Image target detection method, device, equipment and storage medium

Abstract

Embodiments of the present application provide an image target detection method, device and equipment, and a storage medium, and relate to the technical field of artificial intelligence. The image target detection method comprises the following steps: obtaining a to-be-detected image; using an encoding module to extract features of the to-be-detected image to obtain at least one encoding feature map; using a decoding module to densely connect the at least one encoding feature map to obtain a decoding feature map; using a feature enhancement model to perform feature pyramid transformation on the decoding feature map to obtain an enhanced feature map; and using a detection head prediction model to predict the enhanced feature map to obtain a target detection result. In this embodiment, the encoding feature map of the to-be-detected image is densely connected by using an encoding and decoding module to obtain a decoding feature map, the feature extraction capability is enhanced, the decoding feature map is enhanced by using feature pyramid transformation, the detection capability of the subsequent detection head prediction model for multi-scale targets is effectively improved, the detection performance of the SAR target image detection algorithm is improved, and the problem of tight operation capability is alleviated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to an image target detection method and device, equipment and a storage medium. BACKGROUND

[0002] Synthetic Aperture Radar (SAR) is an active microwave remote sensing imaging radar that can observe the earth surface all day long, so SAR plays an important role in ocean monitoring and maritime traffic supervision.

[0003] In related technologies, target detection is performed in a SAR image. On the basis of deep learning-based target detection technology, SAR image target detection also introduces a deep learning-based target detection method. Although deep learning-based SAR image target detection achieves good results, however, due to the different sizes of the detected targets in actual application scenarios, the detection of multi-scale targets still has problems of tight computation and low performance. SUMMARY

[0004] The main purpose of the embodiments of the present application is to provide an image target detection method, device, equipment and storage medium, improve the detection performance of the SAR target image detection algorithm, and alleviate the problem of tight computation.

[0005] To achieve the above purpose, a first aspect of the embodiments of the present application provides an image target detection method, comprising:

[0006] obtaining a to-be-detected image and inputting the to-be-detected image into a trained image detection model, the image detection model comprising: a feature extraction model, a feature enhancement model and a detection head prediction model, the feature extraction model comprising: an encoding module and a decoding module;

[0007] extracting at least one encoding feature map from the to-be-detected image by using the encoding module;

[0008] densely connecting the at least one encoding feature map by using the decoding module to obtain a decoding feature map;

[0009] performing feature pyramid transformation on the decoding feature map by using the feature enhancement model to obtain an enhanced feature map;

[0010] predicting the enhanced feature map by using the detection head prediction model to obtain a target detection result.

[0011] In some embodiments, densely connecting the at least one encoding feature map by using the decoding module to obtain a decoding feature map comprises:

[0012] perform feature map jumping on the at least one encoded feature map according to a dimension of the first encoded feature map to obtain a first decoded feature map;

[0013] perform feature map jumping on the first decoded feature map and the at least one encoded feature map according to a dimension of the second decoded feature map to obtain a second decoded feature map;

[0014] perform feature map jumping on the second decoded feature map and the at least one encoded feature map according to a dimension of the encoded feature map to obtain a decoded feature map.

[0015] In some embodiments, the at least one feature map includes a first encoded feature map, a second encoded feature map, a third encoded feature map, and a fourth encoded feature map, and performing feature map jumping on the at least one encoded feature map according to a dimension of the first encoded feature map to obtain a first decoded feature map includes:

[0016] down-sampling the first encoded feature map to obtain a first down-sampling result;

[0017] down-sampling the second encoded feature map to obtain a second down-sampling result;

[0018] deconvolving the fourth encoded feature map to obtain a fourth deconvolution result;

[0019] multiplying each weight parameter in a preset first weight parameter set with a corresponding first down-sampling result, third encoded feature map, second down-sampling result, and fourth deconvolution result respectively and then adding them to obtain the first decoded feature map.

[0020] In some embodiments, performing feature map jumping on the first decoded feature map and the at least one encoded feature map according to a dimension of the second decoded feature map to obtain a second decoded feature map includes:

[0021] deconvolving the first decoded feature map to obtain a first deconvolution result;

[0022] up-sampling the third encoded feature map to obtain a third up-sampling result;

[0023] multiplying each weight parameter in a preset second weight parameter set with a corresponding first deconvolution result, second encoded feature map, first down-sampling result, and third up-sampling result respectively and then adding them to obtain the second decoded feature map.

[0024] In some embodiments, performing feature map jumping on the second decoded feature map and the at least one encoded feature map according to a dimension of the encoded feature map to obtain a decoded feature map includes:

[0025] deconvolving the second decoded feature map to obtain a second deconvolution result;

[0026] The second encoded feature map is up-sampled to obtain a second up-sampling result;

[0027] The decoding feature map is obtained by multiplying each weight parameter in the preset third weight parameter set with the corresponding second deconvolution result, first encoded feature map, second up-sampling result and third up-sampling result and then adding them.

[0028] In some embodiments, the feature pyramid transformation is performed on the decoding feature map by using the feature enhancement model to obtain an enhanced feature map, including:

[0029] The decoding feature map is compressed to obtain at least one compressed feature map;

[0030] The at least one compressed feature map is convolved to obtain at least one compressed convolution feature map;

[0031] According to the dimension of the enhanced feature map, the at least one compressed convolution feature map is aligned and up-sampled to obtain the enhanced feature map.

[0032] In some embodiments, the detection head prediction model includes a target center position regression prediction module, a target center point offset regression prediction module and a target size regression prediction model, and the target detection result is obtained by predicting the enhanced feature map by using the detection head prediction model, including:

[0033] The target center position output value is obtained by predicting the enhanced feature map by using the target center position regression prediction module;

[0034] The target center point offset output value is obtained by predicting the enhanced feature map by using the target center point offset regression prediction module;

[0035] The target size output value is obtained by predicting the enhanced feature map by using the target size regression prediction model;

[0036] According to the comparison result of the target center position output value and the preset confidence threshold, the detection target is obtained, and the target detection result is the target center position output value, target center point offset output value and target size output value corresponding to the detection target.

[0037] To achieve the above-mentioned purpose, a second aspect of the present application is an image detection device, including:

[0038] An image acquisition unit is configured to acquire a to-be-detected image and input the to-be-detected image into a trained image detection model, wherein the image detection model includes a feature extraction model, a feature enhancement model and a detection head prediction model, and the feature extraction model includes an encoding module and a decoding module.

[0039] An encoding unit is configured to extract features of the to-be-detected image by using the encoding module to obtain at least one encoded feature map.

[0040] a decoding unit configured to densely connect the at least one encoded feature map by using a decoding module to obtain a decoded feature map;

[0041] a feature enhancement unit configured to perform feature pyramid transformation on the decoded feature map by using a feature enhancement model to obtain an enhanced feature map;

[0042] a detection unit configured to predict the enhanced feature map by using a detection head prediction model to obtain a target detection result.

[0043] To achieve the above object, the third aspect of the present application provides an electronic device, comprising:

[0044] at least one memory;

[0045] at least one processor;

[0046] at least one program;

[0047] The program is stored in the memory, and the processor executes the at least one program to implement the method of the first aspect of the present application.

[0048] To achieve the above object, the fourth aspect of the present application provides a storage medium, which is a computer readable storage medium, and the computer readable storage medium stores computer executable instructions, and the computer executable instructions are used to make a computer execute:

[0049] the method of the first aspect.

[0050] The image target detection method, device, equipment and storage medium provided by the embodiment of the present application, by acquiring the to-be-detected image and inputting the to-be-detected image into the trained image detection model, using the encoding module to extract features of the to-be-detected image to obtain at least one encoded feature map, using the decoding module to densely connect the at least one encoded feature map to obtain a decoded feature map, using the feature enhancement model to perform feature pyramid transformation on the decoded feature map to obtain an enhanced feature map, and using the detection head prediction model to predict the enhanced feature map to obtain a target detection result. The embodiment of the present application densely connects the encoded feature map of the to-be-detected image by using the encoding and decoding module to obtain a decoded feature map, enhances the feature extraction capability, and then uses the feature pyramid transformation to enhance the features of the decoded feature map, effectively improves the detection capability of the subsequent detection head prediction model on the multi-scale target, improves the detection performance of the SAR target image detection algorithm, and relieves the problem of tight operation capability. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 is a flowchart of the image target detection method provided by the embodiment of the present application.

[0052] Figure 2The figure detection model schematic diagram of the image target detection method provided by another embodiment of the present application.

[0053] Figure 3 The flow chart of the image target detection method provided by another embodiment of the present application.

[0054] Figure 4 The flow chart of the image target detection method provided by another embodiment of the present application.

[0055] Figure 5 The flow chart of the image target detection method provided by another embodiment of the present application.

[0056] Figure 6 The flow chart of the image target detection method provided by another embodiment of the present application.

[0057] Figure 7 The flow chart of the image target detection method provided by another embodiment of the present application.

[0058] Figure 8 The flow chart of the image target detection method provided by another embodiment of the present application.

[0059] Figure 9 The principle schematic diagram of the image target detection method provided by another embodiment of the present application.

[0060] Figure 10 The structure block diagram of the image detection device provided by an embodiment of the present application.

[0061] Figure 11 The hardware structure schematic diagram of the electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0062] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not intended to limit the present application.

[0063] It should be noted that although the functional modules are divided in the device schematic diagram, and the logical order is shown in the flow chart, in some cases, the steps shown or described can be executed in a manner different from the module division in the device or the order in the flow chart.

[0064] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application, and are not intended to limit the present application.

[0065] First, the several terms involved in the present application are analyzed:

[0066] Synthetic Aperture Radar (SAR) is an active microwave remote sensing imaging radar. SAR imaging technology is widely used in geological disaster monitoring, ocean military monitoring and ocean search and rescue due to its all-weather and all-day characteristics. With the rapid development of spaceborne SAR, many countries have developed their own SAR technology, such as TerraSAR-X of Germany, RADARSAT-2 of Canada and Gaofen-3 of China.

[0067] SAR image target detection algorithm: with the development of deep learning methods, a large number of SAR image target detection networks based on deep learning are proposed, and with its strong feature extraction ability and nonlinear mapping learning ability, it has achieved results that cannot be compared with traditional methods (such as CFAR and its variants). At present, SAR image target detection algorithm is directly or indirectly derived from visible light target detection method. However, in SAR image target detection, due to the multi-scale characteristics of the target in SAR image, the target detection method based on deep learning needs to pay attention to targets of different sizes at the same time. According to the principle of deep learning target detection, multiple downsampling operations need to be performed on the input image to extract higher-level semantic features with a larger receptive field. However, with the increase of the number of downsampling operations on the input image, the information of multi-scale targets is distributed on multiple layers of the target detection network. How to ensure that the target detection network can extract effective multi-scale target information and effectively use the information on different layers of the target detection for multi-scale target detection is the key to improving the performance of multi-scale target detection and the challenge brought by the multi-scale target characteristics of SAR image to the target detection network based on deep learning.

[0068] Target detection algorithm: from the perspective of the detection process, deep learning target detection algorithms can be mainly divided into two categories: anchor-based target detection network and anchor-free target detection network. Among them, the anchor-based target detection network divides the target detection task into two stages: first, a large number of candidate boxes are predicted, and then a classifier is used to determine whether the candidate box belongs to the background or the target category. Anchor-free target detection network (such as YOLO series) regards target detection as a regression task and directly regresses the target box position and target category. Compared with anchor-based target detection network, anchor-free target detection network has the advantages of simple model and fast calculation.

[0069] ResNest model: used for target detection or image segmentation, etc., a Split-Attention attention module is introduced on the basis of the ResNet model, and the essence of Split-Attention can be understood as a sliced attention supervision mechanism. The ResNest model has better image classification performance on the ImageNet dataset, especially the ResNest-50 model. For example, the model using ResNest-50 as the basic skeleton (such as the Faster-RCNN model) is 3.08% higher than the model using ResNet-50 (such as mAP); the model using ResNest-50 as the basic skeleton (such as the DeeplabV3 model) is 3.02% higher than the model using ResNet-50 (such as mIOU).

[0070] Deep semantic segmentation model FPN (Feature Pyramid Network): refers to a feature pyramid network, and its core feature is a feature pyramid, that is, a feature map of different scales, which uses a top-down network similar to a decoder to fuse feature information from different stages. The FPN model structure includes an encoder and a decoder, the decoder uses deconvolution to restore the resolution of the feature map, and the encoder and decoder are connected through a 1x1 convolution kernel.

[0071] In related technologies, target detection is performed in a SAR image, and on the basis of a target detection technology based on deep learning, a target detection method based on deep learning is also introduced for SAR image target detection. Although the SAR image target detection based on deep learning has achieved good results, there are still problems of tight calculation and low performance in the face of actual application scenarios.

[0072] The main reasons are as follows: 1) Most SAR image target detection networks use anchor-based target detection networks, which achieve satisfactory results in detection accuracy, but a large number of predicted candidate boxes in the intermediate process introduce a large amount of calculation unrelated to the predicted target. 2) Most SAR image target detection networks use anchor-based target detection methods, which introduce a large amount of calculation related to anchor boxes, causing model calculation tension. 3) The feature extraction capability of part of the skeleton network of the SAR image target detection network is low. 4) The size of the detected target in the SAR image target detection network is not the same, and the span is very large. The SAR image target detection network in the related art has poor detection performance for the multi-scale target detector, and cannot well complete the SAR image target detection task.

[0073] Based on this, the embodiment of the present application provides an image target detection method, device, equipment and storage medium, the coding feature map of the image to be detected is densely connected by using the coding and decoding module to obtain the decoding feature map, the feature extraction capability is enhanced, and then the feature pyramid transformation is used for feature enhancement on the decoding feature map, the detection capability of the subsequent detection head prediction model on the multi-scale target is effectively improved, the detection performance of the SAR target image detection algorithm is improved, and the problem of tight operation capability is alleviated.

[0074] The embodiment of the present application provides an image target detection method, device, equipment and storage medium, and is specifically explained through the following embodiment. First, the image target detection method in the embodiment of the present application is described.

[0075] The embodiment of the present application can acquire and process related data based on artificial intelligence technology. The artificial intelligence (AI) is to use a digital computer or a machine controlled by a digital computer to simulate, extend and expand human intelligence, perceive environment, acquire knowledge and use knowledge to obtain the best results.

[0076] The artificial intelligence basic technology generally includes technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics, etc. The artificial intelligence software technology mainly includes computer vision technology, robot technology, biometric technology, speech processing technology, natural language processing technology and machine learning / deep learning, etc.

[0077] The image target detection method provided by the embodiment of the present application relates to the field of artificial intelligence technology, especially to the field of data mining technology. The image target detection method provided by the embodiment of the present application can be applied in a terminal, can be applied in a server, and can also be software running in the terminal or the server. The terminal communicates with the server through a network. The image target detection method can be executed by the terminal or the server, or cooperatively executed by the terminal and the server.

[0078] In some embodiments, the terminal can be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart watch, or the like. The server can be a standalone server, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and basic cloud computing services such as big data and artificial intelligence platforms; or a service node in a blockchain system, the service nodes in the blockchain system form a peer-to-peer (P2P) network among each other, and the P2P protocol is an application layer protocol running on the transmission control protocol (TCP) protocol. The server can install a server side, and the terminal can interact with the server side, for example, the server side installs corresponding software, and the software can be an application for implementing the image target detection method, but is not limited to the above forms. The terminal and the server can be connected through a communication connection mode such as Bluetooth, a universal serial bus (USB), or a network, and the present embodiment is not limited herein.

[0079] The present application can be applied in numerous general purpose or special purpose computer system environments or configurations. For example: personal computer, server computer, handheld or portable device, tablet device, multiprocessor system, microprocessor-based system, set top box, programmable consumer electronics, network PC, minicomputer, mainframe computer, distributed computing environments that include any of the above systems or devices, and the like. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. The present application can also be practiced in a distributed computing environment, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.

[0080] Figure 1 is an optional flowchart of the image target detection method provided by the embodiments of the present application, Figure 1 The method in the above embodiment can include, but is not limited to, steps S110 to S150.

[0081] Step S110: acquiring a to-be-detected image and inputting the to-be-detected image into a trained image detection model.

[0082] In an embodiment, the image detection model is first trained by using a large amount of sample image data to obtain a trained image detection model. The sample image data can be a large number of images containing or not containing a target, and the corresponding detection label is: target, other target and position information corresponding to the target.

[0083] In this embodiment, the image detection model comprises a feature extraction model, a feature enhancement model and a detection head prediction model, wherein the feature extraction model comprises an encoding module and a decoding module, and the detection head prediction model comprises a target center position regression prediction module, a target center point offset regression prediction module and a target size regression prediction module. In this embodiment, the image detection model uses a ResNeSt-50 model as a basic skeleton, and is improved on the basis of the ResNeSt-50 model. Figure 2

[0084] Step S120: performing feature extraction on the to-be-detected image by using the encoding module to obtain at least one encoded feature map.

[0085] In an embodiment, the ResNeSt-50 model constitutes the main part of the encoding module, and the implementation of the encoding process includes four stages. The to-be-detected image I is subjected to feature extraction to obtain four encoded feature maps F of different scales. encode For example, the four encoded feature maps can be a first encoded feature map F encode1 , a second encoded feature map F encode2 , a third encoded feature map F encode3 and a fourth encoded feature map F encode4 , wherein the sizes of the encoded feature maps are respectively: 128x128x256, 64x64x512, 32x32x1024 and 16x16x2048. The encoded feature map F encode is represented as:

[0086] F encode ={F encode1 ,F encode2 ,F encode3 ,F encode4}=ResNet50(I)

[0087] For example, in an embodiment, the encoding module implements an encoding process including five stages, and one more initial encoded feature map F encode0 is added on the basis of the above four encoded feature maps, and the size of the initial encoded feature map F encode0 is 128x128x64. The encoded feature map is represented as:

[0088] F encode ={F encode0 ,F encode1 ,F encode2 ,F encode3 ,F encode4}=ResNet50(I)

[0089] It can be understood that the number and size of the encoded feature maps in the above embodiments can be set according to actual needs, and the above embodiments are only illustrative and do not limit the number and size of the encoded feature maps.

[0090] The above encoding module can more effectively extract image features. After obtaining the encoded feature maps output by the encoding module, the embodiments of the present application need to decode them.

[0091] Step S130: using a decoding module to densely connect at least one encoded feature map to obtain a decoded feature map.

[0092] In an embodiment, the encoded feature maps obtained by the encoding module belong to high-dimensional encoded features, and the decoding module maps the high-dimensional encoded features to decoded features rich in high-level semantic information. The decoding process in this embodiment includes three convolution stages, and the decoded feature map is obtained from at least one encoded feature map by using a dense connection method. Finally, the high-dimensional encoded feature map is decoded into three scale decoded feature maps, which are a first decoded feature map, a second decoded feature map, and a decoded feature map.

[0093] In this embodiment, dense connection refers to connecting features in the channel to realize feature reuse, that is, connecting all layers to each other. Specifically, each layer will accept all previous layers as its additional input, and each layer will be connected with all previous layers in the channel dimension and used as the input of the next layer. The size of the feature map of each layer is the same when connected. This way makes the image detection model of the present application achieve better performance with less parameters and computational cost.

[0094] Referring to Figure 3 , step S130 includes but is not limited to steps S131 to S133.

[0095] Step S131: according to the dimension of the first encoded feature map, performing feature map skipping on at least one encoded feature map to obtain a first decoded feature map.

[0096] In an embodiment, since the size (such as dimension) of the feature map of each layer is the same when connected by the dense connection method, only when the size of the feature map of each layer is consistent, the connection in the channel dimension can be performed. Therefore, when calculating the first decoded feature map, the size of the feature map of each layer is adjusted according to the dimension of the first encoded feature map in this embodiment.

[0097] In an embodiment, the encoded feature map F encode includes a first encoded feature map F encode1 , a second encoded feature map F encode2 , and a third encoded feature map F encode3and the fourth encoded feature map F encode4 , the size of the first decoded feature map F1 can be 32x32x256, the size of the second decoded feature map F2 can be 64x64x128, and the size of the decoded feature map F3 can be 128x128x64.

[0098] In an embodiment, referring to Figure 4 , the step S131 includes but is not limited to steps S1311-S1314.

[0099] In step S1311, the first encoded feature map is down-sampled to obtain a first down-sampling result.

[0100] In an embodiment, the down-sampling process is to compress the feature dimension to a specified size, such as half or less of the current input, through one or more convolution layers according to the required dimension, and then perform the pooling operation. In this embodiment, the size of the first encoded feature map F encode1 is 128x128x256, and the first down-sampling result is represented as f down (F encode1 ), the first encoded feature map F encode1 is compressed in feature dimension through one or more convolution layers, and the size can be 32x32x256, wherein f down (·) represents the compressed feature map channel and the down-sampling operation.

[0101] In step S1312, the second encoded feature map is down-sampled to obtain a second down-sampling result.

[0102] In an embodiment, the size of the second encoded feature map F encode2 is 64x64x512, and the second down-sampling result is represented as f down (F encode2 ), the second encoded feature map F encode2 is compressed in feature dimension through one or more convolution layers, and the size can be 32x32x256, wherein f down () represents the compressed feature map channel and the down-sampling operation.

[0103] In step S1313, the fourth encoded feature map is de-convolved to obtain a fourth de-convolution result.

[0104] In an embodiment, deconvolution is used to map the low-dimensional feature map to a high-dimensional input, which is opposite to the convolution operation. For the deconvolution process, the filter after transposition of the convolution process (the parameters of the filter are the same, and the parameter matrix is flipped horizontally and vertically during deconvolution) is multiplied in reverse. In this embodiment, the size of the fourth encoded feature map F encode4 is 16x16x2048, and the fourth deconvolution result is represented as Deconv(F encode4 ). After deconvolution of the fourth encoded feature map F encode4 , the size can be 32x32x256, where Deconv(·) represents the deconvolution operation on the feature map.

[0105] In step S1314, the first decoded feature map is calculated.

[0106] In an embodiment, the first decoded feature map is obtained by multiplying and adding each weight parameter in the preset first weight parameter set S1 and the corresponding first down-sampling result f down (F encode1 ), the third encoded feature map F encode3 , the second down-sampling result f down (F encode2 ), and the fourth deconvolution result Deconv(F encode4 ), respectively. The preset first weight parameter set S1 is used to represent the weight corresponding to the skip connection path in the dense connection process for calculating the first decoded feature map F1. The parameters can be obtained more optimally according to the training process, and the process for calculating the first decoded feature map F1 is represented as:

[0107] F1=Deconv(F encode4 )+χ1f down (F encode1 )+β1f down (F encode2 )+α1F encode3

[0108] The preset first weight parameter set S1 is represented as {1, α1, β1, χ1}. It can be understood that each weight in the preset first weight parameter set S1 corresponds to the first down-sampling result f down (F encode1 ), the third encoded feature map F encode3 , the second down-sampling result f down (F encode2 ), and the fourth deconvolution result Deconv(F encode4 ), but the order of the weights is not specifically limited in this embodiment.

[0109] Since dense connection connects all layers to each other, each layer receives all previous layers as its additional input, and each layer is connected with all previous layers in the channel dimension and serves as the input of the next layer, in an embodiment, after obtaining the first decoding feature map, the second decoding feature map is calculated as input.

[0110] In step S132, the first decoding feature map and the at least one encoding feature map are feature map jumped to obtain the second decoding feature map according to the dimension of the second decoding feature map.

[0111] In an embodiment, referring to Figure 5 , step S132 includes but is not limited to steps S1321-S1323.

[0112] In step S1321, the first decoding feature map is deconvolved to obtain a first deconvolution result.

[0113] In an embodiment, the size of the first decoding feature map F1 can be 32x32x256, the first deconvolution result is represented as Deconv(F1), and after the first decoding feature map F1 is deconvolved, the size can be 64x64x128, wherein Deconv(·) represents a deconvolution operation on a feature map.

[0114] In step S1322, the third encoding feature map is up-sampled to obtain a third up-sampling result.

[0115] In an embodiment, the up-sampling operation uses an interpolation method to up-sample the feature map to a specified size, the size of the third encoding feature map F encode3 is 32x32x1024, and the third up-sampling result is represented as f up (Deconv(F encode3 )), wherein the size of the third encoding feature map F encode3 after up-sampling can be 64x64x128, and f up (·) represents an up-sampling operation on a feature map.

[0116] In step S1323, the second decoding feature map is calculated.

[0117] In an embodiment, the second decoding feature map F2 is obtained by multiplying each weight parameter in the preset second weight parameter set S2 with the corresponding first deconvolution result Deconv(F1), the second encoding feature map F encode2 , the first down-sampling result f down (Deconv(F encode1 )), and the third up-sampling result f up (Deconv(F encode3 )) respectively and then adding them up, wherein the first down-sampling result f down (Deconv(Fencode1 The calculation process of step S1311 is the same as that of step S1311.

[0118] The preset second weight parameter set S2 is used to represent the weight corresponding to the jump path in the dense connection process of calculating the second decoding feature map F2. The parameters can be obtained more optimally according to the training process, and the process of calculating the second decoding feature map F2 is represented as:

[0119] F2 = Deconv(F1) + β2f down (F encode1 ) + α2F encode2 + χ2f up (F encode3 )

[0120] The preset second weight parameter set S2 is represented as {1, α2, β2, χ2}. It can be understood that each weight in the preset second weight parameter set S2 corresponds to the first deconvolution result Deconv(F1), the second encoding feature map F encode2 , the first down-sampling result f down (F encode1 ), and the third up-sampling result f up (F encode3 ) respectively, but the order of the weights is not specifically limited in the embodiment.

[0121] In an embodiment, after obtaining the second decoding feature map F2, the second decoding feature map F2 is taken as an input to calculate a decoding feature map.

[0122] Step S133, according to the dimension of the encoding feature map, the second decoding feature map and at least one encoding feature map are connected to obtain a decoding feature map.

[0123] In an embodiment, referring to Figure 6 , step S133 includes but is not limited to steps S1331 to S1333.

[0124] Step S1331, the second decoding feature map is deconvoluted to obtain a second deconvolution result.

[0125] In an embodiment, the size of the second decoding feature map F2 can be 64x64x128, and the second deconvolution result is represented as Deconv(F2). After the second decoding feature map F2 is deconvoluted, the size can be 128x128x64, wherein Deconv(·) represents a deconvolution operation on the feature map.

[0126] Step S1332, the second encoding feature map is up-sampled to obtain a second up-sampling result.

[0127] In an embodiment, the second encoding feature map Fencode2 The dimensions are: 64×64×512, and the second upsampling result f up (F encode2 ), second encoded feature map F encode2 After upsampling, the size can be: 128×128×64, where f up (·) indicates an upsampling operation on the feature map.

[0128] Step S1333: Calculate the decoded feature map.

[0129] In one embodiment, each weight parameter in the preset third weight parameter set S3 is respectively compared with the corresponding second deconvolution result Deconv(F2) and the first encoded feature map F. encode1 Second upsampling result f up (F encode2 ) and the third upsampling result f up (F encode3 The products are multiplied and then added together to obtain the decoded feature map F3. Among them, the third upsampling result f... up (F encode3 The calculation process is the same as that in step S1322.

[0130] The preset third weight parameter set S3 is used to represent the weights of the jump paths in the dense connection process of the decoded feature map F3. Optimal parameters can be obtained based on the training process. The process of calculating the decoded feature map F3 is expressed as follows:

[0131] F3=Deconv(F2)+α3F encode1 +β3f up (F encode2 )+χ3f up (F encode3 )

[0132] The preset third weight parameter set S3 is represented as {1, α3, β3, χ3}. This can be understood as each weight in the preset third weight parameter set S3 corresponding to the second deconvolution result Deconv(F2) and the first encoded feature map F, respectively. encode1 Second upsampling result f up (F encode2 ) and the third upsampling result f up (F encode3 However, this embodiment does not specifically limit the order of the weights.

[0133] The embodiment adopts the idea of dense connection in the decoding process to connect each layer with all previous layers in the channel dimension, and calculate layer by layer as the input of the next layer, reuse the features lost in the coding and decoding process, and jump the coding feature maps obtained in the coding stage and the decoding feature maps obtained in the decoding stage. Wherein the feature maps of the same dimension are directly added, and the feature maps of different dimensions are adjusted to the same scale by upsampling, downsampling or deconvolution and then added. Finally, the decoding feature map F3 is obtained by using the dense connection method.

[0134] In an embodiment, the decoding feature map F3 is enhanced by using the following step S140.

[0135] Step S140: using a feature enhancement model to perform feature pyramid transformation on the decoding feature map to obtain an enhanced feature map.

[0136] In an embodiment, the pyramid transformation in the feature enhancement model refers to using an FPN network to enhance the decoding feature map. The process can be described as compression-recovery. First, the decoding feature map is compressed to obtain feature maps of different scales, and then the size is recovered by convolution or upsampling to obtain an enhanced feature map. By exchanging information between different layers, communication between different layers is allowed, further promoting information propagation, and the enhanced feature map can better detect significant targets.

[0137] Referring to Figure 7 , step S140 includes but is not limited to steps S141 to S143.

[0138] Step S141: compressing the decoding feature map to obtain at least one compressed feature map.

[0139] In an embodiment, the decoding feature map F3 is subjected to a convolution operation to further compress it to obtain three compressed feature maps of different sizes, namely a first compressed feature map F bottom , a second compressed feature map F middle , and a third compressed feature map F top , wherein the size of the first compressed feature map F bottom is the largest, which is 128x128x256, the size of the second compressed feature map F middle is the second largest, which is 64x64x64, and the size of the third compressed feature map F top is the smallest, which is 32x32x64. The compressed feature maps are represented as:

[0140] {F bottom ,F middle ,F top}=Conv(F3)

[0141] Wherein, Conv(·) represents a convolution operation.

[0142] Step S142, convolution is performed on the at least one compressed feature map to obtain at least one compressed convolution feature map.

[0143] In an embodiment, the first compressed feature map F bottom , the second compressed feature map F middle , and the third compressed feature map F top are respectively adjusted to compressed convolution feature maps of the same size by convolution or up-sampling.

[0144] The specific process is as follows:

[0145] The first compressed feature map F bottom is subjected to convolution to obtain a first compressed convolution feature map F fpn1 , which is expressed as:

[0146] F fpn1 =Conv(F bottom )

[0147] The second compressed feature map F middle is first subjected to up-sampling to adjust the size and then subjected to convolution to obtain a second compressed convolution feature map F fpn2 . The number of times of up-sampling is not limited in this embodiment, and the purpose of up-sampling is to make the size of the second compressed convolution feature map F fpn2 obtained finally the same as that of the first compressed convolution feature map F fpn1 , for example, the up-sampling multiple can be 2. The second compressed convolution feature map F fpn2 is expressed as:

[0148] F fpn2 =Conv(Up ×2 (F middle ))

[0149] Up ×2 (·) represents two times of up-sampling, and the number of times of up-sampling is 2 in this embodiment only for illustration and does not represent a limitation on the up-sampling multiple.

[0150] The third compressed feature map F top is first subjected to up-sampling to adjust the size and then subjected to convolution, and the convolution result is subjected to up-sampling again, and the up-sampling result is subjected to convolution again to obtain a third compressed convolution feature map F fpn3 . The number of times of up-sampling is not limited in this embodiment, and the purpose of up-sampling is to make the size of the third compressed convolution feature map F fpn3 obtained finally the same as that of the second compressed convolution feature map F fpn2 and the first compressed convolution feature map F fpn1The same, for example, the first up-sampling multiple can be 2 times, and the obtained feature map size is the same as the second compressed feature map F middle The same, the second up-sampling is performed again, the up-sampling multiple is 2 times, and the final third compressed convolution feature map F fpn3 is obtained through convolution.

[0151] F fpn3 =Conv(Up ×2 (Conv(Up ×2 (F top )))

[0152] Up ×2 (·) represents two times up-sampling, and the up-sampling multiple is 2 times in the embodiment only for illustration, and does not represent limitation of the up-sampling multiple.

[0153] In step S143, at least one compressed convolution feature map is aligned and up-sampled according to the dimension of the enhanced feature map, and the enhanced feature map is obtained.

[0154] In an embodiment, after three compressed convolution feature maps are obtained, they are added element by element, and then the feature is aligned through convolution and is up-sampled by 2 times, so that the compressed convolution feature maps of different scales are up-sampled to a unified size, and subsequent category prediction is performed, and the enhanced feature map F enhance is obtained.

[0155] F enhance =Up ×2 (Conv(F fpn1 +F fpn2 +F fpn3 ))

[0156] The enhanced feature map F enhance obtained through the above steps is used for target detection, and more accurate results can be obtained, and the problem of poor detection performance when detecting multi-scale targets is further solved. The following step S150 is used for detecting the target in the enhanced feature map. It can be understood that the target can be one or more than one target of different scales, and the number and size of the target are not limited in the embodiment.

[0157] Step S150: The enhanced feature map is predicted by using a detection head prediction model to obtain a target detection result.

[0158] In an embodiment, for target detection, the detection head prediction model adopts a regression prediction model, which includes a target center position regression prediction module, a target center point offset regression prediction module and a target size regression prediction model, which are respectively used for predicting the center position of different targets, the center point offset of different targets and the size of different targets.

[0159] Referring to Figure 8Step S150 includes but is not limited to steps S151-S154.

[0160] In step S151, the target center position regression prediction module is used to predict the enhanced feature map to obtain a target center position output value.

[0161] In an embodiment, the target center position regression prediction module is mainly used to determine the confidence of the corresponding position in the enhanced feature map being the center point of the target, wherein the confidence is the probability value of the corresponding position being the center point of the target string, and the target center position output value output by the target center position regression prediction module is a heat map with a scale of 256x256x1, and the value range of each pixel in the heat map is 0-1, and the value represents the probability size.

[0162] In step S152, the target center point offset regression prediction module is used to predict the enhanced feature map to obtain a target center point offset output value.

[0163] In an embodiment, the target center point offset regression prediction module is used to predict the offset distance of each position relative to the center point of the real target, and the offset distance can be represented by a coordinate point, so the target center point offset output value output by the target center point offset regression prediction module is a 256x256x2 matrix representing the offset distance coordinates of each position.

[0164] In step S153, the target size regression prediction model is used to predict the enhanced feature map to obtain a target size output value.

[0165] In an embodiment, the target size regression prediction model is used to predict the size of the target at each position in the horizontal and vertical directions, and the target size output value output by the target size regression prediction model is a 256x256x2 matrix representing the size values of each position in the horizontal and vertical directions.

[0166] In step S154, a target detection result is obtained.

[0167] In an embodiment, after obtaining the target center position output value, the target center point offset output value, and the target size output value output by the three detection head prediction models, first, the target that meets the pre-set confidence threshold is selected as the detection target according to the comparison result of the target center position output value and the pre-set confidence threshold, and then the values related to the corresponding position of the detection target in the target center point offset output value and the target size output value are selected as the target detection result, i.e., the target detection result is the target center position output value, the target center point offset output value, and the target size output value corresponding to the detection target.

[0168] For example, in an embodiment, the output result of the detection head prediction model is:

[0169] Target center position output value: [confidence z1 of target 1, confidence z2 of target 2, confidence z3 of target 3…], target center point offset output value: [offset distance size (x1, y1) of target 1, offset distance size (x2, y2) of target 2, offset distance size (x3, y3) of target 3, …], target size output value: [size size (c1, c1) of target 1, size size (c2, c2) of target 2, size size (c3, c4) of target 3, …].

[0170] When the confidence z2 of the target 2 is within the range of the preset confidence threshold, the target 2 is the detection target, and the target detection result at this time is: [z2, (x2, y2), (c2, c2)].

[0171] Reference Figure 9 , the schematic diagram of the image target detection method of the embodiment of the present application.

[0172] In this embodiment, Figure 9 The input layer inputs the image to be detected L, whose size is 256x256x3, and then enters the encoding stage. The encoding module encodes the input image to be detected to obtain five encoding feature maps, and the feature maps output by different layers in turn are: the initial encoding feature map F encode0 , the first encoding feature map F encode1 , the second encoding feature map F encode2 , the third encoding feature map F encode3 , and the fourth encoding feature map F encode4 .

[0173] In this embodiment, after entering the decoding stage, the decoding module obtains the first decoding feature map F1, the second decoding feature map F2, and the decoding feature map F3 from the encoding feature maps by using the dense connection method, Figure 9 In which the dashed line between different layers indicates that the feature maps obtained by the two layers have the same size, and the connection line schematically shows the process of dense connection. For example, the connection line between the first encoding feature map F encode1 and the first decoding feature map F1 and the second decoding feature map F2 indicates that the deconvolution operation is performed on the first encoding feature map F encode1 ; the connection line between the second encoding feature map F encode2 and the first decoding feature map F1 indicates the downsampling operation, and the connection line between the second encoding feature map F encode2 and the decoding feature map F3 indicates the upsampling operation on the second encoding feature map F encode2 ; and the connection line between the third encoding feature map F encode3 and the second decoding feature map F2 and the decoding feature map F3 indicates the upsampling operation.

[0174] In this embodiment, after obtaining the decoding feature map F3, a feature enhancement stage is entered, and first, a first compressed feature map F bottom , a second compressed feature map F middle , and a third compressed feature map F top are obtained from the decoding feature map F3, and first, second, and third compressed convolution feature maps F fpn1 , F fpn2 , and F fpn3 are obtained, respectively. enhance .

[0175] In this embodiment, the detection stage detects the enhanced feature map F enhance , and obtains target center position output values (heatmaps in the figure), target center point offset output values (offsets in the figure), and target size output values (sizes in the figure).

[0176] As described above, in the embodiments of the present application, a to-be-detected image is obtained and input into a trained image detection model, an encoding module is used to extract features of the to-be-detected image to obtain at least one encoding feature map, a decoding module is used to densely connect the at least one encoding feature map to obtain a decoding feature map, a feature enhancement model is used to perform feature pyramid transformation on the decoding feature map to obtain an enhanced feature map, and a detection head prediction model is used to predict the enhanced feature map to obtain a target detection result. In the embodiments, the encoding feature map of the to-be-detected image is densely connected by using an encoding and decoding module to obtain a decoding feature map, the feature extraction capability is enhanced, the decoding feature map is enhanced by using feature pyramid transformation, the detection capability of a subsequent detection head prediction model for multi-scale targets is effectively improved, the detection performance of a SAR target image detection algorithm is improved, and the problem of tight operation capability is alleviated.

[0177] According to whether a preset anchor box is needed in a network processing process, target detection networks can be divided into two categories: anchor box-based and anchor box-free target detection networks. In the embodiments of the present application, the image detection model is an anchor box-free target detection network, which has the advantages of strong generalization, easy calculation, and simple network compared with anchor box-based target detection networks.

[0178] The embodiments of the present application also provide an image detection device, which can implement the image target detection method described above, and the device comprises Figure 10 , and the device comprises

[0179] The image acquisition unit 1010 is configured to acquire an image to be detected and input the image to be detected into a trained image detection model, the image detection model comprising: a feature extraction model, a feature enhancement model and a detection head prediction model, the feature extraction model comprising: an encoding module and a decoding module;

[0180] The encoding unit 1020 is configured to perform feature extraction on the image to be detected by using the encoding module to obtain at least one encoded feature map;

[0181] The decoding unit 1030 is configured to perform dense connection on the at least one encoded feature map by using the decoding module to obtain a decoded feature map;

[0182] The feature enhancement unit 1040 is configured to perform feature pyramid transformation on the decoded feature map by using the feature enhancement model to obtain an enhanced feature map;

[0183] The detection unit 1050 is configured to perform prediction on the enhanced feature map by using the detection head prediction model to obtain a target detection result.

[0184] The specific implementation of the image detection device of the embodiment is basically the same as the specific implementation of the image target detection method described above, and will not be repeated here.

[0185] The embodiment of the present application also provides an electronic device, comprising:

[0186] at least one memory;

[0187] at least one processor;

[0188] at least one program;

[0189] The program is stored in the memory, and the processor executes the at least one program to implement the image target detection method described above. The electronic device can be any intelligent terminal including mobile phone, tablet computer, personal digital assistant (Personal Digital Assistant, PDA), vehicle-mounted computer, etc.

[0190] Please refer to Figure 11 , Figure 11 The hardware structure of the electronic device of another embodiment is illustrated, and the electronic device comprises:

[0191] The processor 1101 can be implemented in the form of a general-purpose CPU (Central Processing Unit), a microprocessor, an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits, etc., and is configured to execute related programs to implement the technical solutions provided by the embodiments of the present application;

[0192] The memory 1102 can be implemented in the form of a ROM (Read Only Memory), a static storage device, a dynamic storage device, or a RAM (Random Access Memory), etc. The memory 1102 can store an operating system and other application programs, and when the technical solutions provided by the embodiments of the present specification are implemented by software or firmware, the related program codes are stored in the memory 1102 and are called and executed by the processor 1101 to implement the image target detection method of the embodiments of the present application;

[0193] The input / output interface 1103 is used to realize information input and output;

[0194] The communication interface 1104 is used to realize the communication interaction between the device and other devices, which can realize communication through wired mode (such as USB, network cable, etc.) or wireless mode (such as mobile network, WIFI, Bluetooth, etc.);

[0195] The bus 1105 transmits information between various components (such as the processor 1101, the memory 1102, the input / output interface 1103, and the communication interface 1104) of the device;

[0196] The processor 1101, the memory 1102, the input / output interface 1103, and the communication interface 1104 are connected to each other through the bus 1105 to realize communication connection between them inside the device.

[0197] The embodiments of the present application also provide a storage medium, which is a computer readable storage medium, and the computer readable storage medium stores computer executable instructions for causing a computer to execute the above-mentioned image target detection method.

[0198] The image target detection method, the image detection device, the electronic device, and the storage medium provided by the embodiments of the present application can obtain a to-be-detected image, input the to-be-detected image into a trained image detection model, use an encoding module to extract at least one encoding feature map from the to-be-detected image, use a decoding module to densely connect the at least one encoding feature map to obtain a decoding feature map, use a feature enhancement model to perform feature pyramid transformation on the decoding feature map to obtain an enhanced feature map, and use a detection head prediction model to predict the enhanced feature map to obtain a target detection result. In this embodiment, the encoding feature map of the to-be-detected image is densely connected by using an encoding and decoding module to obtain a decoding feature map, which enhances the feature extraction capability. Then, the decoding feature map is enhanced by using feature pyramid transformation, which effectively improves the detection capability of the subsequent detection head prediction model for multi-scale targets, improves the detection performance of the SAR target image detection algorithm, and alleviates the problem of tight operation capacity.

[0199] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. In addition, the memory can include a high-speed random access memory and can also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some embodiments, the memory can optionally include a memory that is remotely disposed relative to the processor, and these remote memories can be connected to the processor through a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0200] The embodiments described in the embodiments of the present application are for more clearly illustrating the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided by the embodiments of the present application. Those skilled in the art can know that, with the evolution of technology and the appearance of new application scenarios, the technical solutions provided by the embodiments of the present application are also applicable to similar technical problems.

[0201] Those skilled in the art can understand that, Figures 1-11 The technical solutions shown in the above description do not constitute a limitation on the embodiments of the present application, and can include more or fewer steps than shown in the figures, or combine certain steps, or different steps.

[0202] The device embodiments described above are only schematic, and the units described as separate components can or can not be physically separate, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiments of the present application.

[0203] Those skilled in the art can understand that all or some of the steps in the above disclosed method, the functional modules / units in the system and the device can be implemented as software, firmware, hardware and their appropriate combinations.

[0204] The terms "first", "second", "third", "fourth" and the like used in the description of the present application and the above-described drawings (if any) are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0205] It should be understood that, in the present application, "at least one" means one or more, and "multiple" means two or more. "And / or" is used to describe the relationship between associated objects, which means that there can be three relationships, for example, "A and / or B" can mean: only A, only B, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects. "At least one of the following" or similar expressions means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0206] In several embodiments provided by the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only illustrative, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed units can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0207] The units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, that is, they can be located in one place, or they can be distributed on multiple network units. According to actual needs, some or all of the units can be selected to achieve the purpose of the embodiment scheme.

[0208] In addition, the functional units in each embodiment of the present application can be integrated into a processing unit, or each unit can exist physically, or two or more units can be integrated into one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0209] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in other words, the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes multiple instructions for making an electronic device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program storage media.

[0210] The preferred embodiments of the embodiments of the present application are described above with reference to the accompanying drawings, and are not limited to the scope of the embodiments of the present application. Any modifications, equivalent replacements and improvements made by those skilled in the art without departing from the scope and essence of the embodiments of the present application shall be within the scope of the embodiments of the present application.

Claims

1. An image target detection method, characterized in that, include: The image to be detected is acquired and input into a trained image detection model. The image detection model includes a feature extraction model, a feature enhancement model, and a detection head prediction model. The feature extraction model includes an encoding module and a decoding module. The encoding module is used to extract features from the image to be detected to obtain at least one encoded feature map, wherein the at least one encoded feature map includes: a first encoded feature map, a second encoded feature map, a third encoded feature map, and a fourth encoded feature map; The decoding module is used to densely connect at least one of the encoded feature maps to obtain a decoded feature map; The enhanced feature map is obtained by performing a feature pyramid transformation on the decoded feature map using the aforementioned feature enhancement model. The target detection result is obtained by using the detection head prediction model to predict the enhanced feature map; The step of using the decoding module to perform dense concatenation on at least one of the encoded feature maps to obtain a decoded feature map includes: Based on the dimension of the first encoded feature map, the first encoded feature map is downsampled to obtain a first downsampled result; the second encoded feature map is downsampled to obtain a second downsampled result; the fourth encoded feature map is deconvolved to obtain a fourth deconvolution result; and each weight parameter in the preset first weight parameter set is multiplied by the corresponding first downsampled result, the third encoded feature map, the second downsampled result, and the fourth deconvolution result, and then summed to obtain a first decoded feature map. Based on the dimension of the second decoded feature map, the first decoded feature map is deconvolved to obtain the first deconvolution result; the third encoded feature map is upsampled to obtain the third upsampled result; and each weight parameter in the preset second weight parameter set is multiplied by the corresponding first deconvolution result, the second encoded feature map, the first downsampled result, and the third upsampled result, and then added to obtain the second decoded feature map. Based on the dimension of the encoded feature map, the second decoded feature map is deconvolved to obtain a second deconvolution result; the second encoded feature map is upsampled to obtain a second upsampled result; and each weight parameter in the preset third weight parameter set is multiplied by the corresponding second deconvolution result, the first encoded feature map, the second upsampled result, and the third upsampled result, and then added together to obtain the decoded feature map.

2. The image target detection method according to claim 1, characterized in that, The step of using the feature enhancement model to perform feature pyramid transformation on the decoded feature map to obtain the enhanced feature map includes: The decoded feature map is compressed to obtain at least one compressed feature map; Convolution is performed on at least one of the compressed feature maps to obtain at least one compressed convolutional feature map; The enhanced feature map is obtained by aligning and upsampling at least one of the compressed convolutional feature maps according to the dimension of the enhanced feature map.

3. The image target detection method according to any one of claims 1 to 2, characterized in that, The detection head prediction model includes: a target center position regression prediction module, a target center point offset regression prediction module, and a target size regression prediction model. The step of using the detection head prediction model to predict the enhanced feature map to obtain the target detection result includes: The target center position output value is obtained by predicting the enhanced feature map using the target center position regression prediction module. The target center point offset output value is obtained by predicting the enhanced feature map using the target center point offset regression prediction module. The target size output value is obtained by predicting the enhanced feature map using the target size regression prediction model. The target is obtained by comparing the output value of the target center position with the preset confidence threshold. The target detection result is the output value of the target center position, the output value of the target center point offset, and the output value of the target size corresponding to the target.

4. An image detection device, characterized in that, include: An image acquisition unit is used to acquire an image to be detected and input the image to be detected into a trained image detection model. The image detection model includes a feature extraction model, a feature enhancement model, and a detection head prediction model. The feature extraction model includes an encoding module and a decoding module. The encoding unit is used to extract features from the image to be detected using the encoding module to obtain at least one encoded feature map, wherein the at least one encoded feature map includes: a first encoded feature map, a second encoded feature map, a third encoded feature map, and a fourth encoded feature map; A decoding unit is used to densely connect at least one of the encoded feature maps using the decoding module to obtain a decoded feature map; A feature enhancement unit is used to perform a feature pyramid transformation on the decoded feature map using the feature enhancement model to obtain an enhanced feature map. The detection unit is used to predict the enhanced feature map using the detection head prediction model to obtain the target detection result; The step of using the decoding module to perform dense concatenation on at least one of the encoded feature maps to obtain a decoded feature map includes: Based on the dimension of the first encoded feature map, the first encoded feature map is downsampled to obtain a first downsampled result; the second encoded feature map is downsampled to obtain a second downsampled result; the fourth encoded feature map is deconvolved to obtain a fourth deconvolution result; and each weight parameter in the preset first weight parameter set is multiplied by the corresponding first downsampled result, the third encoded feature map, the second downsampled result, and the fourth deconvolution result, and then summed to obtain a first decoded feature map. Based on the dimension of the second decoded feature map, the first decoded feature map is deconvolved to obtain the first deconvolution result; the third encoded feature map is upsampled to obtain the third upsampled result; and each weight parameter in the preset second weight parameter set is multiplied by the corresponding first deconvolution result, the second encoded feature map, the first downsampled result, and the third upsampled result, and then added to obtain the second decoded feature map. Based on the dimension of the encoded feature map, the second decoded feature map is deconvolved to obtain a second deconvolution result; the second encoded feature map is upsampled to obtain a second upsampled result; and each weight parameter in the preset third weight parameter set is multiplied by the corresponding second deconvolution result, the first encoded feature map, the second upsampled result, and the third upsampled result, and then added together to obtain the decoded feature map.

5. An electronic device, characterized in that, Including the processor and memory; The memory is used to store programs; The processor is configured to execute the image target detection method as described in any one of claims 1 to 3 according to the program.

6. A computer-readable storage medium storing computer-executable instructions for performing the image target detection method according to any one of claims 1 to 3.

Images