Ground environment target identification method and device based on DCF-YOLOv8

Abstract

The invention provides a ground environment target identification method and device based on DCF-YOLOv8, and solves the problems of complex background, serious similar target interference and low small target identification precision in target identification in a ground environment. According to the invention, a C2F module in a YOLOv8 model is improved, and a new target identification model DCF-YOLOv8 is provided. In the DCF module, firstly, a decomposition module decomposes a channel dimension of an input feature, and divides the dimension of the input feature into double branches; wherein one branch feature is input into a three-layer bottleneck convolution module; the three layers of bottleneck convolution modules adopt dense connection, so that the later bottleneck convolution module uses low-level feature information extracted by the previous bottleneck convolution module when extracting target features; and the output features of each layer of the two-branch and three-layer bottleneck convolution modules are spliced and are fused in the fusion module to serve as the output features of the DCF module.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing and relates to a ground environment target recognition method and device based on an improved YOLO model. Background Art

[0002] In the field of target recognition, the YOLO structure has been widely used in various application scenarios. However, in complex ground background environments, due to the large number of ground background elements, it is difficult for the target to accurately express its characteristics in the ground background environment, and the existing classic YOLO model needs to be improved.

[0003] YOLOv8 is a cutting-edge model in the field of deep learning, built on the improvements and innovations of YOLOv5, which has been widely used in many fields such as object recognition, image classification, instance segmentation and target tracking. YOLOv8 is easy to use and has multiple pre-trained models for different scenarios.

[0004] Figure 1 The traditional architecture of the YOLOv8 model is shown, which mainly includes the following parts: The input layer receives the input image and performs preprocessing, such as normalization and resizing, to adapt to the model requirements. The feature extraction layer consists of multiple convolutional layers and pooling layers to extract features from the image. YOLOv8 uses more efficient convolution operations and advanced feature extraction techniques to improve model performance. The feature pyramid network (FPN) enables the model to detect objects at different scales through multi-scale feature fusion, thereby improving the detection accuracy of small objects. The detection head includes multiple prediction layers, which output the bounding box coordinates, confidence scores, and category probability distributions of each grid cell. YOLOv8 uses an improved prediction mechanism to improve detection accuracy and speed. Post-processing includes non-maximum suppression (NMS) to remove overlapping detection boxes and retain the detection results with the highest confidence. In addition, steps such as threshold processing are included to obtain the final detection output. The architecture of YOLOv8 improves detection accuracy while maintaining high detection speed through multi-level feature extraction and optimized detection head design, especially the detection performance of small objects is significantly improved.

[0005] The YOLOv8 model uses Mosaic data enhancement technology during training. Its basic idea is to stitch multiple images into a large image, and then randomly select an area as a training sample, thereby increasing the diversity and complexity of the data set. During training, four images are randomly selected from the training data set, and the coordinates of a center point are randomly selected in the large image. Make sure that the selected center point is far enough away from the edge of the image to retain enough background area. The four images are stitched into a large image in a random order, with the selected center point as the center to form a Mosaic image. For the target in the Mosaic image, the coordinate information of its bounding box needs to be adjusted to adapt to the stitched image. At the same time, the category information and other attributes of the target (such as confidence, etc.) need to be adjusted accordingly. After generating the Mosaic image, other data enhancement techniques can be further applied, such as random horizontal flipping, random scaling, color distortion, etc., to increase the diversity and robustness of the data. Mosaic data enhancement improves the generalization ability of the model and has become the preferred algorithm for data preprocessing of most recognition algorithms.

[0006] DarkNet53 is a deep convolutional neural network composed of alternating convolutional layers and pooling layers. It mainly consists of convolutional layers, residual blocks, and fully connected layers. The convolutional layer usually uses a 3x3 convolution kernel with a stride of 1 and uses zero padding to keep the size of the feature map unchanged; the residual unit in the residual block adds the input features to the output features through direct connections across layers. This connection method helps to alleviate the gradient vanishing problem and improves the training speed and accuracy of the network; DarkNet53 usually contains one or more fully connected layers at the end to map the features extracted by the convolutional layer to the output space required for the target recognition task. Inspired by the VGG network model, the Darknet-53 backbone network in YOLOv8 doubles the number of channels after the pooling operation; reduces the amount of computation and compresses features by placing 1×1 convolution kernels between 3×3 convolution kernels; and uses batch normalization for regularization to accelerate model training convergence.

[0007] The CSP module is a feature extraction module that aims to improve the effectiveness of feature extraction by utilizing cross-stage connections. The introduction of cross-stage connections helps to reduce the loss of features during network transmission, while increasing the path of information transmission, thereby improving the network's feature representation ability.

[0008] In YOLOv8, the feature extractor uses CSPDarknet53, and the SPP module accepts feature maps as input and divides them into grids of different sizes. Unlike the traditional SPP module, the SPP module in YOLOv8 uses Max-Pooling and Concatenation operations. For each scale of grid, their features will be max-pooled, and then these pooled feature maps are concatenated together. The purpose of this is to retain the spatial information of feature maps at different scales and fuse them together to improve recognition performance.

[0009] In the target recognition task, due to the different frequencies of occurrence of targets of different categories, class imbalance may occur. At the same time, some targets may be more difficult to distinguish or more prone to errors, which leads to the problem of imbalance between difficult and easy samples. Traditional cross entropy loss may not be effective enough in dealing with these problems, so a more targeted loss function is needed. The DFL (Distribution Focal Loss) loss function is used in YOLOv8. DFL introduces a new adjustment factor to make the loss function more focused on samples that are difficult to handle in target recognition. This adjustment factor takes into account the distribution of each category in target recognition, so that the model pays more attention to categories that are prone to errors in training.

[0010] Based on the high accuracy, fast running speed and high versatility of the YOLOv8 algorithm, the present invention recognizes ground background targets based on the YOLOv8 structure. By improving the C2F module and training on the ground target data set, the recognition model proposed by the present invention has a higher accuracy. Summary of the invention

[0011] In view of this, the present invention provides a ground environment target recognition method based on an improved YOLO model, which can improve the target recognition accuracy in a complex interference environment on the ground.

[0012] In order to solve the above technical problems, the present invention is implemented as follows.

[0013] A ground environment target recognition method based on DCF-YOLOv8, the method adopts the YOLOv8 architecture to recognize ground environment targets, the method comprising:

[0014] Step 1: Construct a dataset for multi-target training and testing in a ground environment;

[0015] Step 2: Design a feature enhancement structure based on low-level feature sharing, namely the DCF module;

[0016] In the DCF module, the channel dimension of the input feature is first decomposed, and the input feature dimension (h, w, c) is divided into two branches with a branch dimension of (h, w, 0.5c); one of the branch features is processed by three consecutive bottleneck convolution modules; the three bottleneck convolution modules are densely connected, and the input of each bottleneck convolution module enters each subsequent bottleneck convolution module, so that the subsequent bottleneck convolution module uses the low-level feature information extracted by the previous bottleneck convolution module when extracting the target feature; the two branch features are feature spliced ​​with the output features of each layer processed by the multi-layer bottleneck convolution module, and the feature dimension is restored to (h, w, c), and then the dual branch feature fusion is realized through a set of convolution calculations as the output feature of the DCF module;

[0017] Step 3: Use the DCF module to replace the original C2F module in the YOLOv8 architecture to obtain the improved YOLOv8 model;

[0018] Step 4: Use the constructed dataset to train and test the improved YOLOv8 model;

[0019] Step 5: Use the trained YOLOv8 model to identify ground environment targets.

[0020] The present invention also provides a ground environment target recognition device based on DCF-YOLOv8, which adopts the YOLOv8 architecture as a basis. The present invention adopts a feature enhancement structure based on low-level feature sharing, namely, a DCF module, to replace the C2F module in the YOLOv8 architecture;

[0021] The DCF module includes a decomposition module, a three-layer bottleneck convolution module, and a fusion module;

[0022] The decomposition module is used to decompose the channel dimension of the input feature and divide the input feature dimension (h, w, c) into two branches; one branch feature is input into the three-layer bottleneck convolution module and the fusion module; the other branch feature is input into the fusion module;

[0023] The three-layer bottleneck convolution module adopts dense connection, and the input of each bottleneck convolution module enters each subsequent bottleneck convolution module, so that the subsequent bottleneck convolution module uses the low-level feature information extracted by the previous bottleneck convolution module when extracting target features; the output features of each layer of the three-layer bottleneck convolution module are input into the fusion module;

[0024] The two branches of the decomposition module output and the output of each layer of the three-layer bottleneck convolution module are concatenated, and the feature dimension is restored to (h, w, c). The features are fused in the fusion module as the output features of the DCF module.

[0025] Preferably, the decomposition module performs a 1 / 2 decomposition on the channel dimension of the input feature, dividing the input feature dimension (h, w, c) into a dual branch with a dimension of (h, w, 0.5c).

[0026] Preferably, the bottleneck convolution module includes 2 convolution layers; the input of the bottleneck convolution module and the output processed by the 2 convolution layers are spliced ​​together as the output of the bottleneck convolution module.

[0027] Preferably, the fusion module is implemented using convolution calculation.

[0028] Beneficial effects:

[0029] (1) Based on the YOLOv8 model structure, the present invention designs a feature enhancement structure based on low-level feature sharing in the feature extraction network, namely the DCF module, and combines it with the rest of YOLOv8 to form a DCF-YOLOv8 structure, which improves the target recognition accuracy in complex environments by enhancing the target's texture, edge and other detailed features.

[0030] (2) The DCF module converts channel features into spatial features through a two-branch feature decomposition operation, thereby better capturing low-level features such as edges and textures.

[0031] (3) Dense connections are used to fuse shallow low-level features with deep abstract features. By concatenating low-level features and high-level features for feature extraction, the model has better feature representation capabilities for multiple targets on the ground battlefield during training.

[0032] (4) The DCF module has a simple structure and is easy to implement. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is the YOLOv8 algorithm architecture diagram.

[0034] Figure 2 for Figure 1 C2F module structure diagram.

[0035] Figure 3 This is a flow chart of the ground environment target recognition method based on DCF-YOLOv8 of the present invention.

[0036] Figure 4 Edge identification graph for the target dataset.

[0037] Figure 5 The figure is a block diagram of the DCF module in the DCF-YOLOv8 model of the present invention. DETAILED DESCRIPTION

[0038] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0039] The present invention provides a ground environment target recognition scheme based on DCF-YOLOv8. The scheme aims at solving the problems of complex background, serious interference of similar targets and low recognition accuracy of small targets in target recognition under ground environment. Based on the YOLOv8 architecture, the C2F module in the YOLOv8 model is improved, and a feature enhancement structure based on low-level feature sharing, namely the DCF module, is designed in the feature extraction network, which is combined with the rest of YOLOv8 to form a new target recognition model, namely DCF-YOLOv8. The DCF module adopted in the scheme constructs a feature extraction network architecture suitable for expressing ground target features by cascading and sharing the low-level structure, thereby improving the target recognition accuracy in complex ground environment.

[0040] The experimental verification proves that the DCF-YOLOv8 model proposed in the present invention has obvious advantages in expressing low-level features such as texture and edge of the target when identifying the target. The algorithm is further tested on the target recognition data set with ground background, and the method of the present invention has higher recognition accuracy and better algorithm robustness.

[0041] Figure 3 The flowchart of the ground environment target recognition method based on DCF-YOLOv8 of the present invention is shown. As shown in the figure, the method includes the following steps:

[0042] Step 1: Construct a dataset for multi-target training and testing in a ground environment.

[0043] Step 2: Design a feature enhancement structure based on low-level feature sharing, namely the DCF module.

[0044] The original C2F module in the YOLOv8 algorithm has good model versatility. The target identified by the algorithm in this paper is a ground combat target dataset. The YOLOv8 algorithm has a high recognition accuracy. The C2F module in the YOLOv8 algorithm aggregates the output feature map to retain the gradient flow information. By analyzing the ground combat dataset, it is found that when identifying targets, the proportion of targets is large and the number is small. When the object is large, low-level features are usually more conducive to model training. Significant edge, texture and shape features can be easily captured by low-level features, thereby achieving accurate recognition, such as Figure 4 High-level features contain more object composition and context information. However, when the recognition targets are large and small in number, high-level features do not show significant advantages.

[0045] The C2F module only considers the output features of each bottleneck. Since shallow networks are easy to extract low-level features, target recognition under ground background relies on low-level features, which are often ignored as the network depth increases. Therefore, the present invention uses dense connections to fuse shallow low-level features with deep abstract features, and proposes a DCF module, i.e., a feature extraction module based on dense connections, so that the model can better extract low-level features of combat targets.

[0046] Figure 5 The implementation scheme of the DCF module of the present invention is shown. Figure 5 As shown, the DCF module includes a decomposition module, a three-layer bottleneck convolution module, and a fusion module.

[0047] First, the decomposition module decomposes the channel dimension of the input feature and divides the input feature dimension (h, w, c) into two branches with branch dimensions of (h, w, 0.5c). One branch feature is input into the three-layer bottleneck convolution module and processed by three consecutive bottleneck convolution modules. The other branch feature is input into the fusion module.

[0048] The three bottleneck convolution modules are densely connected, and the input of each bottleneck convolution module enters each subsequent bottleneck convolution module, so that the subsequent bottleneck uses the low-level feature information extracted by the previous bottleneck when extracting the target feature. The output features of each layer of the three-layer bottleneck convolution module are input into the fusion module.

[0049] The two branches output by the decomposition module and the output of each layer of the three-layer convolution module are concatenated, and the feature dimension is restored to (h, w, c); the features are fused in the fusion module as the output features of the DCF module. The fusion module here can use a set of convolution calculations to achieve dual-branch feature fusion.

[0050] The above operation combines low-level features with high-level features for feature extraction, and the model has better feature representation capabilities for multiple targets on the ground battlefield during training. The channel features are converted into spatial features through the dual-branch feature decomposition operation, thereby better capturing low-level features such as edges and textures.

[0051] Step 3: Use the DCF module to replace the original C2F module in the YOLOv8 architecture to obtain the improved YOLOv8 model;

[0052] Step 4: Use the constructed dataset to train and test the improved YOLOv8 model;

[0053] Step 5: Use the trained YOLOv8 model to identify ground environment targets.

[0054] Based on the above method, the ground environment target recognition device based on DCF-YOLOv8 provided by the present invention is based on the YOLOv8 architecture and adopts a feature enhancement structure based on low-level feature sharing, namely, a DCF module, to replace the C2F module in the YOLOv8 architecture. The specific structure of the DCF module is still referred to Figure 5 , including decomposition module, three-layer bottleneck convolution module, and fusion module.

[0055] The decomposition module is used to decompose the channel dimension of the input feature and divide the input feature dimension (h, w, c) into two branches; one branch feature is input into the three-layer bottleneck convolution module and the fusion module; the other branch feature is input into the fusion module. Preferably, the decomposition module decomposes the channel dimension of the input feature by 1 / 2 and divides the input feature dimension (h, w, c) into two branches with dimensions of (h, w, 0.5c).

[0056] The three-layer bottleneck convolution module uses dense connections. The input of each bottleneck convolution module enters each subsequent bottleneck convolution module, so that the subsequent bottleneck convolution module uses the low-level feature information extracted by the previous bottleneck convolution module when extracting target features; the output features of each layer of the three-layer bottleneck convolution module are input into the fusion module. Among them, the bottleneck convolution module includes 2 layers of convolution layers; the input of the bottleneck convolution module is spliced ​​with the output processed by the 2 layers of convolution layers as the output of the bottleneck convolution module.

[0057] The two branches output by the decomposition module and the outputs of each layer of the three-layer bottleneck convolution module are concatenated, and the feature dimension is restored to (h, w, c). The features are fused in the fusion module as the output features of the DCF module. The fusion module is preferably implemented by convolution calculation.

[0058] The above specific embodiments only describe the design principle of the present invention. The shapes and names of the components in the description may be different and are not limited. Therefore, those skilled in the art in the field of the present invention may modify or replace the technical solutions recorded in the above embodiments; and these modifications and replacements do not deviate from the creative purpose and technical solutions of the present invention and should all fall within the protection scope of the present invention.

Claims

1. A ground environment target recognition method based on DCF-YOLOv8, which adopts the YOLOv8 architecture to recognize ground environment targets, characterized in that: The method includes: Step 1: Construct a dataset for multi-target training and testing in a ground environment; Step 2: Design a feature enhancement structure based on low-level feature sharing, namely the DCF module; In the DCF module, the channel dimension of the input feature is first decomposed, and the input feature dimension (h, w, c) is divided into two branches with a branch dimension of (h, w, 0.5c); one of the branch features is processed by three consecutive bottleneck convolution modules; the three bottleneck convolution modules are densely connected, and the input of each bottleneck convolution module enters each subsequent bottleneck convolution module, so that the subsequent bottleneck convolution module uses the low-level feature information extracted by the previous bottleneck convolution module when extracting the target feature; the two branch features are feature spliced ​​with the output features of each layer processed by the multi-layer bottleneck convolution module, and the feature dimension is restored to (h, w, c), and then the dual branch feature fusion is realized through a set of convolution calculations as the output feature of the DCF module; Step 3: Use the DCF module to replace the original C2F module in the YOLOv8 architecture to obtain the improved YOLOv8 model; Step 4: Use the constructed dataset to train and test the improved YOLOv8 model; Step 5: Use the trained YOLOv8 model to identify ground environment targets.

2. A ground environment target recognition device based on DCF-YOLOv8, the device adopts the YOLOv8 architecture; characterized in that: Adopt a feature enhancement structure based on low-level feature sharing, namely the DCF module, to replace the C2F module in the YOLOv8 architecture; The DCF module includes a decomposition module, a three-layer bottleneck convolution module, and a fusion module; The decomposition module is used to decompose the channel dimension of the input feature and divide the input feature dimension (h, w, c) into two branches; one branch feature is input into the three-layer bottleneck convolution module and the fusion module; the other branch feature is input into the fusion module; The three-layer bottleneck convolution module adopts dense connection, and the input of each bottleneck convolution module enters each subsequent bottleneck convolution module, so that the subsequent bottleneck convolution module uses the low-level feature information extracted by the previous bottleneck convolution module when extracting target features; the output features of each layer of the three-layer bottleneck convolution module are input into the fusion module; The two branches of the decomposition module output and the output of each layer of the three-layer bottleneck convolution module are concatenated for feature concatenation. The feature dimension Restore to (h, w, c), perform feature fusion in the fusion module, and use it as the output feature of the DCF module.

3. The device according to claim 2, characterized in that The decomposition module performs 1 / 2 decomposition on the channel dimension of the input feature, dividing the input feature dimension (h, w, c) into two branches with dimensions (h, w, 0.5c).

4. The device according to claim 2, characterized in that The bottleneck convolution module includes two convolution layers; the input of the bottleneck convolution module and the output processed by the two convolution layers are spliced ​​together as the output of the bottleneck convolution module.

5. The device according to claim 2, characterized in that The fusion module is implemented by convolution calculation.

Images