Small target ship identification method, device, electronic equipment and storage medium

Abstract

The present invention provides a method, device, electronic device, and storage medium for identifying small-target ships, belonging to the field of image recognition technology. The method comprises: obtaining an image to be identified, wherein the image includes a small-target ship to be identified; extracting a feature map of the small-target ship to be identified using a feature extraction module that combines a human vision simulation module and a channel attention mechanism module; and performing feature fusion on the feature map using a path aggregation network that combines the human vision simulation module and a low-level feature enhancement module to obtain the small-target ship to be identified in the image to be identified. The present invention achieves more accurate detection while ensuring a lightweight model, addressing the problem of insufficient small-target detection accuracy in current ship target detection.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a method, device, electronic equipment and storage medium for identifying a small target ship. Background Art

[0002] While deep learning-based object detection has achieved promising results for target classification and localization, missed and false detections of small targets remain a persistent challenge. This is particularly true in the field of ship detection, where ships vary greatly in geometric characteristics such as size and shape. Furthermore, the diverse camera angles, distances, wake waves, and occlusions in their environments contribute to these issues.

[0003] In the existing technology, in order to improve the accuracy and comprehensiveness of the detection of small ship targets, the number of small target samples is expanded through traditional data augmentation, and the number of small targets used in training is increased to improve the model's generalization ability for small targets. Although traditional data augmentation schemes can effectively increase the number of small targets in the data set, they may also introduce potential problems such as noise, which may adversely affect the training effect of the model. The enhanced data may not fully conform to the real-world distribution, which may increase the training time and computational cost of the model. In addition, data augmentation schemes are often difficult to solve the problem that small targets themselves have little information and difficult feature extraction.

[0004] It can be seen that the existing technology still has the technical problem of insufficient detection accuracy for the detection of small ship targets. Summary of the Invention

[0005] In view of this, it is necessary to provide a small target ship identification method, device, electronic equipment and storage medium to solve the technical problem that the existing technology still has insufficient detection accuracy for the detection of small target ships.

[0006] In order to solve the above technical problems, the present invention provides a small target ship identification method, comprising:

[0007] Acquiring an image to be identified, wherein the image to be identified includes a small target ship to be identified;

[0008] A feature extraction module combining a human vision simulation module and a channel attention mechanism module is used to extract a feature map of the small target ship to be identified;

[0009] The feature map is subjected to feature fusion using a path aggregation network that combines the human vision simulation module and the low-level feature enhancement module to obtain a small target ship to be identified in the image to be identified.

[0010] As a possible embodiment of the present invention, in this embodiment, the feature extraction module combining the human visual simulation module and the channel attention mechanism is used to extract the feature map of the small target ship to be identified in the image to be identified, including:

[0011] extracting a first feature vector of the image to be identified using a human visual simulation module;

[0012] Using a channel attention mechanism module to extract a second feature vector of the image to be identified;

[0013] The first feature vector and the second feature vector are fused to obtain a feature extraction vector, and the feature extraction vector is used to extract a feature map of the small target ship to be identified in the image to be identified.

[0014] As a possible implementation manner of the present invention, in this implementation manner, the human vision simulation module includes multiple parallel dilated convolution modules, and the extracting the first feature vector of the image to be recognized using the human vision simulation module includes:

[0015] Performing parallel dilation convolution on the image to be identified using multiple dilation rates to obtain multiple convolution feature vectors;

[0016] The multiple convolution vectors are fused using a preset adaptive fusion matrix, and the fused result is fused with a preset bias matrix to obtain the first eigenvector.

[0017] As a possible implementation manner of the present invention, in this implementation manner, extracting the second feature vector of the image to be recognized by using a channel attention mechanism module includes:

[0018] Performing global average pooling and global maximum pooling on the image to be identified to obtain a first pooling feature vector and a second pooling feature vector;

[0019] A preset linear rectification function is used to fuse the first pooled feature vector and the second pooled feature vector, and a convolution recovery function is used to restore the number of channels to obtain the second feature vector.

[0020] As a possible implementation manner of the present invention, in this implementation manner, the feature graph is subjected to feature fusion using a path aggregation network that combines the human vision simulation module and the low-level feature enhancement module to obtain the small target ship to be identified in the image to be identified, including:

[0021] Using the human vision simulation module and the low-level feature enhancement module to perform feature fusion on the low-level features and deep-level features in the feature map to obtain a feature fusion feature map;

[0022] A preset detection head is used to predict the feature fusion feature map to obtain the small target ship to be identified in the image to be identified.

[0023] As a possible implementation manner of the present invention, in this implementation manner, the preset detection head includes a regression task detection head and a classification task detection head, and the use of the preset detection head to predict the feature fusion feature map to obtain the small target ship to be identified in the image to be identified includes:

[0024] Calculating the regression loss function value of the feature fusion feature map using the regression task detection head;

[0025] Calculate the classification loss function value in the feature fusion feature graph using the classification task detection head;

[0026] The regression loss function value and the classification loss function value are weightedly calculated based on preset weights to obtain a target loss function value.

[0027] As a possible implementation manner of the present invention, in this implementation manner, the use of a preset detection head to predict the feature fusion feature map to obtain the small target ship to be identified in the image to be identified includes:

[0028] Determining classification information and positioning information of the small target ship to be identified in the image to be identified based on the target loss function value;

[0029] The position of the small target ship to be identified in the image to be identified and the type of the small target ship to be identified are determined based on the classification information and the positioning information.

[0030] On the other hand, the present invention also provides a small target ship identification device, comprising:

[0031] An image acquisition module is used to acquire an image to be identified, wherein the image to be identified includes a small target ship to be identified;

[0032] a feature extraction module, configured to extract a feature map of the small target ship to be identified by using a feature extraction module that combines a human visual simulation module and a channel attention mechanism module;

[0033] The feature fusion module is used to perform feature fusion on the feature map using a path aggregation network that combines the human visual simulation module and the low-level feature enhancement module to obtain the small target ship to be identified in the image to be identified.

[0034] On the other hand, the present invention also provides an electronic device, including a memory and a processor, wherein:

[0035] The memory is used to store programs;

[0036] The processor is coupled to the memory and is used to execute the program stored in the memory to implement the steps of the small target ship identification method described in any of the above implementations.

[0037] On the other hand, the present invention also provides a computer-readable storage medium for storing a computer-readable program or instruction, which, when executed by a processor, can implement the steps of the small target ship identification method described in any of the above implementations.

[0038] The beneficial effect of the present invention is as follows: the small target ship identification method provided by the present invention first extracts the feature map of the small target ship to be identified in the image to be identified through a feature extraction module that combines the human vision simulation module and the channel attention mechanism module, which can simulate the effect of different sizes of human visual receptive fields and improve the integrity of the target information extracted by the network. Then, the feature map is subjected to feature fusion by a path aggregation network that combines the human vision simulation module and the low-level feature enhancement module to obtain the small target ship to be identified in the image to be identified, and the low-level feature map is used to enrich the detail information and positioning information of the small target, and the richer positioning information is integrated into the deep feature layer, which not only retains the context information enhancement brought by the feature extraction, but also makes up for the lack of positioning information in the deep feature map, which is beneficial for the path aggregation network to obtain richer context information and effectively reduces the loss of detail information of the deep feature map caused by upsampling, while ensuring the lightweight of the model, achieving a more accurate detection effect, and solving the problem of insufficient accuracy of small target detection in current ship target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0040] Figure 1 A schematic flow chart of a small target ship identification method provided by an embodiment of the present invention;

[0041] Figure 2 A flowchart of a possible implementation method of S102 provided in an embodiment of the present invention;

[0042] Figure 3 A flowchart of a possible implementation method of S201 provided in an embodiment of the present invention;

[0043] Figure 4 An improvement provided by the embodiment of the present invention YOLOv 8. Schematic diagram of the model structure;

[0044] Figure 5 An embodiment of the present invention provides a HVS Schematic diagram of the module structure;

[0045] Figure 6 A flowchart of a possible implementation method of S202 provided in an embodiment of the present invention;

[0046] Figure 7 An embodiment of the present invention provides a VPAM Schematic diagram of the module structure;

[0047] Figure 8 A flowchart of a possible implementation method of S103 provided in an embodiment of the present invention;

[0048] Figure 9 A flowchart of a possible implementation method of S802 provided in an embodiment of the present invention;

[0049] Figure 10 A schematic diagram of a flow chart of a small target ship positioning method provided by an embodiment of the present invention;

[0050] Figure 11 A schematic structural diagram of a small target ship identification device provided by an embodiment of the present invention;

[0051] Figure 12 A schematic structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0053] The terms "first," "second," and so on, used in the embodiments of the present invention are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, technical features designated as "first" or "second" may explicitly or implicitly include at least one such feature.

[0054] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute a separate or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0055] The present invention provides a small target ship identification method, device, electronic equipment and storage medium, which are described below respectively.

[0056] Figure 1 A flow chart of an embodiment of the small target ship identification method provided by the present invention is shown as follows: Figure 1 As shown, the small target ship identification method includes:

[0057] S101, obtaining an image to be identified, where the image to be identified includes a small target ship to be identified;

[0058] S102, extracting a feature map of the small target ship to be identified using a feature extraction module that combines a human vision simulation module and a channel attention mechanism module;

[0059] S103, performing feature fusion on the feature map using a path aggregation network that combines a human vision simulation module and a low-level feature enhancement module to obtain a small target ship to be identified in the image to be identified.

[0060] In the embodiment of the present invention, the small target ship recognition method can be applied to ship recognition in image information such as pictures and videos, and has good recognition accuracy, especially for the recognition of small target ships. Among them, the small target ship generally does not refer to the target ship to be recognized as a small size, but refers to the target ship to be recognized as a distant target in the image to be recognized, and therefore occupies a small area in the image to be recognized, such as the target ship to be recognized that only occupies one grid or several grids when the image to be recognized is grid-divided. Optionally, the small target ship recognition method provided by the embodiment of the present invention can be based on the improved YOLOv 8 model is realized, and the specific improvement method is described later in this invention.

[0061] In the embodiment of the present invention, the image to be identified refers to image information containing the small target ship to be identified, such as a picture or video of the small target ship.

[0062] In the embodiment of the present invention, after adopting the improved YOLOvWhen the 8-model recognizes the image to be recognized, the feature extraction module combines the Human Visual Simulation Module (HVS) and the Channel Attention Mechanism (CAM). Among them, the HVS designs parallel void convolution to simulate the different sizes of human visual receptive fields, and at the same time simulates the eccentricity of human vision for receptive fields of different sizes through an adaptive linear fusion matrix, helping the backbone network to extract more complete small ship target features. Secondly, in order to increase the degree of attention to small target feature information and improve the scale adaptability of the model, the present invention will HVS and CAM Combined with the design of the visual perception attention mechanism module ( Visual Perceptual Attention Mechanism , VPAM ), added to the backbone network to better focus on the details of small targets.

[0063] In the embodiment of the present invention, for the feature map of the small target ship to be identified, in order to take into account the low-level feature map containing more detailed information, a combination of HVS and low-level feature enhancement modules ( Low-level Feature Enhancement , LFE )’s path aggregation network fuses the feature maps, since HVS It does not affect the size of the feature map, so it will not affect the subsequent modules, which is beneficial for the feature fusion network to obtain richer context information and effectively reduces the loss of deep feature map detail information caused by upsampling. LFE The feature information of the feature map is fused to utilize the low-level feature map to enrich the detail information and positioning information of small targets, thereby ensuring the accuracy of small target ship recognition.

[0064] As a possible embodiment of the present invention, in this embodiment, Figure 2 As shown in FIG, a feature extraction module combining a human visual simulation module and a channel attention mechanism is used to extract a feature map of a small target ship to be identified in an image to be identified, including:

[0065] S201, extracting a first feature vector of the image to be recognized using a human visual simulation module;

[0066] S202, using a channel attention mechanism module to extract a second feature vector of the image to be recognized;

[0067] S203 , fusing the first eigenvector and the second eigenvector to obtain a feature extraction vector, and using the feature extraction vector to extract a feature map of the small target ship to be identified in the image to be identified.

[0068] In the embodiment of the present invention, in order to ensure the comprehensiveness of feature extraction, HVS and C AM Extract the first eigenvector and the second eigenvector of the image to be identified, refer to the receptive field mechanism in the human visual perception system, and introduce parallel hole convolution to simulate the receptive field of the human eye, such as Figure 3 As shown, in S201, it includes:

[0069] S301, performing parallel dilation convolution on the image to be recognized using multiple dilation rates to obtain multiple convolution feature vectors;

[0070] S302: fuse multiple convolution vectors using a preset adaptive fusion matrix, and fuse the fusion result with a preset bias matrix to obtain a first eigenvector.

[0071] In the embodiment of the present invention, for the convenience of explanation, Figure 4 As shown, the improved YOLOv8 Model as an example, in YOLOv 8 backbone networks joined HVS and CAM Combined VPAM Module, the input image size in the backbone network of the model is 640*640*3, YOLOv The backbone extraction network of 8 can obtain five effective feature maps of different scales, with sizes of C 1 (320*320*32), C 2 (160*160*64), C 3 (80*80*128), C 4 (40*40*256), C 5 (20*20*512), in the feature map C 1. C 2. C 3. Add the present invention VPAM Modules, specifically, Figure 5 As shown, assuming HVS The input picture of the module is X (H*W*C), for pictures X Void rate d =1, 2, 3, 4 parallel void convolutions, and select a 3*5 size convolution kernel to better adapt to the aspect ratio of the ship, use a convolution kernel with a depth of C / 4 to compress the number of channels to 1 / 4 of the input, and then pass Concat Restore to C , this process does not introduce too many parameters, ensuring the lightweight of parameters. Furthermore, in order to simulate the eccentricity of different receptive fields in human vision, the feature information extracted under different scale receptive fields is fused through the matrix. The size of the matrix is H*W*C, the parameters of the matrix are adaptively adjusted during back propagation to help the model find the best matrix weights suitable for the fusion of receptive fields of different sizes, and finally aggregate the information to the output Y The adaptive fusion formula of features under different scale receptive fields is shown in formula (1):

[0072] (1)

[0073] in, Concat Represents a splicing operation, f represents the adaptive matrix fusion operation, P represents the bias matrix, Conv For convolution operation, after feature fusion, 1×1 convolution is used to compress the number of channels to 1, and then Softmax Function normalization obtains the output first eigenvector of the HVS module , the calculation formula is shown in formula (2):

[0074] (2)

[0075] Further, such as Figure 6 As shown, in step S202, it includes:

[0076] S601, performing global average pooling and global maximum pooling on the image to be recognized to obtain a first pooling feature vector and a second pooling feature vector;

[0077] S602: Using a preset linear rectification function to fuse the first pooled feature vector and the second pooled feature vector, and using a convolution recovery function to restore the number of channels to obtain a second feature vector.

[0078] In the embodiment of the present invention, for CAM Module, assuming the input feature map is X ( H*W*C ), CAM First, global average pooling and global maximum pooling are performed on it in the spatial dimension to obtain two (1*1*C) feature maps. Then, 1*1 convolution is used to replace the original multi-layer perceptron network. The number of channels is first reduced by a 1*1 convolution, and then ReLU Activation function, and finally a 1*1 convolution to restore the number of channels. The output results of the two feature maps are added together, and the Sigmoid The activation function is normalized, and finally we get CAM The output of the module (1*1*C) is calculated as shown in formula (3):

[0079] (3)

[0080] in, is the second eigenvector, Sigmoid is the activation function, MaxPool is the global maximum pooling, AvgPool is global average pooling.

[0081] In the embodiment of the present invention, Figure 7 As shown, CAM Modules and HVS The modules are connected in parallel to reduce the loss of feature information in the network forward propagation. Specifically, first CAM The results obtained Multiply X Get the output , and will pass through HVS The results obtained by the module Multiply X , and get the output , and finally and Fusion, to obtain feature information that is rich in receptive field information of different scales and more focused on beneficial targets Finally, the correlation feature weights are applied to X On the above, we get the output feature map Y The calculation formulas are shown in formulas (4) to (7):

[0082] (4)

[0083] (5)

[0084] (6)

[0085] (7)

[0086] The improved backbone feature extraction network provided by the embodiment of the present invention can connect and fuse contextual information based on the human visual receptive field mechanism, and give small targets more attention weights. Based on this structure, the feature information of small targets can be extracted more accurately and comprehensively.

[0087] As a possible embodiment of the present invention, in this embodiment, Figure 8 As shown in the figure, the feature map is fused using a path aggregation network that combines a human visual simulation module and a low-level feature enhancement module to obtain the small target ship to be identified in the image to be identified, including:

[0088] S801, using a human vision simulation module and a low-level feature enhancement module to fuse low-level features and deep-level features in the feature map to obtain a feature fusion feature map;

[0089] S802: Use a preset detection head to predict the feature fusion feature map to obtain a small target ship to be identified in the image to be identified.

[0090] In the embodiment of the present invention, Figure 4 As shown in the figure, during feature fusion, the feature map F4 is processed by upsampling and backbone network extraction and is compared with the feature map C3. Concat Operation, get the feature map F3, similarly, PAN ( Path Aggregation Network In the path aggregation network structure, a total of six effective feature maps are obtained, namely F5 (20*20*512), F4 (40*40*256), F3 (80*80*128), P3 (80*80*128), P4 (40*40*256), and P5 (20*20*512). The low-level positioning enhancement module designed by the present invention is added between the feature maps C2 and P3. LFE , which integrates richer positioning information into the deep feature layer, retaining the VPAM The context information enhancement brought by it makes up for the lack of positioning information in the deep feature map. LFE Module, since the size of feature maps C2 (160*160*64) and P3 (80*80*128) are inconsistent, firstly, the two feature maps are reset to 80*80*64 feature maps by 1*1 convolution and 3*3 convolution respectively, and then they are Concat Then, adaptive linear transformation is used to perform adaptive feature fusion to form a feature layer with enhanced semantic and positioning information. The calculation formula is shown in formula (8):

[0091] (8)

[0092] Furthermore, between the feature maps C5 and F5, add HVS module, due to HVS The module will not affect the size of the feature map, so it will not affect the subsequent modules. HVS As the first module after the backbone feature extraction network, the module helps the feature fusion network obtain richer context information and effectively reduces the loss of detail information in deep feature maps caused by upsampling.

[0093] In the embodiment of the present invention, the improved neck network in the above embodiment will finally obtain three effective feature maps, which are (80*80*128), P4 (40*40*256), and P5 (20*20*512). The information of these three feature maps detects targets at different scales, and the three feature maps are input into the three preset detection heads respectively to obtain the small target ship to be identified in the image to be identified.

[0094] As a possible implementation of the present invention, in this implementation, the preset detection head includes a regression task detection head and a classification task detection head, such as Figure 9 As shown, a preset detection head is used to predict the feature fusion feature map to obtain the small target ship to be identified in the image to be identified, including:

[0095] S901, using the regression task detection head to calculate the regression loss function value of the feature fusion feature map;

[0096] S902, using the classification task detection head to calculate the classification loss function value in the feature fusion feature graph;

[0097] S903: Perform weighted calculation on the regression loss function value and the classification loss function value based on preset weights to obtain a target loss function value.

[0098] In the embodiment of the present invention, each detection head has two parallel branches, which respectively calculate the loss of the classification task and the localization task. For the classification task, the output is pushed into Softmax In the function, the calculation formula is shown in formula (9):

[0099] (9)

[0100] in, x For input, W is the weight, is the output of the neural network, y is the true distribution;

[0101] For the regression task, a Gaussian function is defined and the calculation formula is shown in formula (10):

[0102] (10)

[0103] in, is a noise value, N represents the Gaussian probability distribution function.

[0104] Assume that the output of the model is and , its conditional probability is shown in formula (11):

[0105] (11)

[0106] Substituting formula (9) and formula (10) into formula (11), and using the maximum likelihood estimation method for formula (11), the result is shown in formula (12):

[0107] (12)

[0108] in, is the loss function value of the regression task, is the loss function value of the classification task, The optimized target loss function is shown in formula (13):

[0109] (13)

[0110] in, is the target loss function, is the loss function for the classification task, is the loss function value of the classification task, and is a hyperparameter.

[0111] In an embodiment of the present invention, the weights of the regression loss function and the classification loss function can be automatically learned simultaneously in model learning. For example, an improved YOLOv8 model is constructed based on the modules adopted in the aforementioned embodiment, and the model is trained based on a preset small target ship dataset, including the following steps:

[0112] Step 1: Convert the format of the small target ship dataset and generate the path files corresponding to the training set and test set samples according to the ratio of training set: test set equal to 9:1.

[0113] Step 2: Set the model's pre-trained weight file to yolov 8_ s.pth .

[0114] Step 3: Set the initial momentum of the model to 0.937, the initial learning rate to 1E-2, and the learning rate decrease method to cos , the gradient descent method is stochastic gradient descent, choose to use Mosaic 、 Mixup and Focus Data enhancement method.

[0115] Step 4, the training phase, uses the multi-task loss function introduced in the previous embodiment to train the model for 100 epochs to obtain the optimal weight file trained by the model.

[0116] The embodiment of the present invention sets weights for different tasks to ensure the accuracy of small target ship identification, and the weights are automatically trained during model learning, which makes the system more applicable to small target ship identification.

[0117] As a possible embodiment of the present invention, in this embodiment, Figure 10 As shown, a preset detection head is used to predict the feature fusion feature map to obtain the small target ship to be identified in the image to be identified, including:

[0118] S1001, determining classification information and positioning information of a small target ship to be identified in the image to be identified based on the target loss function value;

[0119] S1002: Determine the position of the small target ship to be identified in the image to be identified and the type of the small target ship to be identified based on the classification information and the positioning information.

[0120] In the embodiment of the present invention, the trained optimal weight file is used to predict the small target ship in the image to be identified. For example, the image to be identified is passed through the backbone network to generate five valid feature maps, namely C1 (320*320*32), C2 (160*160*64), C3 (80*80*128), C4 (40*40*256), and C5 (20*20*512). The C2, C3, C4, and C5 feature maps are sent to the neck feature fusion network; in the neck network, the feature map C5 is passed through the backbone network. HVS The module obtains the feature map F5, and the feature map C2 is fused with P3 and then LFE The module obtains the feature map ; In addition, through splicing and other operations, feature maps F3, F4, P4, and P5 are obtained. , P4 and P5 are sent to the detection head to calculate the loss value, and finally the classification and positioning information of the target is obtained. The classification information and positioning information are post-processed, including but not limited to non-maximum suppression operations, and the positioning information is mapped to the image to be identified to mark the positioning and classification information of the ship, etc., and finally the position and type of the small target ship in the image to be identified are identified.

[0121] In order to better implement the small target ship identification method in the embodiment of the present invention, based on the small target ship identification method, correspondingly, Figure 11 As shown, an embodiment of the present invention further provides a small target ship identification device, and the small target ship identification device 1100 includes:

[0122] The image acquisition module 1101 is used to acquire an image to be identified, wherein the image to be identified includes a small target ship to be identified;

[0123] A feature extraction module 1102 is configured to extract a feature image of a small target ship to be identified by using a feature extraction module that combines a human vision simulation module and a channel attention mechanism module;

[0124] The feature fusion module 1103 is used to perform feature fusion on the feature map using a path aggregation network that combines a human visual simulation module and a low-level feature enhancement module to obtain a small target ship to be identified in the image to be identified.

[0125] The small target ship identification device 1100 provided in the above embodiment can implement the technical solution described in the above small target ship identification method embodiment. The specific implementation principles of the above modules or units can refer to the corresponding contents in the above small target ship identification method embodiment, which will not be repeated here.

[0126] The small target ship recognition device provided by the present invention first extracts a feature map of a small target ship to be identified in an image to be identified by using a feature extraction module that combines a human vision simulation module and a channel attention mechanism module. This can simulate the effect of different sizes of human visual receptive fields, and at the same time simulate the eccentricity of human vision for receptive fields of different sizes. Combining the two, the effect of simulating the human visual receptive field is finally achieved, thereby improving the integrity of target information extracted by the network. Then, the feature map is subjected to feature fusion using a path aggregation network that combines a human vision simulation module and a low-level feature enhancement module to obtain the small target ship to be identified in the image to be identified. The low-level feature map is used to enrich the detail information and positioning information of the small target, and the richer positioning information is integrated into the deep feature layer. This not only retains the context information enhancement brought by the feature extraction, but also makes up for the lack of positioning information in the deep feature map, which is conducive to the feature fusion network obtaining richer context information and effectively reduces the loss of detail information of the deep feature map caused by upsampling. While ensuring the lightweight model, a more accurate detection effect is achieved, thereby solving the problem of insufficient detection accuracy of small targets in current ship target detection.

[0127] like Figure 12 As shown, the present invention also provides an electronic device 1200. The electronic device 1200 includes a processor 1201, a memory 1202 and a display 1203. Figure 12 Only some of the components of the electronic device 1200 are shown, but it should be understood that it is not required to implement all of the shown components, and more or fewer components may be implemented instead.

[0128] In some embodiments, the processor 1201 may be a central processing unit (CPU), a microprocessor, or other data processing chip, configured to execute program codes or process data stored in the memory 1202, such as the small target ship identification method of the present invention.

[0129] In some embodiments, processor 1201 may be a single server or a server group. The server group may be centralized or distributed. In some embodiments, processor 1201 may be local or remote. In some embodiments, processor 1201 may be implemented on a cloud platform. In one embodiment, the cloud platform may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, multiple clouds, or any combination thereof.

[0130] In some embodiments, the memory 1202 may be an internal storage unit of the electronic device 1200, such as a hard disk or memory of the electronic device 1200. In other embodiments, the memory 1202 may also be an external storage device of the electronic device 1200, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the electronic device 1200.

[0131] Furthermore, the memory 1202 may include both an internal storage unit of the electronic device 1200 and an external storage device. The memory 1202 is used to store application software installed on the electronic device 1200 and various data.

[0132] In some embodiments, display 1203 can be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 1203 is used to display information about electronic device 1200 and to display a visual user interface. Components 1201-1203 of electronic device 1200 communicate with each other via a system bus.

[0133] In one embodiment, when the processor 1201 executes the small target ship identification program in the memory 1202, the following steps may be implemented:

[0134] Acquire an image to be identified, where the image to be identified includes a small target ship to be identified;

[0135] A feature extraction module that combines a human vision simulation module and a channel attention mechanism module is used to extract the feature map of the small target ship to be identified in the image to be identified;

[0136] The feature map is fused using a path aggregation network that combines a human visual simulation module and a low-level feature enhancement module to obtain the small target ship to be identified in the image to be identified.

[0137] It should be understood that, when the processor 1201 executes the small target ship identification program in the memory 1202 , in addition to the above functions, it can also implement other functions. For details, please refer to the description of the corresponding method embodiment above.

[0138] Furthermore, the embodiment of the present invention does not specifically limit the type of the electronic device 1200 mentioned. The electronic device 1200 may be a portable electronic device such as a mobile phone, a tablet computer, a personal digital assistant (PDA), a wearable device, a laptop computer, or the like. Exemplary embodiments of portable electronic devices include, but are not limited to, portable electronic devices equipped with IOS, Android, Microsoft, or other operating systems. The above-mentioned portable electronic devices may also be other portable electronic devices, such as a laptop computer with a touch-sensitive surface (e.g., a touch panel). It should also be understood that in some other embodiments of the present invention, the electronic device 1200 may not be a portable electronic device, but a desktop computer with a touch-sensitive surface (e.g., a touch panel).

[0139] Accordingly, an embodiment of the present invention also provides a computer-readable storage medium, which is used to store computer-readable programs or instructions. When the program or instructions are executed by a processor, it can implement the steps or functions of the small target ship identification method provided in the above-mentioned method embodiments.

[0140] Those skilled in the art will appreciate that all or part of the process flow of the above-described method embodiment can be implemented by instructing related hardware (such as a processor, controller, etc.) through a computer program, and the computer program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a magnetic disk, an optical disk, a read-only memory, or a random access memory.

[0141] The above is a detailed introduction to the small target ship identification method, device, electronic device and storage medium provided by the present invention. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A small target ship identification method, characterized in that: include: Acquiring an image to be identified, wherein the image to be identified includes a small target ship to be identified; extracting a first feature vector of the image to be identified using a human visual simulation module; Using a channel attention mechanism module to extract a second feature vector of the image to be identified; fusing the first feature vector and the second feature vector to obtain a feature extraction vector, and using the feature extraction vector to extract a feature map of the small target ship to be identified in the image to be identified; The human vision simulation module includes a plurality of parallel dilated convolution modules, and the step of extracting the first feature vector of the image to be identified by using the human vision simulation module includes: Performing parallel dilation convolution on the image to be identified using multiple dilation rates to obtain multiple convolution feature vectors; fusing the plurality of convolution feature vectors using a preset adaptive fusion matrix, and fusing the fused structure with a preset bias matrix to obtain the first feature vector; The feature map is subjected to feature fusion using a path aggregation network that combines the human vision simulation module and the low-level feature enhancement module to obtain a small target ship to be identified in the image to be identified.

2. The small target ship identification method according to claim 1, characterized in that: The step of extracting the second feature vector of the image to be identified by using a channel attention mechanism module includes: Performing global average pooling and global maximum pooling on the image to be identified to obtain a first pooling feature vector and a second pooling feature vector; A preset linear rectification function is used to fuse the first pooled feature vector and the second pooled feature vector, and a convolution recovery function is used to restore the number of channels to obtain the second feature vector.

3. The small target ship identification method according to claim 1, characterized in that: The method of performing feature fusion on the feature map using a path aggregation network that combines the human vision simulation module and the low-level feature enhancement module to obtain a small target ship to be identified in the image to be identified includes: Using the human vision simulation module and the low-level feature enhancement module to perform feature fusion on the low-level features and deep-level features in the feature map to obtain a feature fusion feature map; A preset detection head is used to predict the feature fusion feature map to obtain the small target ship to be identified in the image to be identified.

4. The small target ship identification method according to claim 3, characterized in that: The preset detection head includes a regression task detection head and a classification task detection head. The preset detection head is used to predict the feature fusion feature map to obtain the small target ship to be identified in the image to be identified, including: Calculating the regression loss function value of the feature fusion feature map using the regression task detection head; Calculate the classification loss function value in the feature fusion feature graph using the classification task detection head; The regression loss function value and the classification loss function value are weightedly calculated based on preset weights to obtain a target loss function value.

5. The small target ship identification method according to claim 4, characterized in that: The method of using a preset detection head to predict the feature fusion feature map to obtain the small target ship to be identified in the image to be identified includes: Determining classification information and positioning information of the small target ship to be identified in the image to be identified based on the target loss function value; The position of the small target ship to be identified in the image to be identified and the type of the small target ship to be identified are determined based on the classification information and the positioning information.

6. A small target ship identification device, characterized in that: include: An image acquisition module is used to acquire an image to be identified, wherein the image to be identified includes a small target ship to be identified; A feature extraction module, configured to extract a first feature vector of the image to be identified using a human visual simulation module; Using a channel attention mechanism module to extract a second feature vector of the image to be identified; fusing the first feature vector and the second feature vector to obtain a feature extraction vector, and using the feature extraction vector to extract a feature map of the small target ship to be identified in the image to be identified; The human vision simulation module includes a plurality of parallel dilated convolution modules, and the step of extracting the first feature vector of the image to be identified by using the human vision simulation module includes: Performing parallel dilation convolution on the image to be identified using multiple dilation rates to obtain multiple convolution feature vectors; fusing the plurality of convolution feature vectors using a preset adaptive fusion matrix, and fusing the fused structure with a preset bias matrix to obtain the first feature vector; The feature fusion module is used to perform feature fusion on the feature map using a path aggregation network that combines the human visual simulation module and the low-level feature enhancement module to obtain the small target ship to be identified in the image to be identified.

7. An electronic device, characterized in that: comprising a memory and a processor, wherein, The memory is used to store programs; The processor is coupled to the memory and is used to execute the program stored in the memory to implement the steps in the small target ship identification method according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that Used to store computer-readable programs or instructions, which, when executed by a processor, can implement the steps of the small target ship identification method described in any one of claims 1 to 5 above.

Images