Small target detection method for complex background remote sensing image
By introducing Res2Unit network, dense connection network and fourth scale into the YOLO-V5 algorithm, the problem of small object detection accuracy of remote sensing images in complex backgrounds is solved, and higher detection accuracy and accuracy are achieved.
Patent Information
- Application Number
- CN202411611158.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-12
- Publication Date
- 2025-02-18
AI Technical Summary
复杂背景下遥感图像的小目标检测存在检测精度低、检测误差大、漏检率高等问题，限制了遥感图像的进一步应用。
Using a small object detection method for complex background remote sensing images based on YOLO-V5 algorithm, multi-scale feature extraction performance and detection accuracy are improved by building a Res2Unit network, establishing a dense connection network and the fourth scale.
It improves the detection accuracy and accuracy of small remote sensing targets in complex backgrounds, enhances the comprehensive performance of the algorithm, and effectively solves the problems of large detection errors and high missed detection rates.
Smart Images
Figure CN119478719A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of remote sensing small target detection algorithms in complex environments, and in particular relates to a small target detection method for remote sensing images with complex backgrounds. Background Art
[0002] High-altitude remote sensing image data is mainly obtained from satellite or aircraft imaging systems. It has been widely used in weapon guidance, emergency rescue, smart agriculture and other fields. It has become an important information acquisition tool and has gradually become the focus of research. However, when remote sensing satellites take images at high altitudes, they are affected by many factors, which may cause various geometric deformations, distortions, blurs and noises in the images. These factors include atmospheric conditions, changes in lighting and environmental interference, which make it difficult to accurately detect remote sensing targets under complex backgrounds. At present, the field of remote sensing image target detection still faces many challenges such as low detection accuracy, large detection errors and high missed detection rates. The existence of these problems limits the further application of remote sensing images. Summary of the invention
[0003] In view of this, the present invention provides a small target detection method for remote sensing images with complex backgrounds to solve the above-mentioned problems described in the background technology.
[0004] In order to achieve the above object, the present invention is achieved through the following technical solutions:
[0005] The present invention provides a small target detection method for a remote sensing image with a complex background, comprising the following steps:
[0006] Step 1: Build the basic YOLO-V5 algorithm network structure;
[0007] Step 2: Construct a Res2Unit network based on the residual network idea;
[0008] Step 3: Establish a densely connected network and the fourth scale to effectively and accurately detect small remote sensing targets under complex backgrounds;
[0009] Step 4: Use a variety of remote sensing small target image datasets as training sets and test sets to train and test the improved YOLO-V5 target detection algorithm.
[0010] In the above-mentioned small target detection method for remote sensing images with complex backgrounds, as a preferred solution, the process of building the basic YOLO-V5 algorithm network structure includes:
[0011] S11, transform the target detection problem into a regression problem and divide the image into a fixed number of grids;
[0012] S12. Use the YOLO-V5 algorithm to predict N bounding boxes of the network unit, and predict the position (x, y, w, h) of the bounding box and its confidence p;
[0013] The network output of the S13 and YOLO-V5 algorithms contains a total of S×S grid cells, and the output tensor size is S×S×(N×5+C).
[0014] In the above-mentioned small target detection method for remote sensing images with complex backgrounds, as a preferred solution, the position (x, y, w, h) of the bounding box and its confidence p, (x, y) represents the center of the box relative to the boundary of the grid unit, and w and h represent the width and height of the bounding box relative to the entire image; the confidence p is the probability that an object exists in each bounding box, defined as
[0015]
[0016] Among them, IoU is the intersection of the predicted box and the true value, which is a number between 0 and 1.
[0017] In the above-mentioned small target detection method for remote sensing images with complex backgrounds, as a preferred solution, the construction of a Res2Unit network based on the residual network idea specifically does not include:
[0018] S21, construct a hierarchical residual class connection in a residual network to complete feature extraction, and name it Res2Unit;
[0019] S22. Several small residual terms are introduced into the residual unit to achieve the purpose of expanding the receptive field of each layer; S23. Construct the output formula: Where * represents convolution; Conv represents the convolution layer of 3×3 convolution kernel; x1, x2, x3, x4 are the input feature maps that are evenly divided into 4 sub-features after passing through the convolution layer (1×1).
[0020] In the above-mentioned small target detection method for remote sensing images with complex backgrounds, as a preferred solution, the step 3 of establishing a densely connected network and a fourth scale specifically includes:
[0021] S31. Build a basic DenseNet model;
[0022] S32, extract feature maps of different scales through different Dense Blocks;
[0023] S33, extract a high-resolution feature map to introduce the fourth scale feature to capture finer-grained information;
[0024] S34, fuse feature maps of different scales.
[0025] In the above-mentioned small target detection method for remote sensing images with complex backgrounds, as a preferred solution, the Dense Block in S32 uses a Conv(1x1xM)-BN-ReLU-Conv(3x3xM)-BN-ReLU structure, in which the 1x1 convolution is used to reduce the number of input feature maps and reduce the amount of calculation.
[0026] In the above-mentioned small target detection method for remote sensing images with complex backgrounds, as a preferred solution, the fourth scale in step S33 is a feature map that is downsampled four times.
[0027] The present invention provides a small target detection method and application in remote sensing images with complex backgrounds, which has the following beneficial effects:
[0028] 1. In order to improve the multi-scale feature extraction performance of the target detection algorithm, the present invention proposes "Res2Unit" to replace the deep residual unit in the original feature extraction network of YOLO-V5 based on the idea of residual network. Since a hierarchical residual class connection is constructed in a residual network to complete feature extraction, multi-scale features can be represented with finer granularity and the receptive field range of each layer is expanded;
[0029] 2. To avoid gradient attenuation of the algorithm network, the present invention uses "Dense Block" in the densely connected network to replace the convolutional layer in the detection layer. Dense Block contains multiple layers, and the feature map size and dimension of each layer are the same. The layers are densely connected, which can more effectively utilize feature information and improve the overall performance of the algorithm network;
[0030] 3. In order to improve the comprehensive detection performance of complex remote sensing small targets, the fourth scale is introduced in the sampling of the YOLO-V5 algorithm framework. Compared with the traditional three-scale detection layer for multi-scale target detection, for remote sensing image datasets in complex environments, the number of small targets far exceeds the number of medium-sized targets and large targets. The introduction of the fourth scale can better obtain finer-grained features and more detailed location information in the feature map, effectively improving the accuracy of algorithm detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 The overall method flow chart provided by the present invention;
[0032] Figure 2 It is a network structure block diagram of the improved YOLO-V5 of the design example of the present invention;
[0033] Figure 3 It is the Res2Unit network structure diagram proposed by the present invention;
[0034] Figure 4This is the Dense Block network structure diagram proposed in the present invention. DETAILED DESCRIPTION
[0035] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. 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.
[0036] Exemplary embodiments of the present invention will be described in detail with reference to the following drawings.
[0037] The exemplary embodiments of the present invention are described below in conjunction with specific circumstances:
[0038] Example 1: Combination Figure 1-Figure 4 As shown, the present invention provides a small target detection method for a remote sensing image with a complex background, comprising the following steps:
[0039] Step 1: Build the basic YOLO-V5 algorithm network structure;
[0040] The core idea of the YOLO-V5 algorithm is to transform the target detection problem into a regression problem, dividing the image into a fixed number of grids (such as S×S). For each grid cell, the YOLO-V5 algorithm predicts N bounding boxes; for each bounding box, the YOLO-V5 algorithm predicts the position of the bounding box (x, y, w, h) and its confidence p. (x, y) represents the center of the box relative to the boundary of the grid cell, while w and h represent the width and height of the bounding box relative to the entire image. The confidence p is the probability that an object exists in each bounding box, defined as
[0041] Among them, IoU is the intersection of the predicted box and the true value, which is a number between 0 and 1; the closer it is to 1, the closer the predicted bounding box is to the true value.
[0042] The network output of the YOLO-V5 algorithm contains a total of S×S grid cells, each of which predicts N bounding boxes. Each bounding box predicts 5 values (x, y, w, h, p). In addition, each grid cell needs to predict C categories. Therefore, the output tensor size of the YOLO-V5 algorithm network is S×S×(N×5+C).
[0043] The architecture of the YOLO-V5 algorithm can be divided into three key parts: the backbone (feature extractor network), the neck, and the head. The backbone is a deep convolutional neural network that extracts features from the input image and is usually composed of multiple layers of convolution and pooling operations that progressively downsample the input image. The purpose of the backbone is to extract high-level features that can be used to identify objects in the image.
[0044] The neck is a set of intermediate layers connecting the backbone and the head, usually composed of convolutional layers with different kernel sizes to capture features at different scales. The core task of the neck is to capture high and low features from the backbone and combine them to achieve the purpose of improving the accuracy of object detection.
[0045] The head is the last part of the structure and is used to predict the class probabilities and bounding boxes of the detected objects. It usually consists of a set of fully connected layers that take the outputs from the neck and produce the final predictions. The head also includes a set of anchor boxes, which have the effect of predefined boundaries and are used to detect objects of different sizes and shapes.
[0046] Step 2: Construct a Res2Unit network based on the residual network idea;
[0047] The residual network in the YOLO-V5 structure effectively alleviates the gradient attenuation problem of the network on the basis of strengthening the feature extraction network, but the residual network still represents multi-scale features in a hierarchical manner, which will result in the features of each layer not being fully utilized. To solve this problem, a new residual unit connection method is proposed to extract features. In this method, a hierarchical residual-like connection is constructed in a residual network to complete feature extraction, and it is named Res2Unit. Res2Unit represents multi-scale features in a finer granularity and expands the receptive field of each layer. Based on the core idea of Res2Unit, several small residual terms are introduced into the residual unit to achieve the purpose of expanding the receptive field of each layer. Compared with the basic residual unit, the "Res2Unit" structure proposed in this paper is shown in the attached figure. Figure 3 shown.
[0048] In "Res2Unit", the input feature map is evenly divided into 4 sub-features after passing through the convolution layer (1×1). The sub-features are named x1, x2, x3, and x4, and their dimensions are the same. It should be noted that the significant difference between the sub-feature map and the input feature map is that it only contains 1 / 4 of the channels. It is worth noting that y1, y2, y3, and y4 are in the attached Figure 3 It can be expressed as
[0049]
[0050] Where * represents convolution; Conv represents a convolution layer with a 3×3 convolution kernel.
[0051] From the attached Figure 3 It is not difficult to find from the structural comparison diagram that compared with the residual unit, the improved "Res2Unit" can better utilize contextual information and effectively improve the detection performance of small targets and targets disturbed by the environment. In addition, the extraction of multi-scale features also enhances the semantic representation of the network to a certain extent.
[0052] Step 3: Establish a densely connected network and the fourth scale to effectively and accurately detect small remote sensing targets under complex backgrounds;
[0053] Establish a densely connected network and the fourth scale, consider factors such as the size of small targets, lighting changes, and background color proximity, and effectively and accurately detect small remote sensing targets in complex backgrounds;
[0054] The step three comprises the following steps:
[0055] S31. Build a basic DenseNet model; in the YOLO-V5 algorithm structure diagram, each detection layer contains no less than 7 convolutional layers;
[0056] S32, extract feature maps of different scales through different Dense Blocks; introduce the concept of densely connected network (DenseNet). Figure 4 In, x i represents the output feature map; H0(·) and H i (·) refers to the convolutional layer and the transmission layer respectively. Each layer in a densely connected network is connected to all other layers, which means that each layer will receive all the feature maps of the previous layer. With its unique structure, DenseNet can effectively alleviate the gradient attenuation of the network. At the same time, since each layer in a densely connected network receives all the feature maps of the previous layer, the transmission of network features will be enhanced to a certain extent, and the parameter volume of the network model will also be reduced.
[0057] It is worth noting that H i (·), i=1,2,3,4 is defined as: Conv(1x1xM)-BN-ReLU-Conv(3x3xM)-BN-ReLU. The increment of each layer of feature map of "DenseBlock*i" can be calculated according to the input and output dimensions of DenseBlock. Compared with the C3 module in the YOLO-V5 algorithm, the introduced DenseBlock can better alleviate gradient fading and greatly reduce the parameters of the network.
[0058] S33. Extract a high-resolution feature map and introduce the fourth-scale feature to capture finer-grained information. In the YOLO-V5 algorithm, three scale detection layers are used to detect multi-scale targets, in which the scales of 32x, 16x, and 8x in the feature map are downsampled to be responsible for the detection of large targets, medium targets, and small targets, respectively. However, it is worth noting that for remote sensing image datasets in complex environments, the number of small targets far exceeds the number of medium and large targets. On this basis, in order to obtain finer-grained features and more detailed location information, the fourth scale of the feature map that is downsampled 4 times is selected as the new detection layer and integrated into the improved YOLO-V5 algorithm network structure. The specific improved YOLO-V5 algorithm network structure diagram is shown in the attached figure. Figure 2 shown.
[0059] S34, fuse feature maps of different scales.
[0060] Assume that we already have a model that can output multiple feature maps of different scales. These feature maps may come from different convolutional layers or different DenseBlocks. Use the upsampling method to enlarge the low-resolution feature map to the same size as the high-resolution feature map, and then concatenate them to form a new feature map.
[0061] Step 4: Use a variety of remote sensing small target image datasets as training sets and test sets to train and test the improved YOLO-V5 target detection algorithm.
[0062] Specifically: In order to improve the comprehensive detection performance of the improved YOLO-V5 algorithm for remote sensing small targets in complex environments, a variety of remote sensing small target image data sets were selected as training sets and test sets, and the improved YOLO-V5 target detection algorithm was trained and tested to achieve the ultimate goal of real-time remote sensing small target detection.
[0063] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.
[0064] The above specific examples are used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.
Claims
1. A small target detection method for remote sensing images with complex backgrounds, characterized in that: The following steps are involved: Step 1: Build the basic YOLO-V5 algorithm network structure; Step 2: Construct a Res2Unit network based on the residual network idea; Step 3: Establish a densely connected network and the fourth scale to effectively and accurately detect small remote sensing targets under complex backgrounds; Step 4: Use a variety of remote sensing small target image data sets as training sets and test sets to train and test the improved YOLO-V5 target detection algorithm.
2. The small target detection method for remote sensing images with complex background according to claim 1, characterized in that: The process of building the basic YOLO-V5 algorithm network structure includes: S11, transform the target detection problem into a regression problem and divide the image into a fixed number of grids; S12. Use the YOLO-V5 algorithm to predict N bounding boxes of the network unit, and predict the position (x, y, w, h) of the bounding box and its confidence p; The network output of the S13 and YOLO-V5 algorithms contains a total of S×S grid cells, and the output tensor size is S×S×(N×5+C).
3. The small target detection method in a remote sensing image with a complex background according to claim 2, characterized in that: The position (x, y, w, h) of the bounding box and its confidence p, (x, y) represents the center of the box relative to the boundary of the grid cell, while w and h represent the width and height of the bounding box relative to the entire image; the confidence p is the probability that an object exists in each bounding box, defined as Among them, IoU is the intersection of the predicted box and the true value, which is a number between 0 and 1.
4. The small target detection method in a remote sensing image with a complex background according to claim 1, characterized in that: The construction of a Res2Unit network based on the residual network idea specifically does not include: S21, construct a hierarchical residual class connection in a residual network to complete feature extraction, and name it Res2Unit; S22, several small residual terms are introduced in the residual unit to achieve the purpose of expanding the receptive field of each layer; S23. Construct output formula: Where * represents convolution; Conv represents the convolution layer of 3×3 convolution kernel; x1, x2, x3, x4 are the input feature maps that are evenly divided into 4 sub-features after passing through the convolution layer (1×1).
5. The small target detection method in a remote sensing image with a complex background according to claim 1, characterized in that: The step 3 of establishing a densely connected network and a fourth scale specifically includes: S31. Build a basic DenseNet model; S32, extract feature maps of different scales through different Dense Blocks; S33, extract a high-resolution feature map to introduce the fourth scale feature to capture finer-grained information; S34, fuse feature maps of different scales.
6. The small target detection method in a remote sensing image with a complex background according to claim 5, characterized in that: The Dense Block in S32 is used The structure of Conv(1x1xM)-BN-ReLU-Conv(3x3xM)-BN-ReLU, in which 1x1 convolution is used to reduce the number of input feature maps and reduce the amount of calculation.
7. The small target detection method in a remote sensing image with a complex background according to claim 5, characterized in that: The fourth scale in step S33 is a feature map that is downsampled four times.