Panoramic image target detection method and system based on improved YOLOv7

Abstract

This application discloses a method and system for panoramic image object detection based on an improved YOLOv7. The method includes the following steps: acquiring real panoramic video images to obtain an image dataset; performing feature annotation on the image dataset to construct an object detection dataset; improving the YOLOv7 network to construct an initial model; training the initial model based on the object detection dataset to obtain a panoramic image object detection model; and performing object detection based on the panoramic image object detection model. This application can address the lack of data for panoramic image feature research and the problem of inaccurate detection of objects in panoramic images that are too small.

Description

Technical Field

[0001] The present application belongs to the field of target detection technology, and specifically relates to a panoramic image target detection method and system based on improved YOLOv7. Background Art

[0002] Panoramic cameras are now widely used in fields such as intelligent ships, intelligent driving, and airport security. The panoramic images they capture are the cornerstone of technologies such as information perception, communication and navigation, route planning, and early warning. However, these technologies also present new challenges. Objects in panoramic images are extremely small compared to the overall image, and existing object detection technologies are inaccurate for small objects. Therefore, it is crucial to leverage computer vision technology to develop and design an object detection algorithm for panoramic images.

[0003] Common object detection techniques can be categorized into two main types: two-stage detection and one-stage detection. Two-stage detection methods, such as algorithms like R-CNN, Fast R-CNN, and Faster R-CNN, first generate a set of candidate detection boxes and then predict the location and category of each box. One-stage detection methods, such as object detection algorithms like SDD and YOLO, predict the location and category directly within the detection network.

[0004] The above aspects have the following problems in constructing panoramic image target detection: First, the above-mentioned current target detection methods are only improved and optimized for common data sets, such as PASCAL VOC and MS COCO, and lack data sets suitable for panoramic images. Second, the accuracy of small target detection in the existing YOLO target detection series is low, and most of the targets in panoramic images are small objects, which affects the accuracy of panoramic image target detection. The present invention provides a panoramic image target detection method based on improved YOLOv7, which collects panoramic images through a panoramic camera to construct a feature dataset that meets the panoramic image requirements, and constructs a target detection method for panoramic images based on the YOLOv7 framework. Finally, the above-mentioned data sets and evaluation indicators are used to verify the present invention. Summary of the Invention

[0005] This application aims to address the shortcomings of the existing technology and proposes a panoramic image target detection method and system based on improved YOLOv7. The ECA attention module and VIT module are added to the backbone network, the ECA attention module is added to the neck network, and the small target detection layer is added to the head network. A target detection model for panoramic images is constructed, and a real panoramic video image dataset is collected using the established panoramic camera. The trained network model is verified and verified through target detection evaluation indicators to obtain the final detection model. Finally, the detection model is used to perform panoramic image target detection.

[0006] To achieve the above objectives, this application provides the following solutions:

[0007] The panoramic image target detection method based on the improved YOLOv7 includes the following steps:

[0008] Collect real panoramic video images to obtain image datasets;

[0009] Performing feature annotation on the image dataset to construct a target detection dataset;

[0010] Improve the YOLOv7 network and build an initial model;

[0011] Training the initial model based on the target detection dataset to obtain a panoramic image target detection model;

[0012] Target detection is performed based on the panoramic image target detection model.

[0013] Preferably, the feature annotation method includes:

[0014] Classify target location labels;

[0015] The target location label format is set to a YOLO format label with a center point plus height and width;

[0016] The image dataset is labeled with targets based on target position labels to obtain a target detection dataset.

[0017] Preferably, the improvement method of the YOLOv7 network includes:

[0018] Add the first ECA attention module and ViT module to the backbone network;

[0019] Add a second ECA attention module to the neck network;

[0020] Add a small object detection layer to the head network.

[0021] Preferably, the training method includes:

[0022] Set the configuration file for network training and set the environment for network training;

[0023] Dividing the target detection dataset into a training set and a test set according to a preset ratio;

[0024] Training the initial model based on the training set to obtain a trained model;

[0025] The trained model is tested based on the test set and verified using target detection evaluation indicators to obtain the panoramic image target detection model.

[0026] This application also provides a panoramic image target detection system based on improved YOLOv7, including: an image acquisition system, a feature annotation system, a model construction system, a model training system and a target detection system;

[0027] The image acquisition system is used to acquire real panoramic video images to obtain an image data set;

[0028] The feature annotation system is used to perform feature annotation on the image dataset to construct a target detection dataset;

[0029] The model building system is used to improve the YOLOv7 network and build an initial model;

[0030] The model training system is used to train the initial model based on the target detection dataset to obtain a panoramic image target detection model;

[0031] The target detection system is used to perform target detection based on the panoramic image target detection model.

[0032] Preferably, the feature annotation system includes: a label classification subsystem, a label setting subsystem and an annotation subsystem;

[0033] The label classification subsystem is used to classify target location labels;

[0034] The label setting subsystem is used to set the format of the target position label to a YOLO format label with a center point plus height and width;

[0035] The labeling subsystem is used to perform target labeling on the image dataset based on the target position label to obtain a target detection dataset.

[0036] Preferably, the model building system includes: a backbone network optimization subsystem, a neck network optimization subsystem and a head network optimization subsystem;

[0037] The backbone network optimization subsystem is used to add a first ECA attention module and a ViT module to the backbone network;

[0038] The neck network optimization subsystem is used to add a second ECA attention module in the neck network;

[0039] The head network optimization subsystem is used to add a small target detection layer to the head network.

[0040] Preferably, the model training system includes: a parameter configuration subsystem, a data partitioning subsystem, a training subsystem and a testing subsystem;

[0041] The parameter configuration subsystem is used to set the configuration file for network training and set the environment for network training;

[0042] The data partitioning subsystem is used to divide the target detection data set into a training set and a test set according to a preset ratio;

[0043] The training subsystem is used to train the initial model based on the training set to obtain a trained model;

[0044] The testing subsystem is used to test the trained model based on the test set and verify it through target detection evaluation indicators to obtain the panoramic image target detection model.

[0045] Compared with the prior art, the present invention has the following advantages:

[0046] This application can solve the problem of lack of data for panoramic image feature research; it can also solve the problem of inaccurate detection of targets in panoramic images that are too small. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solution of the present application, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0048] Figure 1 A schematic diagram of a method flow chart of an embodiment of the present application;

[0049] Figure 2 This is a schematic diagram of the ECA attention module structure of an embodiment of the present application;

[0050] Figure 3 This is a schematic diagram of the ViT module structure of an embodiment of the present application;

[0051] Figure 4 A schematic diagram of the overall network structure of an embodiment of the present application;

[0052] Figure 5 This is a schematic diagram of the CBS structure of an embodiment of the present application;

[0053] Figure 6 This is a schematic diagram of the E-ELAN structure of an embodiment of the present application;

[0054] Figure 7 This is a schematic diagram of the MP structure of an embodiment of the present application;

[0055] Figure 8 This is a schematic diagram of the E-ELAN-H structure of an embodiment of the present application;

[0056] Figure 9This is a schematic diagram of the SPPCCSPC structure of an embodiment of the present application;

[0057] Figure 10 Schematic diagram of the system structure of an embodiment of the present application. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0059] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0060] Example 1

[0061] In this embodiment, if Figure 1 As shown in FIG, the panoramic image target detection method based on the improved YOLOv7 includes the following steps:

[0062] S1. Collect real panoramic video images to obtain an image dataset.

[0063] In this embodiment, taking the airport as an example, image data actually captured by the airport's panoramic camera is collected. Common targets in the image data mainly include airplanes and vehicles, and an image dataset is constructed based on the target characteristics in the image data.

[0064] S2. Feature labeling of image datasets and constructing target detection datasets.

[0065] The feature labeling method includes: classifying the target position labels; setting the format of the target position labels to a YOLO format label with a center point plus height and width; and labeling the image dataset based on the target position labels to obtain a target detection dataset.

[0066] In this example, the target location labels are categorized as airplanes and cars, and the location labels are formatted as a YOLO label with a center point plus height and width. The location information labels were annotated using the labeling program Labelimg, resulting in a target detection dataset. A total of 100 images were annotated, with 8,660 objects annotated. According to the definition of a small target, an object is considered small if it occupies less than 0.1 of the image size. In this dataset, there are 8,648 small targets, representing 99.8% of the dataset.

[0067] S3. Improve the YOLOv7 network and build an initial model.

[0068] The improvement methods of the YOLOv7 network include: adding the first ECA attention module and ViT module to the backbone network; adding the second ECA attention module to the neck network; and adding a small target detection layer to the head network.

[0069] In this embodiment, the attention network mechanism makes the model more interested in a certain type of target, reduces useless information to increase detection capabilities, and thus improves the network's ability to detect small target objects in panoramic images. The ECA attention module obtains each channel feature value by average pooling on the main input edge. The channel feature value uses a 1×1 convolution kernel to realize the channel weight between feature vectors. The feature map channel weight is obtained through the sigmoid activation function, and is multiplied with the module input feature map to obtain the ECA module output feature map. The network structure is as follows: Figure 2 As shown in the structure diagram, GAP is the global average pooling operation, and K is the convolution kernel size of the one-dimensional convolution operation. The formula is as follows:

[0070]

[0071] In the formula: is the output feature map of the ECA module, Y is the input feature map of the ECA module, sigmoid is the activation operation, Conv1 is the 1D convolution operation, and AvgPool is the global average pooling operation.

[0072] The Transformer algorithm is a neural network based on the self-attention mechanism. Initially applied to natural language processing, its powerful feature representation capabilities have led to its expansion into computer vision tasks, where it has achieved outstanding performance. Traditional convolution operations obtain weights through local receptive fields and use aggregation operations to share these weights across the entire feature map. The Vision Transformer (ViT) model, based on the Self-Attention architecture, employs a weighted average operation based on its own context. By calculating attention weights for related pixels, it adaptively captures different regions within the global receptive field, capturing more features. Its core concept is to segment the input image into several small patches and then encode these patches using the Transformer model to generate the image's feature map. The ViT algorithm takes an RGB image of size H×W×C as input. It first evenly segments the image into several P×P patches, resulting in N patches of size HW / P^2. Each patch is considered a sequence, with each pixel considered an element in the sequence, resulting in N sequences of length P×P. Next, the ViT algorithm encodes these N sequences using the Transformer model. Each encoding layer contains multiple layers of multi-head self-attention modules, normalization layers, and residual connections for feature extraction. An N×D matrix is obtained, where D is the dimension of the feature. Then, the feature representation of these N small blocks is obtained through a simple splicing operation to obtain the feature representation of the entire image. At this time, the feature image has better generalization performance and interpretability. This embodiment adds a ViT module to the backbone network to increase the feature map features, thereby improving the detection algorithm's performance in detecting small targets in panoramic images. The network structure is as follows: Figure 3 shown.

[0073] The YOLOv7 framework downsamples the original image by 8x, 16x, and 32x before sending it to the feature detection network. This results in a 20×20 large object detection feature map, a 40×40 medium object detection map, and an 80×80 small object detection map. Due to their small size, the YOLOv7 object detection algorithm often overlooks or misses small objects, affecting detection accuracy and recall. Since all objects detected in panoramic images are small, their pixel features are similar to the background, which can lead to overlooking or missing detections by the object detection algorithm. In this embodiment, a new scale feature map is added to the original algorithm. The original image is downsampled by 4x, 8x, 16x, and 32x, resulting in a 20×20 large object detection feature map, a 40×40 medium object detection feature map, an 80×80 small object detection feature map, and a 160×160 small object detection feature map. Specifically, the feature map is upsampled by the ELAN-H layer and then fused with the first ELAN layer in the backbone network to generate a new feature map. The generated feature map is passed through ELAN to obtain a larger feature layer. This feature layer serves as the new head detection layer for small object detection. Further downsampling is performed to integrate shallower detail features into the neck feature fusion network. These four detection layers improve the accuracy and recall of small object detection in panoramic images, making the algorithm more applicable to panoramic images.

[0074] The overall improved YOLOv7 network structure is the same as the YOLOv7 network framework. The network consists of a backbone network, a neck network, and a head network. Figure 4 shown.

[0075] The backbone network includes CBS structure (such as Figure 5 As shown), E-ELAN structure (as shown Figure 6 As shown), MP1 structure (as Figure 7 The ViT architecture is shown in Figure 2. The CBS architecture consists of a convolutional module, a normalization module, and a SiLU activation function. The E-ELAN architecture builds on the original ELAN architecture in YOLOv7 by adding an ECA attention module to the input. The MP1 architecture consists of a max pooling module and a stride-2 convolution, combined with two downsampling methods, enhancing network learning without disrupting the architecture.

[0076] The neck network includes FPN structure, PAN structure, MP2 structure, E-ELAH-H structure (such as Figure 8 As shown) and SPPCCSPC structure (as Figure 9(as shown). The FPN structure is a top-down feature pyramid, utilizing top-down methods to improve small object detection. The PAN structure is a bottom-up feature pyramid, leveraging information from lower layers to upper layers to improve occluded object detection. The E-ELAN-H structure builds on the original ELAN-H structure in YOLOv7 by adding an ECA attention module to the input. The SPPCCSPC structure uses three CBS operations on the input feature map, followed by three max pooling operations combined using a concat operation. The SPPCCSPC structure then undergoes two CBS operations before being combined with the input using a concat operation. This module is a legacy YOLOv7 module.

[0077] In the head network, the YOLOv7 framework downsamples the original image by 8x, 16x, and 32x before sending it to the feature detection network, resulting in a 20×20 large target detection feature map, a 40×40 medium target feature detection map, and an 80×80 small target feature detection map. In target detection algorithms, it is believed that although deep convolutions provide rich semantic information, they provide less positional information and some small target information is lost. Based on the fact that most targets in panoramic images are small, a new scale feature map is added to the original algorithm. The original image is downsampled by 4x, 8x, 16x, and 32x, resulting in a 20×20 large target detection feature map, a 40×40 medium target feature detection map, an 80×80 small target feature detection map, and a 160×160 small target feature detection map, which are then sent to the detection network. This embodiment improves the network structure for panoramic image target features, which is conducive to improving the overall performance of panoramic image target detection methods.

[0078] S4. Train the initial model based on the target detection dataset to obtain a panoramic image target detection model.

[0079] The training method includes: setting a configuration file for network training and setting an environment for network training; dividing the target detection data set into a training set and a test set according to a preset ratio; training the initial model based on the training set to obtain a trained model; testing the trained model based on the test set and verifying it through target detection evaluation indicators to obtain a panoramic image target detection model.

[0080] In this embodiment, the improved YOLOv7 data and model configuration file need to be changed before training, the number of object categories in the data file is changed to 3, and the object category names are modified in the category list names. The anchor box sizes are set to (5, 9), (6, 14), (43, 15), (12, 16), (19, 36), (40, 28), (36, 75), (76, 55), (72, 146), (142, 110), (192, 243), (459, 401) in the model configuration file. The feature scale and attention module are added according to the improved network structure, the ECA model code is added to the common.py file, the ECA structure is introduced in the yolo.py file, and the yolo.py file is run to detect whether the network changes are correct.

[0081] Next, we set up the network training environment. In this example, the network environment is the Sugon cloud computing service system. Each node is equipped with a 32-core 2.5GHz x86 processor and an NVIDIA Tesla V100 accelerator card. Each node is equipped with two 16GB DDR4 2666ECC REG memory sticks and two Sugon Parastor300S parallel storage systems to provide large-capacity data storage. For network communication, the cluster uses a full-line, non-blocking 200Gb HDR Infiniband dedicated computing network and the PyTorch 1.9.0 deep learning framework. The specific configuration information is shown in Table 1-3.

[0082] Table 1

[0083] Network test hardware environment configuration

[0084]

[0085] Table 2

[0086] Network test software partial environment configuration

[0087]

[0088] Table 3

[0089] YOLOv7 improves some model parameter settings

[0090]

[0091] The target detection dataset is divided into 7:3 ratio, that is, 70 images are used as training set and 30 images are used as test set. The initial model is trained with the training set to obtain the trained model; the trained model is tested based on the test set.

[0092] In this example, Recall, Precision, F1, and mAP are used as evaluation indicators for the model, and the IOU threshold is set to 0.5. F1 is the harmonic mean of Recall and Precision, which can provide a more accurate response for the model. mAP is the average of the average accuracy of multiple targets under different Recall conditions. The definitions of Recall, Precision, F1, and mAP are as follows:

[0093]

[0094]

[0095]

[0096]

[0097] Among them, TP represents the number of correctly identified positive samples, FP represents the number of incorrectly identified positive samples, FN represents the number of missed positive samples, and m represents the number of identified categories.

[0098] S5. Perform target detection based on the panoramic image target detection model.

[0099] Example 2

[0100] In this embodiment, in order to verify the effectiveness of the improved panoramic image target detection model proposed in this application, the network input image resolution is 640×640×3 and trained for 5000 epochs. The detection accuracy of YOLOv3, YOLOv4, YOLOv5l, YOLOv7, and improved YOLOv7 networks are compared. The experimental results are shown in Table 4:

[0101] Table 4

[0102]

[0103] Experimental results show that the improved YOLOv7 panoramic image target detection model proposed in this application has the best performance in all the evaluation indicators in this embodiment. This application improves the mAP0.5 of YOLOv3 by 21.1%, the mAP0.5 of YOLOv4 by 14.1%, the mAP0.5 of YOLOv5l by 11.5%, and the mAP0.5 of YOLOv7 by 18.1%. The panoramic image detection results are analyzed using YOLOv3, YOLO, YOLOv5l, YOLOv7 and the improved YOLOv7 panoramic image target detection model. The results show that the targets that the YOLO algorithm failed to detect can be detected by the improved YOLOv7 panoramic image target detection model, and the detection accuracy is higher.

[0104] Example 3

[0105] In this embodiment, if Figure 10 As shown in FIG, a panoramic image target detection system based on improved YOLOv7 includes: an image acquisition system, a feature annotation system, a model construction system, a model training system, and a target detection system.

[0106] The image acquisition system is used to collect real panoramic video images and obtain image data sets.

[0107] The feature annotation system is used to perform feature annotation on image datasets and construct target detection datasets.

[0108] The feature labeling system includes: a label classification subsystem, a label setting subsystem and a labeling subsystem; the label classification subsystem is used to classify the target position labels; the label setting subsystem is used to set the format of the target position labels to a YOLO format label with a center point plus height and width; the labeling subsystem is used to label the image dataset based on the target position labels to obtain a target detection dataset.

[0109] The model building system is used to improve the YOLOv7 network and build an initial model.

[0110] The model construction system includes: a backbone network optimization subsystem, a neck network optimization subsystem and a head network optimization subsystem; the backbone network optimization subsystem is used to add the first ECA attention module and the ViT module to the backbone network; the neck network optimization subsystem is used to add the second ECA attention module to the neck network; the head network optimization subsystem is used to add a small target detection layer to the head network.

[0111] The model training system is used to train the initial model based on the target detection dataset to obtain a panoramic image target detection model.

[0112] The model training system includes: a parameter configuration subsystem, a data partitioning subsystem, a training subsystem, and a testing subsystem; the parameter configuration subsystem is used to set the configuration file for network training and set the network training environment; the data partitioning subsystem is used to divide the target detection data set into a training set and a test set according to a preset ratio; the training subsystem is used to train the initial model based on the training set to obtain a trained model; the testing subsystem is used to test the trained model based on the test set, and verify it through target detection evaluation indicators to obtain a panoramic image target detection model.

[0113] The target detection system is used to perform target detection based on the panoramic image target detection model.

[0114] The embodiments described above are merely descriptions of the preferred embodiments of the present application and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements made to the technical solutions of the present application by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present application.

Claims

1. A panoramic image target detection method based on improved YOLOv7, characterized in that: The following steps are involved: Collect real panoramic video images to obtain image datasets; Performing feature annotation on the image dataset to construct a target detection dataset; Improve the YOLOv7 network and build an initial model; Training the initial model based on the target detection dataset to obtain a panoramic image target detection model; Performing target detection based on the panoramic image target detection model; The improvement method of the YOLOv7 network includes: Add the first ECA attention module and ViT module to the backbone network; Add a second ECA attention module to the neck network; Add a small object detection layer to the head network; The ECA attention module obtains the eigenvalue of each channel by average pooling on the main input side. The channel eigenvalue uses a 1×1 convolution kernel to realize the channel weight between feature vectors. After the sigmoid activation function, the feature map channel weight is obtained. The ECA module output feature map is multiplied by the module input feature map. In the structure diagram, GAP is the global average pooling operation, and K is the convolution kernel size of the one-dimensional convolution operation. The formula is as follows: In the formula: is the output feature map of the ECA module, Y is the input feature map of the ECA module, sigmoid is the activation operation, Conv1 is the 1D convolution operation, and AvgPool is the global average pooling operation; The backbone network consists of CBS, E-ELAN, MP1, and ViT structures. The CBS structure consists of a convolution module, a normalization module, and a SiLU activation function. The E-ELAN structure adds an ECA attention module to the input based on the original ELAN structure in YOLOv7. The MP1 structure is composed of a maximum pooling module and a convolution with a stride of 2, and is recombined into two downsampling methods to enhance the network's learning ability without breaking the structure. The neck network includes FPN structure, PAN structure, MP2 structure, E-ELAH-H structure and SPPCCSPC structure; among them, the FPN structure is a top-down feature pyramid, which uses the upper adoption method to improve the detection ability of small targets. PAN is a bottom-up feature pyramid, which uses the lower-level information to pass to the upper layer to improve the detection ability of occluded targets. The E-ELAN-H structure is based on the original ELAN-H structure in YOLOv7, and adds an ECA attention module to the input. The SPPCCSPC structure input feature map undergoes three CBS operations, and the three maximum pooling operations are combined using the concat operation. After two CBS operations, it is combined with the input using the concat operation. This module is the original module of YOLOv7; In the head network, the YOLOv7 framework downsamples the original image by 8x, 16x, and 32x before sending it to the feature detection network, resulting in a 20×20 large target detection feature map, a 40×40 medium target feature detection map, and an 80×80 small target feature detection map. The target detection algorithm considers that although deep convolution has rich semantic information, it lacks positional information and some small target information is lost. Given that most targets in panoramic images are small, a new scale feature map is added to the original algorithm. The original image is downsampled by 4x, 8x, 16x, and 32x, resulting in a 20×20 large target detection feature map, a 40×40 medium target feature detection map, an 80×80 small target feature detection map, and a 160×160 small target feature detection map, which are then sent to the detection network. The feature annotation method includes: Classify target location labels; The target location label format is set to a YOLO format label with a center point plus height and width; Performing target labeling on the image dataset based on the target position label to obtain a target detection dataset; The training method includes: Set the configuration file for network training and set the environment for network training; Dividing the target detection dataset into a training set and a test set according to a preset ratio; Training the initial model based on the training set to obtain a trained model; The trained model is tested based on the test set and verified using target detection evaluation indicators to obtain the panoramic image target detection model.

2. A panoramic image target detection system based on improved YOLOv7, characterized by: include: Image acquisition system, feature annotation system, model building system, model training system and target detection system; The image acquisition system is used to acquire real panoramic video images to obtain an image data set; The feature annotation system is used to perform feature annotation on the image dataset to construct a target detection dataset; The model building system is used to improve the YOLOv7 network and build an initial model; The model training system is used to train the initial model based on the target detection dataset to obtain a panoramic image target detection model; The target detection system is used to perform target detection based on the panoramic image target detection model; The feature annotation system includes: a label classification subsystem, a label setting subsystem and an annotation subsystem; The label classification subsystem is used to classify target location labels; The label setting subsystem is used to set the format of the target position label to a YOLO format label with a center point plus height and width; The labeling subsystem is used to perform target labeling on the image dataset based on the target position label to obtain a target detection dataset; The model building system includes: a backbone network optimization subsystem, a neck network optimization subsystem and a head network optimization subsystem; The backbone network optimization subsystem is used to add a first ECA attention module and a ViT module to the backbone network; The neck network optimization subsystem is used to add a second ECA attention module in the neck network; The head network optimization subsystem is used to add a small target detection layer to the head network; The model training system includes: a parameter configuration subsystem, a data partitioning subsystem, a training subsystem and a testing subsystem; The parameter configuration subsystem is used to set the configuration file for network training and set the environment for network training; The data partitioning subsystem is used to divide the target detection data set into a training set and a test set according to a preset ratio; The training subsystem is used to train the initial model based on the training set to obtain a trained model; The testing subsystem is used to test the trained model based on the test set and verify it through target detection evaluation indicators to obtain the panoramic image target detection model.

Images