Ship target detection method and system based on multi-scale Gaussian context feature excitation

Abstract

The invention relates to the technical field of remote sensing image processing, in particular to a ship target detection method and system based on multi-scale Gaussian context feature excitation, a ship target detection model is constructed based on a YOLOv8 network, and the ship target detection model is trained by using remote sensing sample data, in the ship target detection model, a backbone network is combined with a space and channel attention mechanism to carry out Gaussian context conversion and realize remote sensing image context feature excitation, and a neck network obtains a feature map fused with remote sensing image target semantics and spatial position information by using a multi-scale feature adaptive fusion mechanism; and inputting a to-be-detected remote sensing image into the trained ship target detection model, and identifying and extracting a ship target in the to-be-detected remote sensing image by using the ship target detection model. According to the method, identifiability expression of the foreground target can be realized, remote sensing image ship target detection efficiency and precision are improved, and the method has a good application prospect in the field of remote sensing image target detection.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and particularly relates to a ship target detection method and system based on multi-scale Gaussian context feature excitation. Background Art

[0002] As an important tool for maritime transportation and resource exploration, the intelligent detection and recognition of ships are widely applied in fields such as target reconnaissance, port change monitoring, traffic management, emergency rescue, and combating illegal activities. With the continuous improvement of the quality of remote sensing images obtained by high-resolution satellites, the target details are more abundant, providing sufficient data support for ship target detection.

[0003] Traditional ship target detection methods rely on manually extracting features, with poor real-time performance and low efficiency. Moreover, manual extraction mainly performs target detection through shallow features, and cannot fully consider deep information such as target semantics, making it difficult for their detection performance and detection effect to meet the actual application requirements. In recent years, deep learning methods based on convolutional neural networks (CNNs) have increasingly shown advantages in target detection, mainly divided into anchor-based and anchor-free target detection algorithms. Typical anchor-based target detection methods mainly include R-CNN, SPP-net, Faster R-CNN, YOLOv5, SSD, and RetinaNet, etc. However, these detection methods still face problems such as small ship target scales, few features, and variability in remote sensing images, as well as interference from complex environmental backgrounds (such as clouds and sea clutter). On the one hand, convolutional neural networks are limited by the local context features of convolutional kernels and cannot effectively capture the global context information in images. However, this is essential for understanding the target semantics in complex backgrounds. Scholars have enhanced the representation ability of the network by introducing channel attention modules such as SENet, but usually introduce a large number of parameters, and the complex relationship between the global context learned by the fully connected layer or linear transformation and the attention activation value is not deep enough. On the other hand, the shallow features in convolutional neural networks have high spatial resolution, rich texture information, and accurate spatial information, but the receptive field is small; the deep features have a larger receptive field and richer semantic information, but the resolution is low. Therefore, there is an urgent need for a remote sensing image target detection means to improve the accuracy and practicality of detecting ship targets of different sizes in complex scenarios. Summary of the Invention

[0004] Therefore, the present invention provides a ship target detection method and system based on multi-scale Gaussian context feature excitation to solve the problem that the detection accuracy and efficiency need to be improved due to small target scales, few features, and interference from complex environmental backgrounds in the existing ship target detection task of remote sensing images.

[0005] According to the design solution provided by the present invention, on the one hand, a ship target detection method based on multi-scale Gaussian context feature excitation is provided, including:

[0006] Construct a ship target detection model based on the YOLOv8 network, and train the ship target detection model using remote sensing sample data. Among them, the ship target detection model includes a backbone network for extracting features from the input remote sensing image, a neck network for fusing and enhancing the extracted features, and a detection head for detecting and outputting the target using the extracted features. The backbone network combines spatial and channel attention mechanisms to perform Gaussian context transformation and realizes remote sensing image context feature excitation. The neck network uses a multi-scale feature adaptive fusion mechanism to obtain a feature map fused with remote sensing image target semantics and spatial position information;

[0007] Input the remote sensing image to be detected into the trained ship target detection model, and use the ship target detection model to identify and extract the ship target in the remote sensing image to be detected.

[0008] As the ship target detection method based on multi-scale Gaussian context feature excitation of the present invention, further, the backbone network includes: a plurality of convolutional layers that respectively perform convolutional operations on the input remote sensing image to extract different scale features of the input image, a C2f component for performing feature transformation processing on the extracted remote sensing image features, a fast spatial pyramid pooling layer SPPF that enlarges the receptive field of the remote sensing image and pools the region of interest into a fixed size, and a Gaussian context transformation component that combines spatial and channel attention mechanisms to excite the remote sensing image context features.

[0009] As the ship target detection method based on multi-scale Gaussian context feature excitation of the present invention, further, combining spatial and channel attention mechanisms to excite the remote sensing image context features includes:

[0010] First, use the channel attention mechanism to assign weights to the channels of the extracted remote sensing image features, and obtain intermediate features of the remote sensing image based on the feature channel weights;

[0011] Then, fuse the intermediate features of the remote sensing image with the extracted remote sensing image features, and use the spatial attention mechanism to process the fused remote sensing image features to obtain the final activated features of the remote sensing image.

[0012] As the ship target detection method based on multi-scale Gaussian context feature excitation of the present invention, further, using the channel attention mechanism to assign weights to the channels of the extracted remote sensing image features includes:

[0013] Perform global information aggregation on the extracted features in space;

[0014] Calculate the global context feature mean based on the aggregated global information, and normalize the deviation of the global context feature mean using the global context standard deviation factor;

[0015] Use the Gaussian function to transform and activate the normalized global context feature mean to obtain the activation value of the channel attention weight, and multiply the activation value by the extracted remote sensing image features to obtain the intermediate features of the remote sensing image.

[0016] As the ship target detection method based on multi-scale Gaussian context feature excitation of the present invention, further, use the spatial attention mechanism to process the fused remote sensing image features, including:

[0017] For the fused remote sensing image features, generate two feature maps respectively through global max pooling operation and global average pooling operation, and splice the two feature maps according to the number of channels;

[0018] Use convolution operation to reduce the dimension of the spliced feature map, and use the Sigmoid activation function to generate the spatial attention features of the reduced-dimensional feature map, so as to multiply the spatial attention features with the fused remote sensing image features element by element and obtain the final activated features of the remote sensing image.

[0019] As the ship target detection method based on multi-scale Gaussian context feature excitation of the present invention, further, the multi-scale feature adaptive fusion mechanism weights and fuses the features of different levels of the remote sensing image through a bilateral fusion path and a skip connection path, wherein the bilateral fusion path is used to fuse the spatial resolution feature information of the shallow layer and the semantic feature information of the deep layer, and the bilateral fusion path includes a top-down semantic feature transfer path and a bottom-up spatial feature transfer path; the skip connection path is used to strengthen the feature correlation between the network input node and the network output node at the same layer, and the skip connection path is an additional cross-scale connection path added between the network input node and the network output node at the same layer.

[0020] As the ship target detection method based on multi-scale Gaussian context feature excitation of the present invention, further, the process of weighted fusion of features of different levels of the remote sensing image is expressed as: where w i represents the weight of the i-th feature layer, ε is a minimum value used to avoid numerical instability, F In represents the input feature, and F Out represents the weighted fusion feature result.

[0021] On the other hand, the present invention also provides a ship target detection system based on multi-scale Gaussian context feature excitation, comprising: a model construction module and a target detection module, wherein,

[0022] The model construction module is used to construct a ship target detection model based on the YOLOv8 network and train the ship target detection model using remote sensing sample data. Among them, the ship target detection model includes a backbone network for extracting features from the input remote sensing image, a neck network for fusing and enhancing the extracted features, and a detection head for detecting and outputting targets using the extracted features. The backbone network combines spatial and channel attention mechanisms to perform Gaussian context transformation and achieve remote sensing image context feature excitation. The neck network uses a multi-scale feature adaptive fusion mechanism to obtain a feature map that fuses the semantic and spatial position information of the remote sensing image target;

[0023] The target detection module is used to input the remote sensing image to be detected into the trained ship target detection model and use the ship target detection model to identify and extract the ship target in the remote sensing image to be detected.

[0024] Advantages of the present invention:

[0025] The present invention combines spatial and channel attention mechanisms to perform Gaussian context transformation of remote sensing image features, realizes the discriminative expression of foreground targets, and alleviates problems such as complex backgrounds and susceptibility to interference from similar targets in remote sensing images; introduces a multi-scale feature adaptive weighted fusion mechanism to construct a feature map with more sufficient information expression, and realizes the accurate representation of target semantics and spatial positions. Further experimental data show that compared with the reference method, the proposed scheme in this case has significant advantages in detection accuracy and efficiency, and the mAP value reaches 96.8%. The practicability and generalization of the proposed scheme are verified by selecting remote sensing images in different real scenarios, and it has good application prospects in the field of remote sensing image target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 Schematic diagram of the ship target detection process based on multi-scale Gaussian context feature excitation in the embodiment;

[0027] Figure 2 Schematic diagram of the ship target detection model architecture in the embodiment;

[0028] Figure 3 Schematic diagram of the working process of Gaussian context transformation SC-GCT in the embodiment;

[0029] Figure 4 Schematic diagram of the working process of multi-scale feature adaptive weighted bilateral fusion MFAF in the embodiment;

[0030] Figure 5Schematic illustration of the legend of the HRSC2016 dataset in the embodiment

[0031] Figure 6 Schematic comparison of ship target detection results in different scenarios in the embodiment

[0032] Figure 7 Schematic comparison of the detection results of the proposed solution in the embodiment with the ground truth

[0033] Figure 8 Schematic analysis of the detection effect of the proposed solution in the embodiment

[0034] Figure 9 Schematic illustration of ship target detection results on small-size images in the embodiment

[0035] Figure 10 Schematic illustration of ship target detection results on large-size images in the embodiment Detailed implementation manners

[0036] To make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and technical solutions

[0037] In view of the characteristics of high-resolution remote sensing images, such as large image width, complex background, and rich detailed features of ship targets, an embodiment of the present invention provides a ship target detection method based on multi-scale Gaussian context feature excitation, as Figure 1 shown, including:

[0038] S101. Construct a ship target detection model based on the YOLOv8 network, and train the ship target detection model using remote sensing sample data. Among them, the ship target detection model includes a backbone network for extracting features from the input remote sensing image, a neck network for fusing and enhancing the extracted features, and a detection head for detecting and outputting targets using the extracted features. The backbone network combines spatial and channel attention mechanisms to perform Gaussian context transformation and realize remote sensing image context feature excitation. The neck network uses a multi-scale feature adaptive fusion mechanism to obtain a feature map fused with remote sensing image target semantic and spatial position information

[0039] S102. Input the remote sensing image to be detected into the trained ship target detection model, and use the ship target detection model to identify and extract ship targets in the remote sensing image to be detected

[0040] Among them, the backbone network can be designed to include: multiple convolutional layers that respectively perform convolutional operations on the input remote sensing image to extract features of different scales of the input image, a C2f component that performs feature transformation processing on the extracted remote sensing image features, a fast spatial pyramid pooling layer SPPF that enlarges the receptive field of the remote sensing image and pools the region of interest into a fixed size, and a Gaussian context transformation component that stimulates the context features of the remote sensing image by combining the spatial and channel attention mechanisms.

[0041] As Figure 2 shown, first, the image is sent into the backbone network for feature extraction. Deep features of the target are extracted through multiple Conv and C2f modules, and then processed by the SPPF module to enlarge the receptive field and pool the region of interest into a fixed size. An SC-GCT module is added after the SPPF module to achieve context feature excitation through a Gaussian function that satisfies the negative correlation relationship between the preset global context and the attention activation value, improving the target detection efficiency. The output result of the Backbone is passed to the Neck layer, where an MFAF is introduced to replace the original Feature Pyramid Network FPN and PAN to construct a method for multi-scale feature expression, fully combining the high spatial resolution features of the shallow layer of the convolutional neural network and the strong semantic features of the deep layer, and considering the contribution degree of different input feature maps to the output features, to obtain a feature map with more sufficient information expression.

[0042] The Gaussian context transformation module SC-GCT that combines the spatial and channel attention mechanisms, as Figure 3 shown, combines feature excitation in both channel and spatial attention aspects. The input image is first processed by the channel attention mechanism, including three operations: Global Context Aggregation (GCA), normalization, and Gaussian Context Excitation (GCE).

[0043] Global semantic aggregation GCA operation: GCA aggregates global information in the space of the feature map to capture per-channel statistical information. The input feature map is X ∈ R C×H×W , and the global context information is expressed as:

[0044]

[0045] where C is the number of channels, and H and W are the width and height of the feature map.

[0046] Normalization operation: Define the mean shift as z - μ, where represents the mean of the global context z. Introduce a specific instance factor σ, that is, the standard deviation of the global context, to stabilize the distribution and ensure a mean of 0 and a variance of 1.

[0047]

[0048] GCE operation: The transformation and activation operations are implemented through a function f(·) based on the negative correlation hypothesis. To meet the negative correlation hypothesis, a continuous function needs to be found that satisfies the following conditions:

[0049] a) should have a value range within (0, 1] to be used as the attention value, that is

[0050] b) When takes the value of 0, obtains the unique maximum value of 1;

[0051] c) When is less than 0, increases monotonically; when is greater than 0, decreases monotonically;

[0052] d) There is

[0053] Among the known functions, the Gaussian function satisfies all the above conditions. Therefore, the following definition of the GCE operation is obtained, where a is the amplitude of the Gaussian function, set to 1 to meet condition a); b represents the mean of the Gaussian function, set to 0 to meet condition b); c represents the standard deviation of the Gaussian function, which controls the diversity of the channel attention map. The larger the standard deviation, the smaller the diversity of the activation values between channels.

[0054]

[0055] Through setting, G can be simplified to Equation (5), where c is a constant or a learnable parameter, and g is the attention activation value. It is multiplied with the original feature map to obtain the feature map adjusted by the channel attention mechanism.

[0056]

[0057] The features generated by the above operations are fused with the original features and output to the spatial attention mechanism. Through global max pooling and global average pooling operations, two feature maps of size H×W×1 are generated. The two feature maps are concatenated according to the number of channels to obtain richer spatial information, and then reduced to a single H×W×1 feature map through a 7×7 convolution. Then, the Sigmiod activation function is used to generate the spatial attention features to highlight the key regions in the feature map. Finally, the features are multiplied element-wise with the input features to obtain the final feature activation map.

[0058] Among them, the multi-scale feature adaptive fusion mechanism performs weighted fusion on the features of different levels of remote sensing images through the bilateral fusion path and the skip connection path. Among them, the bilateral fusion path is used to fuse the spatial resolution feature information of the shallow layer and the semantic feature information of the deep layer, and the bilateral fusion path includes a top-down semantic feature transfer path and a bottom-up spatial feature transfer path; the skip connection path is used to strengthen the feature correlation between the network input node and the network output node at the same layer, and the skip connection path is an additional cross-scale connection path added between the network input node and the network output node at the same layer.

[0059] As Figure 4 shown, in the embodiment of this case, by introducing two-way paths, namely top-down and bottom-up connections, the feature information of different levels is effectively integrated, and the feature weights are adaptively learned to better meet the requirements of different object detection and segmentation tasks. At the same time, its feature fusion and recombination mechanism further improves the quality and diversity of feature representation.

[0060] Compared with the PAN network, the specific improvements of MFAF in the solution of this case are as follows:

[0061] (1) Establish a bilateral fusion path: Establish a top-down semantic feature transfer path and a bottom-up spatial feature transfer path to fuse the high spatial resolution information of the shallow layer and the strong semantic information of the deep layer, and achieve higher-level feature fusion; at the same time, remove the nodes with only one-sided input, such as Figure 4 the two nodes missing in (b) compared with Figure 4 (a) in. Since these nodes only receive one input edge and lack feature fusion, their contribution to a network aiming to fuse different features is very small. Therefore, removing these nodes has little impact on the overall performance of the network, and it also helps to simplify the network structure and improve the calculation efficiency.

[0062] (2) Establish a skip connection path: If the original input node and the output node are at the same layer, to strengthen the feature correlation between these two types of nodes, an additional edge is added between them, such as Figure 4 the additional black dotted cross-scale connection added in (b) compared with Figure 4 (a) in, which realizes richer feature integration while keeping the cost controllable.

[0063] (3) Adopt a multi-scale feature adaptive weighted fusion method: By assigning different weights to each feature layer, the network adaptively learns the importance of different-level features, so as to better fuse features of different scales and improve the learning ability of the network. Among them, the fused feature can be expressed as:

[0064]

[0065] In formula (6), w i represents the weight, which can ensure that w i ≥0 under the action of the activation function ReLu. ε is a minimum value used to avoid numerical instability. F In represents the input feature, and F Out represents the result of weighted feature fusion.

[0066] Furthermore, based on the above method, an embodiment of the present invention also provides a ship target detection system based on multi-scale Gaussian context feature excitation, including: a model construction module and a target detection module, where,

[0067] The model construction module is used to construct a ship target detection model based on the YOLOv8 network and train the ship target detection model using remote sensing sample data. The ship target detection model includes a backbone network for feature extraction of the input remote sensing image, a neck network for fusing and enhancing the extracted features, and a detection head for detecting and outputting the target using the extracted features. The backbone network jointly performs Gaussian context conversion with spatial and channel attention mechanisms and realizes remote sensing image context feature excitation. The neck network uses a multi-scale feature adaptive fusion mechanism to obtain a feature map fused with remote sensing image target semantics and spatial position information;

[0068] The target detection module is used to input the remote sensing image to be detected into the trained ship target detection model and identify and extract the ship target in the remote sensing image to be detected using the ship target detection model.

[0069] To verify the effectiveness of the solution in this case, further explanation will be given below in combination with experimental data:

[0070] As Figure 5 shown, ship target detection experiments were carried out using the images provided by the public dataset HRSC2016. The image size was 300×300 - 1500×900 pixels, and the target resolution was 0.4 meters to 2 meters. By screening out 610 unlabeled images, converting the annotation format, and re-dividing the dataset, a training set including 512 images, a validation set, and a prediction set of 526 images were formed for the experiment. In addition, remote sensing images of ports in regions A, B, C, and D, etc., were downloaded through a public data platform. The image size was in the range of 800×500 to 8000×8000 pixels to verify the effectiveness and generalization of the solution in this case on satellite remote sensing ship target images in different scenarios.

[0071] Table 1 details the server configuration details and the specific configuration of the virtual environment used in the experiment.

[0072] Table 1 Experimental environment configuration

[0073]

[0074] The experiment uses the YOLOv8n network as the basic framework, and the parameter settings are shown in Table 2.

[0075] Table 2 Experimental parameter settings

[0076]

[0077]

[0078] The evaluation criteria widely recognized in the field of object detection, such as Average Precision (AP), mean Average Precision (mAP), Precision P, Recall R, and F1-score, etc., are used as the evaluation indicators for the experimental results.

[0079] Precision P: It reflects the proportion of true positive samples among all samples predicted as positive. The calculation formula is as follows:

[0080]

[0081] Recall R: It measures the proportion of the number of correctly predicted positive samples to the total number of actual positive samples. The calculation formula is as follows:

[0082]

[0083] Among them, TP represents the number of results correctly predicted as positive samples, FP represents the number of negative samples predicted as positive classes, and FN represents the number of positive samples predicted as negative classes.

[0084] The F1-score is the harmonic mean of precision and recall. The higher its value, the better the performance of the model.

[0085]

[0086] AP refers to the average precision of a single category, and mAP is the average of the APs of all categories. The calculation methods are as shown in equations (11) and (12), where m is the number of categories.

[0087]

[0088] mAP is specifically divided into mAP 50 and mAP50:95. mAP 50 refers to the mAP value when the Intersection of Union (IoU) is set to 0.5; mAP50:95 represents the mAP values at different IoU thresholds (from 0.5 to 0.95, with a step size of 0.05).

[0089] To verify the effectiveness and superiority of the solution in this case, multiple deep learning methods conducted on the HRSC2016 dataset were selected for comparison. Among them, REG-BB and SRBBS-FRCNN are ship category-level bounding box detection algorithms officially provided for the HRSC2016 dataset. The experimental results are shown in Table 3. It can be seen from Table 3 that the ship target detection method designed and introduced with the MFAF and SC-GCT mechanisms in YOLOv8 has the highest accuracy, and the mAP value reaches 96.8%, verifying the effectiveness of the solution in this case.

[0090] Table 3 Comparative experiments on ship target detection in the HRSC2016 dataset

[0091]

[0092]

[0093] Figure 6 and Figure 7 shows the visualization results of the solution in this case for ship target detection. As Figure 6 can be seen, some ship targets that were not detected by the original network framework of YOLOv8 were detected using the solution in this case. Thus, it can be known that compared with YOLOv8, the solution in this case enhances the detection ability of small-sized and dark-toned ship targets through the MFAF and SC-GCT modules, and alleviates the interference of complex backgrounds to a certain extent. Figure 7 The green boxes in the left figure are the true annotation target positions in the dataset. It can be seen that even in the face of complex scenarios where only part of the hull is shown and there is no annotation information, the algorithm can still achieve high-precision detection, proving its ability to learn generalizable features. Figure 8 shows the result image with poor detection effect of the solution in this case. It can be seen from the figure that in complex backgrounds (such as sea clutter, port debris, etc.), there are still missed detections and false detections in the algorithm, and further optimization is needed to improve the robustness.

[0094] To verify the effectiveness and superiority of the MFAF and SC-GCT mechanisms, two groups of comparative experiments were designed to verify their detection effects on ship targets on the HRSC2016 dataset respectively.

[0095] The CBAM and SC-GCT mechanisms were used as comparative experiments. CBAM is a convolutional attention mechanism module that activates features along two independent dimensions: channels and space. The experimental results are shown in Table 4. It can be seen from the table that after adding the SC-GCT mechanism to YOLOv8, the mAP50 accuracy reached the highest, at 96.2%, an increase of 0.7% compared to the original network accuracy, verifying the effectiveness of this mechanism. However, under strict evaluation index conditions, that is, mAP50:95 accuracy, the detection effects of the CBAM and SC-GCT methods decreased relative to the YOLOv8 network, but SC-GCT was still 3.3% more accurate than the CBAM method. After analysis, it is known that CBAM was proposed for ship target detection in natural scenes. In natural scenes, the target images are relatively large. Using global pooling to map the feature map dimension to 1×1 can still extract rich feature information. However, when expressing the features of small-sized targets such as ships in remote sensing images, more feature information will be lost, resulting in relatively poor target detection accuracy.

[0096] Table 4 Comparative experiments of SC-GCT on the HRSC2016 dataset

[0097]

[0098] To verify the effectiveness of the MFAF mechanism, YOLOv8 was selected as the baseline method and compared with multi-scale fusion methods such as CARAFE, ASFF, AFPN, and HS-FPN. The experimental results are shown in Table 5. The YOLOv8+MFAF method achieved the highest accuracy in mAP50 and mAP50:95, at 96.6% and 89.0% respectively, an increase of 1.1% and 1.5% compared to the YOLOv8 framework, indicating the advantages of MFAF in feature fusion and the effectiveness of the weighted fusion strategy. After analysis, it is known that the HRSC2016 dataset contains large-sized targets such as aircraft carriers, and the information in the p3 layer is rich. Adding the information in the p2 layer may cause information redundancy, resulting in a decrease in the target detection accuracy of YOLOv8+p2 compared to the original network. Since the YOLOv8+HS-FPN method is mainly used to solve the multi-scale problem in the white blood cell dataset in the medical field, and this experiment is for detecting ship targets in remote sensing images. Although both are small targets in their respective fields, the image size of ship targets is an order of magnitude larger than that of white blood cells. Therefore, this method is not very applicable to ship target detection. AFPN uses a progressive feature fusion method, that is, first fuses low-level high-spatial-resolution features, and then fuses high-level strong-semantic features. And subsequent feature operations pay more attention to the semantic features of the target while ignoring the spatial features. Therefore, the detection results of YOLOv8+AFPN are relatively poor.

[0099] Table 5 Comparative experiments of MFAF on the HRSC2016 dataset

[0100]

[0101] Remote sensing images of different scales are obtained using a public remote sensing image download platform. Images of about 1000×1000 pixels are defined as small-size images, and images larger than 2000×2000 pixels are defined as large-scale images. From the ship target detection results of small-size images (as shown in Table 6), it can be seen that using the solution of this case to detect targets in images of about 1000×1000 pixels takes about 50 ms, and the ship targets detected by the algorithm are basically consistent with the ship targets interpreted manually. Thus, it can be seen that the model trained on the HRSC2016 dataset through the solution of this case can be effectively used for ship target detection in other scenarios, verifying the practicability and generalization of the solution of this case.

[0102] Table 6 Detection Results of Small-Size Remote Sensing Images

[0103]

[0104] Using the solution of this case to detect ship targets in large-scale remote sensing images covering multiple types of ports obtained by downloading, the results are shown in Table 7. It can be seen from this that the highest precision rate of ship target detection in a single image can reach 100%, the recall rate reaches 94%, and the F1 score remains above 0.70. At the same time, the algorithm has a high detection efficiency. However, due to the influence of the number of targets, features, and background complexity, the detection and processing time for images of the same size (3840×2160 pixels) in different ports is about 1 s to 4 s. When the image size doubles to 7936×7424 pixels, the detection time also doubles to about 9 s.

[0105] Table 7 Detection Results of Large-Scale Remote Sensing Images

[0106]

[0107]

[0108] From the visualization results, in small-size remote sensing images, as Figure 9 shown, the ship targets account for a relatively large proportion in the image, and their geometric shapes, edge contours and other features are prominent, with high detection accuracy; in large-scale remote sensing images, as Figure 10 shown, combined with the analysis in Table 7, it can be obtained that the solution of this case shows strong detection ability when dealing with ship targets with weak feature responses and small pixel proportions, effectively overcoming the challenges faced by traditional methods when processing images with large image widths and low-resolution images. However, the detection results are still inevitably affected by factors such as occlusion between targets and similar features of targets, resulting in phenomena such as missed detection and misdetection. As Figure 10In the F Port Area scene shown in (f), within Area ①, the detection confidence of the five small boats docked at the port shore reached 90%, and small boats sailing at sea can also be accurately detected (Area ②). However, there are situations where the area composed of the bow and part of the hull of a large ship is misidentified as the entire ship (Area ③), and non-ship objects with similar color distributions, textures, and other visual characteristics (such as the jetty port in Area ④) are misdetected as ships. This phenomenon also indicates the research direction for further improving the performance of the detector in this case (resisting interference from similar backgrounds).

[0109] The above experimental data shows that compared with the reference method, the solution in this case has significant advantages in detection accuracy and efficiency on the HRSC2016 dataset, with the mAP value reaching 96.8%, and has good application prospects in the field of remote sensing image target detection.

[0110] Unless otherwise specifically stated, the relative steps, numerical expressions, and values of the components and steps described in these embodiments do not limit the scope of the present invention.

[0111] Each embodiment in this specification is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0112] The units and method steps of each example described in combination with the embodiments disclosed in this article can be implemented by electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described according to their functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Those of ordinary skill in the art can use different methods to implement the described functions for each specific application, but such implementation is not considered to exceed the scope of the present invention.

[0113] Those of ordinary skill in the art can understand that all or part of the steps in the above method can be completed by instructing relevant hardware through a program, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a magnetic disk, or an optical disc, etc. Optionally, all or part of the steps of the above embodiments can also be implemented using one or more integrated circuits. Correspondingly, each module / unit in the above embodiments can be implemented in the form of hardware or in the form of a software functional module. The present invention is not limited to any specific form of the combination of hardware and software.

[0114] Finally, it should be noted that the above-described embodiments are only specific embodiments of the present invention, which are used to illustrate the technical solutions of the present invention, rather than limiting it. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that any person skilled in the technical field of the present invention can still modify the technical solutions described in the foregoing embodiments, or can easily conceive of changes, or perform equivalent replacements on some of the technical features; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A ship target detection method based on multi-scale Gaussian context feature excitation, characterized in that Comprising: Construct a ship target detection model based on the YOLOv8 network, and train the ship target detection model using remote sensing sample data. Among them, the ship target detection model includes a backbone network for extracting features from the input remote sensing image, a neck network for fusing and enhancing the extracted features, and a detection head for detecting and outputting targets using the extracted features. The backbone network combines spatial and channel attention mechanisms to perform Gaussian context transformation and realize the excitation of remote sensing image context features. The neck network uses a multi-scale feature adaptive fusion mechanism to obtain a feature map fused with the semantic and spatial location information of the remote sensing image target; Input the remote sensing image to be detected into the trained ship target detection model, and use the ship target detection model to identify and extract the ship targets in the remote sensing image to be detected.

2. The ship target detection method based on multi-scale Gaussian context feature excitation according to claim 1, wherein, The backbone network includes: multiple convolutional layers that respectively perform convolutional operations on the input remote sensing image to extract features of different scales of the input image, a C2f component for performing feature transformation processing on the extracted remote sensing image features, a fast spatial pyramid pooling layer SPPF that enlarges the receptive field of the remote sensing image and pools the region of interest into a fixed size, and a Gaussian context transformation component that combines spatial and channel attention mechanisms to stimulate the context features of the remote sensing image.

3. The ship target detection method based on multi-scale Gaussian context feature excitation according to claim 1 or 2, characterized in that, Combining spatial and channel attention mechanisms to stimulate the context features of the remote sensing image includes: First, use the channel attention mechanism to assign weights to the channels of the extracted remote sensing image features, and obtain the intermediate features of the remote sensing image based on the feature channel weights; Then, fuse the intermediate features of the remote sensing image with the extracted remote sensing image features, and use the spatial attention mechanism to process the fused remote sensing image features to obtain the final activated features of the remote sensing image.

4. The ship target detection method based on multi-scale Gaussian context feature excitation according to claim 3, characterized in that Using the channel attention mechanism to assign weights to the channels of the extracted remote sensing image features includes: Performing global information aggregation on the extracted features in the spatial domain; Calculating the mean value of the global context features based on the aggregated global information, and performing normalization processing on the offset of the global context feature mean value using the global context standard deviation factor; Using the Gaussian function to transform and activate the mean value of the globally context features after normalization processing to obtain the activation value of the channel attention weight, so as to obtain the intermediate features of the remote sensing image by multiplying the activation value with the extracted remote sensing image features.

5. The ship target detection method based on multi-scale Gaussian context feature excitation according to claim 3, characterized in that Using the spatial attention mechanism to process the fused remote sensing image features includes: For the fused remote sensing image features, generate two feature maps through global maximum pooling operation and global average pooling operation respectively, and splice the two feature maps according to the number of channels; Use convolutional operation to reduce the dimension of the spliced feature map, and use the Sigmoid activation function to generate the spatial attention features of the reduced-dimension feature map, so as to multiply the spatial attention features with the fused remote sensing image features element by element and obtain the final activated features of the remote sensing image.

6. The ship target detection method based on multi-scale Gaussian context feature excitation according to claim 1, wherein The multi-scale feature adaptive fusion mechanism performs weighted fusion on features at different levels of remote sensing images through a bilateral fusion path and a skip connection path. Among them, the bilateral fusion path is used to fuse the spatial resolution feature information of the shallow layer and the semantic feature information of the deep layer, and the bilateral fusion path includes a top-down semantic feature transfer path and a bottom-up spatial feature transfer path; the skip connection path is used to strengthen the feature correlation between the input node and the output node of the network model at the same layer, and the skip connection path is an additional cross-scale connection path added between the network input node and the network output node at the same layer.

7. The ship target detection method based on multi-scale Gaussian context feature excitation according to claim 6, characterized in that The process of weighted fusion of different hierarchical features of remote sensing images is expressed as: where w i represents the weight of the i-th feature layer, ε is a minimum value used to avoid numerical instability, and F In represents the input feature, and F Out represents the feature result of weighted fusion.

8. A ship target detection system based on multi-scale Gaussian context feature excitation, characterized in that, It includes: a model construction module and an object detection module, where The model construction module is used to construct a ship object detection model based on the YOLOv8 network and train the ship object detection model using remote sensing sample data. Among them, the ship object detection model includes a backbone network for feature extraction of the input remote sensing image, a neck network for fusing and enhancing the extracted features, and a detection head for detecting and outputting the object using the extracted features. The backbone network combines the spatial and channel attention mechanisms to perform Gaussian context transformation and realize the excitation of the context features of the remote sensing image. The neck network uses the multi-scale feature adaptive fusion mechanism to obtain a feature map fused with the semantic and spatial position information of the remote sensing image target. The object detection module is used to input the remote sensing image to be detected into the trained ship object detection model, and use the ship object detection model to identify and extract the ship object in the remote sensing image to be detected.

9. An electronic device, characterized in that, It includes: At least one processor, and a memory coupled to the at least one processor; Among them, the memory stores a computer program, and the computer program can be executed by the at least one processor to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, A computer program is stored in the computer-readable storage medium, and when the computer program is executed, it can implement the method according to any one of claims 1 to 7.

Images