network. The backbone network integrates a parallel feature extraction module to extract multi-scale features from the input image through parallel branching and frequency domain self-attention mechanism, and performs feature enhancement by fusing convolution and self-attention, outputting three feature maps to the neck network. The neck network fuses the three input feature maps and outputs four fused features to the head network. At the same time, the third feature map output by the backbone network is directly output to the head network, and the head network outputs the detection result.

[0020] like Figure 2In the backbone network, the input image passes through the following layers in sequence: first convolutional layer Conv1, second convolutional layer Conv2, first feature extraction module PFS(F)1, third convolutional layer Conv3, second feature extraction module PFS(F)2, fourth convolutional layer Conv4, first deep feature extraction module PFS(T)1, fifth convolutional layer Conv5, second deep feature extraction module PFS(T)2, spatial pyramid pooling and convolutional self-attention fusion module SPPFC, and C2PSA module. Among them, the second feature extraction module PFS(F)2 extracts multi-scale features through parallel branching and frequency domain self-attention mechanism and outputs the first feature map to the neck network. The first deep feature extraction module PFS(T)1 outputs the second feature map to the neck network. The C2PSA module outputs the third feature map to the neck network.

[0021] like Figure 3 The first feature extraction module PFS(F)1 and the second feature extraction module PFS(F)2 adopt the same structure. The second feature extraction module PFS(F)2 includes a sixth convolutional layer Conv6, a first segmentation layer Split1, a first bottleneck module FBottleneck1 based on the frequency domain main attention mechanism, a second bottleneck module FBottleneck2 based on the frequency domain main attention mechanism, a first concatenation layer Concat1, a seventh convolutional layer Conv7, and a first parallel branch module ParnetBlock1. The sixth convolutional layer Conv6 convolves the input feature map and outputs the convolved features. Figure 1 The first segmentation layer Split1 and the first concatenation layer Concat1 are used to segment the convolutional feature map and output the first feature map. The first feature map passes through the first bottleneck module FBottleneck1 based on the frequency domain main attention mechanism and outputs the second feature map. The second feature map passes through the second bottleneck module FBottleneck2 based on the frequency domain main attention mechanism and outputs the third feature map, which is then fed to the first concatenation layer Concat1. The first concatenation layer Concat1 concatenates the feature maps and then passes them through the seventh convolutional layer Conv7 and the first parallel branch module ParnetBlock1 to output the first feature map.

[0022] like Figure 4 The first bottleneck module FBottleneck1 based on the frequency domain main attention mechanism and the second bottleneck module FBottleneck2 based on the frequency domain main attention mechanism adopt the same structure. The first bottleneck module FBottleneck1 based on the frequency domain main attention mechanism includes an eighth convolutional layer Conv8 and a frequency domain self-attention solver FSAS. After the first feature map passes through the eighth convolutional layer Conv8 and the frequency domain self-attention solver FSAS, intermediate features are output. The intermediate features are concatenated with the first feature map to output the second feature map.

[0023] like Figure 5 The first deep feature extraction module PFS(T)1 and the second deep feature extraction module PFS(T)2 adopt the same structure. The first deep feature extraction module PFS(T)1 includes a ninth convolutional layer Conv9, a second splitting layer Split2, a first residual feature extraction module C3k_F1, a second residual feature extraction module C3k_F12, a second concatenation layer Concat2, a tenth convolutional layer Conv10, and a second parallel branch module ParnetBlock2. The ninth convolutional layer Conv9 convolves the input feature map and outputs the convolved feature map. Figure 2 The second segmentation layer Split2 and the second concatenation layer Concat2 are given. After the second segmentation layer Split2 performs segmentation, it outputs a fourth feature map, which is then given to the first residual feature extraction module C3k_F1 and the second concatenation layer Concat2. The first residual feature extraction module C3k_F1 outputs a fifth feature map, which is then given to the second residual feature extraction module C3k_F12. The second residual feature extraction module C3k_F12 outputs a sixth feature map, which is then given to the second concatenation layer Concat2. The concatenated features output by the second concatenation layer Concat2 are then passed through the tenth convolutional layer Conv10 and the second parallel branch module ParnetBlock2 to output a second feature map.

[0024] like Figure 6 The first residual feature extraction module C3k_F1 and the second residual feature extraction module C3k_F12 have the same structure. The first residual feature extraction module C3k_F1 includes an eleventh convolutional layer Conv11, a third FBottleneck module, a fourth FBottleneck module, a third concatenation layer Concat3, a twelfth convolutional layer Conv12, and a thirteenth convolutional layer Conv13. The fourth feature map is output as a second intermediate feature after passing through the eleventh convolutional layer Conv11. The second intermediate feature is then output as a third intermediate feature after passing through the third FBottleneck module and the fourth FBottleneck module. The second intermediate feature is then output as a fourth intermediate feature after passing through the twelfth convolutional layer Conv12. The output feature of the third concatenation layer Concat3 is then output as a fifth feature map after passing through the thirteenth convolutional layer Conv13.

[0025] like Figure 7The Spatial Pyramid Pooling and Convolutional Self-Attention Fusion Module (SPPFC) includes a first switchable dilated convolutional layer (SAConv1), a second 2D max pooling layer (MaxPool2d1), a third 2D max pooling layer (MaxPool2d2), a fourth concatenation layer (Concat4), a CAFM module, and a second switchable dilated convolutional layer (SAConv2). The first switchable dilated convolutional layer (SAConv1) performs switchable dilation convolution on the input feature map and outputs the dilated convolutional features to the second 2D max pooling layer (MaxPool2d1) and the fourth concatenation layer (Concat4). The second 2D max pooling layer (MaxPool2d1) outputs the dilated features... The features after the first pooling are fed into the second two-dimensional max pooling layer MaxPool2d2 and the fourth concatenation layer Concat4. The second two-dimensional max pooling layer MaxPool2d2 outputs the features after the second pooling and feeds them into the third two-dimensional max pooling layer MaxPool2d3 and the fourth concatenation layer Concat4. The third two-dimensional max pooling layer MaxPool2d3 outputs the features after the third pooling and feeds them into the fourth concatenation layer Concat4. The output features of the fourth concatenation layer Concat4 are fed into the second switchable dilated convolutional layer SAConv2 through the CAFM module. The output features of the second switchable dilated convolutional layer SAConv2 are used as the output of the spatial pyramid pooling and convolutional self-attention fusion module SPPFC.

[0026] In step 2, high-precision and high-efficiency detection is achieved through modular innovation and loss function optimization. Parallel feature extraction is based on ParNet blocks and the frequency domain self-attention solver FSAS. The frequency domain self-attention solver FSAS contains three parallel branches: 1x1 convolution, 3x3 convolution, and SSE branch (Skip-Squeeze-and-Excitation). By reducing attention complexity through frequency domain computation, it enhances multi-scale feature extraction capabilities without significantly increasing network depth or parameters. The parallel feature extraction module PFS improves the model's robustness in noisy environments by fusing local details and global context through parallel branches. The Spatial Pyramid Pooling and Convolutional Self-Attention Fusion Module (SPPFC) builds upon the SPPF module in YOLOv11 by introducing the Convolutional Self-Attention Fusion Module (CAFM) and the Switchable Dilated Convolution Module (SAConv). CAFM combines the local feature extraction capabilities of convolution with the global dependency capture capabilities of self-attention. SAConv expands the receptive field by dynamically adjusting the dilation rate, enhancing feature discriminativity and reducing spatial information loss during pooling operations, thus optimizing small target detection. The head network employs five detection heads to process targets at different scales, outputting prediction results for targets of varying scales. Features are aggregated via a bidirectional path through the PAN network to improve multi-scale detection performance.

[0027] Step 3: Train the YOLO-EX model using the training dataset to obtain the trained YOLO-EX model.

[0028] In step S3, the YOLO-EX model is trained using the Inner-WIoU loss function. The Inner-WIoU loss function dynamically adjusts the IoU weights to optimize training, resulting in the trained YOLO-EX model. The Inner-WIoU loss function introduces a scaling factor Ratio=1.2 to dynamically adjust the IoU weights and employs a gradient enhancement strategy to reduce competition from high-quality anchor boxes and the negative gradient impact of low-quality samples, thus improving the model's adaptability to scale changes. Weighting of the inner overlapping region enhances the attention given to ordinary anchor boxes.

[0029] Step 4: Use the trained YOLO-EX model to obtain the target detection results for the image to be detected.

[0030] In step 4, the trained YOLO-EX model can be deployed on drones or satellite platforms to achieve real-time target detection, and the output includes bounding boxes and class probabilities.

[0031] This image small target detection method based on the improved YOLOv11 algorithm constructs a YOLO-EX model, combining a parallel feature extraction module and a spatial pyramid pooling and convolutional self-attention fusion module (SPPFC). This effectively handles complex backgrounds and noise, achieving high-precision and high-efficiency image target detection. The Inner-WIoU loss function enhances model training stability and exhibits strong robustness. Experimental results show that the proposed method achieves an AP50 of 84.4% on the SIMD dataset, a 3.0% improvement over YOLOv11n; and an mAP of 65.6% on the DIOR dataset, a 10.7% improvement, demonstrating significant performance in small target detection. Figure 8 and Figure 9 .Depend on Figure 8 It can be seen that the proposed method can accurately locate multi-scale targets simultaneously in complex backgrounds, maintaining good detection performance even in densely distributed, partially occluded, and noisy / shadow-affected areas; especially for small or elongated targets, it can still output clear and stable prediction boxes, further demonstrating the robustness and practical value of the proposed method in handling multi-scale targets and suppressing image noise. Figure 9It can be seen that when multiple scale targets exist simultaneously in an image, the YOLO-EX model of the embodiment can handle targets of different sizes more effectively. In scenarios with a large number of targets, the YOLO-EX model of the embodiment has a stronger ability to focus on small targets and can reduce false detections. Overall comparison results show that in complex remote sensing scenarios with multiple targets and multiple scales, the YOLO-EX model of the embodiment has stronger target perception and discrimination capabilities than the baseline YOLOv11n. The comparative experimental results of the YOLO-EX model of the embodiment and existing models on SIMD are shown in Table 1.

[0032] Table 1. Comparative experimental results of the YOLO-EX model in this embodiment and existing models on SIMD.

[0033] As shown in Table 1, the YOLO-EX model in this embodiment achieves only 5.4M parameters, 14.2G FLOPs, and an inference speed of 6.0ms, which is superior to models such as YOLO-HR-n and has high industrial application value. The proposed image small target detection method based on the improved YOLOv11 algorithm is experimentally verified using industrial remote sensing detection as an example, as follows: Experimental environment configuration: Ubuntu 24.04 system, PyTorch 2.4 framework, CPU Intel Xeon E5-2650, GPU NVIDIA RTX 2080Ti.

[0034] Data preparation: Load the SIMD or DIOR dataset, adjust the image size to 1024x1024 and 800x800 respectively, set the batch size to 8, the epochs to 200, and do not use pre-trained weights.

[0035] Model Training: The comparison model uses the YOLOv11n model. The YOLO-EX model in this example integrates a parallel feature extraction module (parameters increased by 1.3M), a spatial pyramid pooling and convolutional self-attention fusion module (SPPFC) (parameters increased by 1.5M), and the loss function is Inner-WIoU. Training hyperparameters: learning rate 0.01, momentum 0.937, weight decay 0.0005, warm-up phase 3 times. Iterative training until convergence, output detection results are shown in Table 2: Table 2. Comparison of YOLO-EX model of the embodiment with existing models on DIOR.

[0036] As shown in Table 2, on the SIMD dataset, the AP50 of the YOLO-EX model in the example is 84.4%, and the mAP is 68.0%; on the DIOR dataset, the AP50 is 88.1%, and the mAP is 65.6%.

[0037] The results of the ablation experiment of this image small target detection method based on the improved YOLOv11 algorithm in the embodiment are shown in Table 3: Table 3 Ablation Experiment Results of the Example Method

[0038] As can be seen from Table 3, the methods in the embodiments show that each module makes a significant contribution. For example, the parallel feature extraction module PFS alone improves AP50 by 2.2%.

[0039] This image small target detection method based on the improved YOLOv11 algorithm addresses challenges such as high resolution, complex backgrounds, multi-scale targets, and dense distribution of small targets in images. It significantly improves multi-scale and small target detection performance by introducing a designed parallel feature extraction module (PFS), a spatial pyramid pooling and convolutional self-attention fusion module (SPPFC), an Inner-WIoU loss function, and an additional detection head. The PFS module expands the receptive field without increasing network depth through a parallel branching structure, reducing computational complexity. The SPPFC module integrates convolution and self-attention mechanisms, enhancing feature extraction capabilities for complex scenes. The Inner-WIoU loss function dynamically adjusts IoU weights to reduce the impact of scale differences on detection accuracy. Experiments on the SIMD and DIOR datasets show that the YOLO-EX model in this embodiment outperforms mainstream models such as YOLOv11n and YOLO-HR-n in both AP50 and mAP metrics, and exhibits smaller parameter increments and higher real-time performance. This invention provides an efficient and robust detection tool for image analysis. Its modular design can be extended to other detection tasks and is suitable for applications such as inspection and testing, industrial quality inspection, environmental monitoring, medical imaging inspection, or agricultural monitoring.

[0040] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A method for detecting small targets in images based on an improved YOLOv11 algorithm, characterized in that, Includes the following steps: Step 1: Acquire images, perform preprocessing, and obtain preprocessed images to form the training dataset; Step 2: Construct an image small target detection model based on the improved YOLOV11 algorithm, namely the YOLO-EX model. The YOLO-EX model includes a backbone network, a neck network, and a head network. The backbone network extracts multi-scale features from the input image through parallel branching and frequency domain self-attention mechanism, and performs feature enhancement by fusing convolution and self-attention. It outputs three feature maps to the neck network. The neck network fuses the three input feature maps and outputs four fused features to the head network. At the same time, it directly outputs the third feature map from the backbone network to the head network. The head network outputs the detection results. Step 3: Train the YOLO-EX model using the training dataset to obtain the trained YOLO-EX model; Step 4: Use the trained YOLO-EX model to obtain the target detection results for the image to be detected.

2. The image small target detection method based on the improved YOLOv11 algorithm as described in claim 1, characterized in that: In the backbone network, the input image passes through the following layers in sequence: first convolutional layer Conv1, second convolutional layer Conv2, first feature extraction module PFS(F)1, third convolutional layer Conv3, second feature extraction module PFS(F)2, fourth convolutional layer Conv4, first deep feature extraction module PFS(T)1, fifth convolutional layer Conv5, second deep feature extraction module PFS(T)2, spatial pyramid pooling and convolutional self-attention fusion module SPPFC, and C2PSA module. Among them, the second feature extraction module PFS(F)2 extracts multi-scale features through parallel branching and frequency domain self-attention mechanism and outputs the first feature map to the neck network. The first deep feature extraction module PFS(T)1 outputs the second feature map to the neck network. The C2PSA module outputs the third feature map to the neck network.

3. The image small target detection method based on the improved YOLOv11 algorithm as described in claim 2, characterized in that: The first feature extraction module PFS(F)1 and the second feature extraction module PFS(F)2 adopt the same structure. The second feature extraction module PFS(F)2 includes a sixth convolutional layer Conv6, a first segmentation layer Split1, a first bottleneck module FBottleneck1 based on the frequency domain main attention mechanism, a second bottleneck module FBottleneck2 based on the frequency domain main attention mechanism, a first concatenation layer Concat1, a seventh convolutional layer Conv7, and a first parallel branch module ParnetBlock1. The sixth convolutional layer Conv6 convolves the input feature map and outputs the convolved feature map to the first concatenation layer. The system consists of a splitting layer (Split1) and a first concatenation layer (Concat1). The first splitting layer (Split1) segments the convolutional feature map and outputs a first feature map. The first feature map then passes through a first bottleneck module (FBottleneck1) based on a frequency domain main attention mechanism and outputs a second feature map. The second feature map then passes through a second bottleneck module (FBottleneck2) based on a frequency domain main attention mechanism and outputs a third feature map, which is then fed into the first concatenation layer (Concat1). The first concatenation layer (Concat1) concatenates the features and then passes through a seventh convolutional layer (Conv7) and a first parallel branch module (ParnetBlock1) to output the first feature map.

4. The image small target detection method based on the improved YOLOv11 algorithm as described in claim 3, characterized in that: The first bottleneck module FBottleneck1 based on the frequency domain main attention mechanism and the second bottleneck module FBottleneck2 based on the frequency domain main attention mechanism adopt the same structure. The first bottleneck module FBottleneck1 based on the frequency domain main attention mechanism includes an eighth convolutional layer Conv8 and a frequency domain self-attention solver FSAS. After the first feature map passes through the eighth convolutional layer Conv8 and the frequency domain self-attention solver FSAS, intermediate features are output. The intermediate features are concatenated with the first feature map to output the second feature map.

5. The image small target detection method based on the improved YOLOv11 algorithm as described in claim 2, characterized in that: The first deep feature extraction module PFS(T)1 and the second deep feature extraction module PFS(T)2 adopt the same structure. The first deep feature extraction module PFS(T)1 includes a ninth convolutional layer Conv9, a second segmentation layer Split2, a first residual feature extraction module C3k_F1, a second residual feature extraction module C3k_F12, a second concatenation layer Concat2, a tenth convolutional layer Conv10, and a second parallel branch module ParnetBlock2. The ninth convolutional layer Conv9 convolves the input feature map and outputs the convolved feature map. The second segmentation layer Split2, the first concatenation layer Concat2, the second concatenation layer Concat2, the third concatenation layer Conv10, the tenth convolutional layer Conv10, and the second parallel branch module ParnetBlock2. The second concatenation layer Concat2 and the second segmentation layer Split2 output a fourth feature map after segmentation, which is then fed to the first residual feature extraction module C3k_F1. The second concatenation layer Concat2 outputs a fifth feature map, which is then fed to the second residual feature extraction module C3k_F12. The second residual feature extraction module C3k_F12 outputs a sixth feature map, which is then fed to the second concatenation layer Concat2. The concatenated features output by the second concatenation layer Concat2 are then processed by the tenth convolutional layer Conv10 and the second parallel branch module ParnetBlock2 to output a second feature map.

6. The image small target detection method based on the improved YOLOv11 algorithm as described in claim 5, characterized in that: The first residual feature extraction module C3k_F1 and the second residual feature extraction module C3k_F12 have the same structure. The first residual feature extraction module C3k_F1 includes an eleventh convolutional layer Conv11, a third FBottleneck module, a fourth FBottleneck module, a third concatenation layer Concat3, a twelfth convolutional layer Conv12, and a thirteenth convolutional layer Conv13. The fourth feature map is output as a second intermediate feature after passing through the eleventh convolutional layer Conv11. The second intermediate feature is then passed through the third FBottleneck module and the fourth FBottleneck module in sequence, and output as a third intermediate feature to the third concatenation layer Concat3. The second intermediate feature is then passed through the twelfth convolutional layer Conv12, and output as a fourth intermediate feature to the third concatenation layer Concat3. The output feature of the third concatenation layer Concat3 is then passed through the thirteenth convolutional layer Conv13 to output the fifth feature map.

7. A method for detecting small targets in images based on an improved YOLOv11 algorithm as described in any one of claims 2-6, characterized in that: The Spatial Pyramid Pooling and Convolutional Self-Attention Fusion Module (SPPFC) includes a first switchable dilated convolutional layer (SAConv1), a second 2D max pooling layer (MaxPool2d1), a third 2D max pooling layer (MaxPool2d2), a fourth concatenation layer (Concat4), a CAFM module, and a second switchable dilated convolutional layer (SAConv2). The first switchable dilated convolutional layer (SAConv1) performs switchable dilated convolution on the input feature map and outputs the dilated convolutional features to the second 2D max pooling layer (MaxPool2d1) and the fourth concatenation layer (Concat4). The second 2D max pooling layer (MaxPool2d1) outputs the dilated features... The features after the first pooling are fed into the second two-dimensional max pooling layer MaxPool2d2 and the fourth concatenation layer Concat4. The second two-dimensional max pooling layer MaxPool2d2 outputs the features after the second pooling and feeds them into the third two-dimensional max pooling layer MaxPool2d3 and the fourth concatenation layer Concat4. The third two-dimensional max pooling layer MaxPool2d3 outputs the features after the third pooling and feeds them into the fourth concatenation layer Concat4. The output features of the fourth concatenation layer Concat4 are fed into the second switchable dilated convolutional layer SAConv2 through the CAFM module. The output features of the second switchable dilated convolutional layer SAConv2 are used as the output of the spatial pyramid pooling and convolutional self-attention fusion module SPPFC.

8. A method for detecting small targets in images based on an improved YOLOv11 algorithm as described in any one of claims 2-6, characterized in that: The head network uses five detectors to process targets of different scales and outputs prediction results for targets of different scales.

9. A method for detecting small targets in images based on an improved YOLOv11 algorithm as described in any one of claims 2-6, characterized in that: In step S3, the YOLO-EX model is trained using the Inner-WIoU loss function to obtain the trained YOLO-EX model.

Images