An infrared small target detection method, device, computing equipment and storage medium

Abstract

The application provides an infrared small target detection method and device, a computing device and a storage medium. The method comprises: acquiring an infrared image of a target to be detected; inputting the infrared image into a pre-constructed basic network structure for feature extraction to obtain at least one low-level feature map and one high-level feature map; inputting the high-level feature map into a multi-receptive field model to obtain a first feature map; wherein the multi-receptive field model is used to expand the receptive field through a hollow convolution to obtain the first feature map rich in semantic information; and inputting the first feature map and the at least one low-level feature map into a fusion model according to a preset cascading rule to obtain a target detection result. The infrared small target detection method provided by the scheme can improve the detection accuracy of small targets and reduce the false alarm rate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection, and in particular to an infrared small target detection method and device, a computing device and a storage medium. BACKGROUND

[0002] Infrared detection technology has the advantage of not being affected by environmental factors, and is increasingly widely used in military fields and industrial, medical and other livelihood fields. Infrared small targets are a type of target that lacks color texture and other characteristics, has extremely small size and low contrast. Due to the small number of available features, low image signal-to-noise ratio, complex imaging environment and high degree of similarity of the background, it is difficult to detect. However, as an important branch of infrared image detection, infrared small target detection has important applications in early warning of targets, detection of unmanned aerial vehicles and guidance of missiles.

[0003] Currently, the position information and geometric information of small targets are obtained by the method of infrared small target segmentation. However, the current target segmentation algorithm has weak feature extraction capability for target background, and there is a certain error in the fusion of small targets and background, resulting in low precision of the obtained position information and shape information of small targets. SUMMARY

[0004] The embodiments of the present application provide an infrared small target detection method, device, computing device and storage medium, which improve the detection precision of small targets and reduce the false alarm rate.

[0005] In a first aspect, the embodiments of the present application provide an infrared small target detection method, comprising:

[0006] obtaining an infrared image of a target to be detected;

[0007] inputting the infrared image into a pre-constructed basic network structure for feature extraction to obtain at least one low-level feature map and one high-level feature map;

[0008] inputting the high-level feature map into a multi-receptive field model to obtain a first feature map; wherein the multi-receptive field model is used to expand the receptive field by a hole convolution to obtain the first feature map with rich semantic information;

[0009] inputting the first feature map and the at least one low-level feature map into a fusion model according to a pre-set cascading rule to obtain a target detection result.

[0010] Optionally, the step of inputting the infrared image into the pre-constructed basic network structure for feature extraction to obtain at least one low-level feature map and one high-level feature map comprises:

[0011] The basic network structure comprises a first convolutional layer, a second convolutional layer and a third convolutional layer;

[0012] convolve the infrared image by using the first convolutional layer to obtain a first low-layer feature map;

[0013] convolve the first low-layer feature map by using the second convolutional layer to obtain a second low-layer feature map;

[0014] convolve the second low-layer feature map by using the third convolutional layer to obtain a high-layer feature map.

[0015] Optionally, the multi-receptive field model comprises at least two convolutional modules with different receptive fields in parallel;

[0016] The convolutional module comprises a plurality of dilated convolutions with different dilation rates in series.

[0017] Optionally, the multi-receptive field model comprises a first convolutional module, a second convolutional module, a third convolutional module, a fourth convolutional module and a fifth convolutional module in parallel;

[0018] The first convolutional module comprises four dilated convolutions with dilation rates of 1, 2, 3 and 6 in series, and the convolution kernel of the dilated convolution is 3*3;

[0019] The second convolutional module comprises four dilated convolutions with dilation rates of 1, 2, 5 and 7 in series, and the convolution kernel of the dilated convolution is 3*3;

[0020] The third convolutional module comprises four dilated convolutions with dilation rates of 1, 3, 5 and 9 in series, and the convolution kernel of the dilated convolution is 3*3;

[0021] The fourth convolutional module comprises four dilated convolutions with dilation rates of 1, 3, 7 and 11 in series, and the convolution kernel of the dilated convolution is 3*3;

[0022] The fifth convolutional module comprises four dilated convolutions with dilation rates of 1, 3, 8 and 13 in series, and the convolution kernel of the dilated convolution is 3*3.

[0023] Optionally, the preset cascade rule comprises:

[0024] input the second low-layer feature map and the first feature map processed by upsampling into a fusion model to obtain a first fusion feature map;

[0025] input the first low-layer feature map and the first fusion feature map processed by upsampling into a fusion model to obtain a second fusion feature map;

[0026] perform upsampling processing on the second fusion feature map to obtain the target detection result.

[0027] Optionally, the second low-level feature map and the first feature map after upsampling processing are input into a fusion model to obtain a first fusion feature map, including:

[0028] The first feature map after upsampling processing is enhanced through a channel attention mechanism to obtain a first enhanced feature map;

[0029] The second low-level feature map is enhanced through a position attention mechanism to obtain a second enhanced feature map;

[0030] The first enhanced feature map and the second enhanced feature map are fused to obtain the first fusion feature map; wherein the first fusion feature map is obtained through the following formula:

[0031]

[0032] Wherein, the X MSM is used to represent the first fusion feature map, the X l is used to represent the second low-level feature map, the X h is used to represent the first feature map, and the G PA (X l ) is used to represent the second enhanced feature map, and the G CA (X h ) is used to represent the first enhanced feature map.

[0033] Optionally, the first low-level feature map and the first fusion feature map after upsampling processing are input into a fusion model to obtain a second fusion feature map, including:

[0034] The first fusion feature map after upsampling processing is enhanced through a channel attention mechanism to obtain a third enhanced feature map;

[0035] The first low-level feature map is enhanced through a position attention mechanism to obtain a fourth enhanced feature map;

[0036] The third enhanced feature map and the fourth enhanced feature map are fused to obtain the second fusion feature map; wherein the second fusion feature map is obtained through the following formula:

[0037]

[0038] Wherein, the X' MSM is used to represent the second fusion feature map, the X' l is used to represent the first low-level feature map, the X' h is used to represent the first fusion feature map, and the G' PA (X' l) for characterizing the fourth enhanced feature map, the G' CA (X′ h ) for characterizing the third enhanced feature map.

[0039] In a second aspect, the embodiments of the present application further provide an infrared small target detection device, comprising:

[0040] an acquisition module, configured to acquire an infrared image of a target to be detected;

[0041] a first feature extraction module, configured to input the infrared image into a pre-constructed basic network structure for feature extraction, to obtain at least one low-level feature map and one high-level feature map;

[0042] a second feature extraction module, configured to input the high-level feature map into a multi-receptive field model to obtain a first feature map; wherein the multi-receptive field model is configured to expand a receptive field through a dilated convolution, to obtain the first feature map rich in semantic information;

[0043] a fusion module, configured to input the first feature map and the at least one low-level feature map into a fusion model according to a preset cascading rule, to obtain a target detection result.

[0044] Optionally, the second feature extraction module comprises at least two convolution modules in parallel, each of which has a different receptive field; and each of the convolution modules comprises dilated convolutions in series, each of which has a different dilation rate.

[0045] In a third aspect, the embodiments of the present application further provide a computing device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the infrared small target detection method of any one of the above aspects.

[0046] In a fourth aspect, the embodiments of the present application further provide a computer readable storage medium, which stores a computer program, and when the computer program is executed in a computer, the computer executes the infrared small target detection method of any one of the above aspects.

[0047] This invention provides an infrared small target detection method, apparatus, computing device, and storage medium. The method detects infrared small targets using a multi-receptive-field model and a fusion model that fuses data according to cascade rules. First, an infrared image is input into a basic network structure to obtain feature maps output from different layers. The high-level feature maps from the lower layers are then processed by the multi-receptive-field model to obtain a first feature map rich in semantic information within a large receptive field. Then, the fusion model is used to fuse the first feature map with low-level feature maps output from other layers of the basic network structure according to preset cascade rules to obtain the target detection result. In this way, the method fully utilizes small target information and background information, effectively acquires background features, and fuses contextual semantics, improving the accuracy of small target detection and reducing the false alarm rate. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of an infrared small target detection method provided in an embodiment of the present invention;

[0050] Figure 2 This invention provides an infrared image of a target to be detected and the corresponding target detection result, according to an embodiment of the present invention.

[0051] Figure 3 This is an infrared image of another target to be detected and its corresponding target detection result provided in an embodiment of the present invention;

[0052] Figure 4 This is another infrared image of a target to be detected and its corresponding target detection result provided by an embodiment of the present invention;

[0053] Figure 5 This is another infrared image of a target to be detected and its corresponding target detection result provided by an embodiment of the present invention;

[0054] Figure 6 This is a hardware architecture diagram of a computing device provided in an embodiment of the present invention;

[0055] Figure 7 This is a structural diagram of an infrared small target detection device provided in an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0057] Please refer to Figure 1 This invention provides an infrared small target detection method, which includes:

[0058] Step 100: Acquire an infrared image of the target to be detected;

[0059] Step 102: Input the infrared image into the pre-constructed basic network structure for feature extraction to obtain at least one low-level feature map and one high-level feature map.

[0060] Step 104: Input the high-level feature map into the multi-receptive-field model to obtain the first feature map; wherein, the multi-receptive-field model is used to expand the receptive field through dilated convolution to obtain the first feature map with rich semantic information.

[0061] Step 106: Input the first feature map and at least one low-level feature map into the fusion model according to the preset cascading rules to obtain the target detection result.

[0062] In this embodiment of the invention, infrared small targets are detected using a multi-receptive-field model and a fusion model that fuses data according to cascade rules. First, the infrared image is input into a basic network structure to obtain feature maps output from different layers. The high-level feature maps from the lower layers are then processed by the multi-receptive-field model to obtain a first feature map rich in semantic information within a large receptive field. Then, the fusion model is used to fuse the first feature map with low-level feature maps output from other layers of the basic network structure according to a preset cascade rule to obtain the target detection result. Thus, this method fully utilizes small target information and background information, effectively acquires background features, fuses contextual semantics, improves the accuracy of small target detection, and reduces the false alarm rate.

[0063] The following description Figure 1 The execution method of each step is shown.

[0064] First, for step 100, an infrared image of the target to be detected is acquired, specifically a single-frame infrared image of the small target to be detected.

[0065] For step 102, the infrared image is input into a pre-constructed basic network structure for feature extraction to obtain at least one low-level feature map and one high-level feature map, including:

[0066] The basic network structure includes a first convolutional layer, a second convolutional layer, and a third convolutional layer;

[0067] The infrared image is convolved using the first convolutional layer to obtain the first low-level feature map;

[0068] The second convolutional layer is used to perform a convolution operation on the first low-level feature map to obtain the second low-level feature map;

[0069] A third convolutional layer is used to perform a convolution operation on the second low-level feature map to obtain a high-level feature map.

[0070] It should be noted that the high-level feature maps are obtained from the outputs of the lower layers in the basic network structure. The difference between the number of convolutional layers in the basic network structure and the number of fusion operations in the fusion model is 1.

[0071] In this invention, the more convolutional layers the basic network structure includes, the more times the fusion model performs fusion, and the higher the accuracy of the output target detection results. The inventors discovered that when the basic network structure includes 3 convolutional layers, it can ensure both high accuracy of the target detection results and high efficiency of the detection method, thus shortening the detection time.

[0072] In step 104, the multi-receptive-field model includes at least two convolutional modules with different receptive fields connected in parallel;

[0073] The convolution module consists of cascaded dilated convolutions with different dilation rates.

[0074] In a preferred embodiment, the dilation rate of the dilated convolution in each convolutional module satisfies the following formula:

[0075] M i =max[M i+1 -2r i M i+1 -2(M i+1 -r i ),r i ], i < n; and M2 ≤ k, M n =r n (1)

[0076] Among them, M i M represents the maximum value of the dilation rate of the i-th dilated convolution in any convolutional module; i <M i+1 The kernel size of a dilated convolution is k*k; n represents the number of dilated convolutions cascaded in this convolutional module; r i This is used to represent the actual dilation rate of the i-th dilated convolution in a convolutional module.

[0077] In a preferred embodiment, in step 104, the multi-receptive field model includes a first convolutional module, a second convolutional module, a third convolutional module, a fourth convolutional module, and a fifth convolutional module connected in parallel;

[0078] The first convolutional module includes four dilated convolutions connected in series with dilation rates of 1, 2, 3, and 6, and the kernel of each dilated convolution is 3*3.

[0079] The second convolutional module includes four dilated convolutions connected in series with dilation rates of 1, 2, 5, and 7, and the kernel of each dilated convolution is 3*3.

[0080] The third convolution module includes four dilated convolutions connected in series with dilation rates of 1, 3, 5, and 9, and the kernel of each dilated convolution is 3*3.

[0081] The fourth convolutional module includes four dilated convolutions connected in series with dilation rates of 1, 3, 7, and 11, and the kernel of each dilated convolution is 3*3.

[0082] The fifth convolutional module includes four dilated convolutions connected in series with dilation rates of 1, 3, 8, and 13, and the kernel of each dilated convolution is 3*3.

[0083] Specifically, in step 104, when using a multi-receptive-field model to extract background features, the high-level feature map is first convolved once with a 1*1 convolution, and then five parallel convolutional modules are used. Each convolutional module uses a different dilation rate, so that the receptive field can utilize every pixel in the image and obtain background features at different scales and in different receptive fields, thereby enhancing the utilization rate of image information. Then, the five feature maps with different scales are concatenated using concat and then convolved once more with a 1*1 convolution, so that information from multiple receptive fields is fused, enriching the background features.

[0084] While existing ASPP structures also use multiple parallel dilated convolutions with different dilation rates to obtain background features at different scales, they suffer from gridding issues, hindering the extraction of effective image information. Therefore, the multi-receptive-field model of this invention, through cascaded dilated convolutions, performs denser pixel sampling on the feature map compared to the ASPP structure. It employs multiple dilated convolutions with different dilation rates for feature extraction, reducing the gridding problem while expanding the receptive field. This makes the network more effective at extracting multi-scale features, obtaining better background information, and improving the segmentation of small targets.

[0085] In this invention, a multi-receptive field model is used to extract background features, which effectively integrates feature information obtained under different receptive fields, effectively improves the utilization rate of background information, and thus enables accurate detection of small targets.

[0086] For step 106, the preset cascading rules include:

[0087] The second low-level feature map and the upsampled first feature map are input into the fusion model to obtain the first fused feature map;

[0088] The first low-level feature map and the first fused feature map after upsampling are input into the fusion model to obtain the second fused feature map;

[0089] The second fused feature map is upsampled to obtain the target detection result.

[0090] In step 106, the second low-level feature map and the upsampled first feature map are input into the fusion model to obtain the first fused feature map, including:

[0091] The first enhanced feature map is obtained by enhancing the first feature map after upsampling through a channel attention mechanism;

[0092] The second low-level feature map is enhanced by a positional attention mechanism to obtain the second enhanced feature map;

[0093] The first enhanced feature map and the second enhanced feature map are fused to obtain the first fused feature map; wherein the first fused feature map is obtained by the following formula:

[0094]

[0095] Among them, X MSM X is used to characterize the first fused feature map. l X is used to characterize the second low-level feature map. h G is used to characterize the first feature map. PA (X l G is used to characterize the second enhanced feature map. CA (X h ) is used to characterize the first enhanced feature map.

[0096] In step 106, the first low-level feature map and the first fused feature map after upsampling are input into the fusion model to obtain the second fused feature map, including:

[0097] The first fused feature map, after upsampling, is enhanced using a channel attention mechanism to obtain a third enhanced feature map;

[0098] The first low-level feature map is enhanced using a positional attention mechanism to obtain the fourth enhanced feature map;

[0099] The third and fourth enhanced feature maps are fused to obtain the second fused feature map; the second fused feature map is obtained by the following formula:

[0100]

[0101] Where, X′ MSM Used to characterize the second fused feature map, X′ l Used to characterize the first low-level feature map, X′ h Used to characterize the first fused feature map, G′ PA (X′ l ) is used to characterize the fourth enhanced feature map, G′ CA (X′ h ) is used to characterize the third enhanced feature map.

[0102] In this invention, a fusion model is used to fuse feature maps containing deep semantic information and feature maps containing low-level semantic information. Since low-level semantic information contains a large amount of target location information, a positional attention mechanism is first applied to the second low-level feature map to obtain an output second enhanced feature map. Deep semantic information contains more background information, and the background information contained in different channels is inconsistent. Therefore, a channel attention mechanism is applied to the first feature map to output a first enhanced feature map to pay attention to the differences between channels. Then, the first and second enhanced feature maps are fused to obtain a first fused feature map. Finally, the fused first feature map, as the feature map containing deep semantic information, is fused with the low-level feature map containing low-level semantic information output by the convolutional layer adjacent to the bottom layer in the basic network structure. This cascading loop continues until fusion with all low-level feature maps is achieved, and then the target detection result is output. Thus, the fusion model uses the positional attention mechanism to multiply with high-dimensional information, enhancing the small target information in the high-dimensional information. After adding it with low-dimensional information, it is multiplied with the channel attention mechanism, as shown in the above formulas (2) and (3), so that the focus is on the channel that needs attention, thereby improving the detection accuracy of small targets.

[0103] It should be noted that during the upsampling stage, the proposed fusion model guides the fusion of high-resolution images using low-resolution images to obtain better small target localization information. The high-resolution images are weighted according to different channels, improving the network's utilization of important features. The final binary image output by the fusion model is the target detection result for infrared small targets, including the position and shape of the infrared small targets.

[0104] It should be noted that in this invention, by changing the size of the infrared image of the target to be detected (input_size), the detection speed and detection accuracy can be adjusted according to the actual application requirements.

[0105] In a preferred embodiment, the infrared small target detection method employs a basic network structure including three convolutional layers, a multi-receptive field model including five parallel convolutional modules as described in the above embodiment, and a fusion model based on the above formulas (2) and (3). Based on this infrared small target detection method, four targets are detected to obtain their corresponding detection results, such as... Figures 2 to 4 As shown. Figure 2 and Figure 3 The image shows the original infrared image of a small target existing at the edge of an object (left image) and the target detection result (right image). Figure 4 and Figure 5 The original infrared image (left) of a small target existing in background clutter and its target detection result (right) are shown. Obviously, when detecting the edge of an object and background clutter, although the existing technology has the problem of imperfect background extraction and often produces a lot of false alarms, the infrared small target detection method of the present invention can still achieve good detection results.

[0106] like Figure 6 , Figure 7 As shown, this embodiment of the invention provides an infrared small target detection device. The device embodiment can be implemented through software, hardware, or a combination of both. From a hardware perspective, as... Figure 6 The diagram shown is a hardware architecture diagram of a computing device housing an infrared small target detection device provided in an embodiment of the present invention. (Except for...) Figure 6 In addition to the processor, memory, network interface, and non-volatile memory shown, the computing device in the embodiment may also include other hardware, such as a forwarding chip responsible for processing packets. Taking software implementation as an example, such as... Figure 7 As shown, as a logical device, it is formed by the CPU of its computing device reading the corresponding computer program from the non-volatile memory into the memory for execution. This embodiment provides an infrared small target detection device, including: an acquisition module 700, a first feature extraction module 702, a second feature extraction module 704, and a fusion module 706;

[0107] The acquisition module 700 is used to acquire infrared images of the target to be detected;

[0108] The first feature extraction module 702 is used to input the infrared image into a pre-constructed basic network structure for feature extraction to obtain at least one low-level feature map and one high-level feature map.

[0109] The second feature extraction module 704 is used to input the high-level feature map into the multi-receptive field model to obtain the first feature map; wherein, the multi-receptive field model is used to expand the receptive field through dilated convolution to obtain the first feature map with rich semantic information.

[0110] The fusion module 706 is used to input the first feature map and at least one low-level feature map into the fusion model according to a preset cascading rule to obtain the target detection result.

[0111] In some specific implementations, the acquisition module 700 can be used to perform the above step 100, the first feature extraction module 702 can be used to perform the above step 102, the second feature extraction module 704 can be used to perform the above step 104, and the fusion module 706 can be used to perform the above step 106.

[0112] In some specific implementations, the basic network structure includes a first convolutional layer, a second convolutional layer, and a third convolutional layer; the first feature extraction module 702 is also used to perform the following operations:

[0113] The infrared image is convolved using the first convolutional layer to obtain the first low-level feature map;

[0114] The second convolutional layer is used to perform a convolution operation on the first low-level feature map to obtain the second low-level feature map;

[0115] A third convolutional layer is used to perform a convolution operation on the second low-level feature map to obtain a high-level feature map.

[0116] In some specific implementations, the multi-receptive-field model includes at least two convolutional modules with different receptive fields connected in parallel.

[0117] The convolution module consists of cascaded dilated convolutions with different dilation rates;

[0118] In each convolutional module, the dilation rate of the dilated convolution satisfies the following formula:

[0119] M i =max[M i+1 -2r i M i+1 -2(M i+1 -r i ),r i ], i < n; and M2 ≤ k, M n =r n

[0120] Among them, M i M represents the maximum value of the dilation rate of the i-th dilated convolution in any convolutional module; i <M i+1 The kernel size of a dilated convolution is k*k; n represents the number of dilated convolutions cascaded in this convolutional module; r i This is used to represent the actual dilation rate of the i-th dilated convolution in a convolutional module.

[0121] In some specific implementations, the multi-receptive-field model includes a first convolutional module, a second convolutional module, a third convolutional module, a fourth convolutional module, and a fifth convolutional module connected in parallel;

[0122] The first convolutional module includes four dilated convolutions connected in series with dilation rates of 1, 2, 3, and 6, and the kernel of each dilated convolution is 3*3.

[0123] The second convolutional module includes four dilated convolutions connected in series with dilation rates of 1, 2, 5, and 7, and the kernel of each dilated convolution is 3*3.

[0124] The third convolution module includes four dilated convolutions connected in series with dilation rates of 1, 3, 5, and 9, and the kernel of each dilated convolution is 3*3.

[0125] The fourth convolutional module includes four dilated convolutions connected in series with dilation rates of 1, 3, 7, and 11, and the kernel of each dilated convolution is 3*3.

[0126] The fifth convolutional module includes four dilated convolutions connected in series with dilation rates of 1, 3, 8, and 13, and the kernel of each dilated convolution is 3*3.

[0127] In some specific implementations, the fusion module 706 is also used to perform the following operations:

[0128] The first enhanced feature map is obtained by enhancing the first feature map after upsampling through a channel attention mechanism;

[0129] The second low-level feature map is enhanced by a positional attention mechanism to obtain the second enhanced feature map;

[0130] The first enhanced feature map and the second enhanced feature map are fused to obtain the first fused feature map; wherein the first fused feature map is obtained by the following formula:

[0131]

[0132] Among them, X MSM X is used to characterize the first fused feature map. l X is used to characterize the second low-level feature map. h G is used to characterize the first feature map. PA (X l G is used to characterize the second enhanced feature map. CA (X h ) is used to characterize the first enhanced feature map;

[0133] The first fused feature map, after upsampling, is enhanced using a channel attention mechanism to obtain a third enhanced feature map;

[0134] The first low-level feature map is enhanced using a positional attention mechanism to obtain the fourth enhanced feature map;

[0135] The third and fourth enhanced feature maps are fused to obtain the second fused feature map; the second fused feature map is obtained by the following formula:

[0136]

[0137] Where, X′ MSM Used to characterize the second fused feature map, X′ l Used to characterize the first low-level feature map, X′ h Used to characterize the first fused feature map, G′ PA (X′ l ) is used to characterize the fourth enhanced feature map, G′ CA (X′ h ) is used to characterize the third enhanced feature map;

[0138] The second fused feature map is upsampled to obtain the target detection result.

[0139] It is understood that the structures illustrated in the embodiments of the present invention do not constitute a specific limitation on an infrared small target detection device. In other embodiments of the present invention, an infrared small target detection device may include more or fewer components than illustrated, or combine some components, or split some components, or arrange different components. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0140] The information interaction and execution process between the modules in the above-mentioned device are based on the same concept as the method embodiment of the present invention, and the specific details can be found in the description of the method embodiment of the present invention, and will not be repeated here.

[0141] This invention also provides a computing device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements an infrared small target detection method according to any embodiment of this invention. For example, the computer program is implemented using Python programming and can run on a 1060Ti graphics card.

[0142] This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform an infrared small target detection method according to any embodiment of this invention.

[0143] Specifically, a system or apparatus equipped with a storage medium may be provided, on which software program code implementing the functions of any of the embodiments described above is stored, and the computer (or CPU or MPU) of the system or apparatus may read and execute the program code stored in the storage medium.

[0144] In this case, the program code read from the storage medium can itself implement the function of any of the above embodiments, and therefore the program code and the storage medium storing the program code constitute part of the present invention.

[0145] Examples of storage media used to provide program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer via a communication network.

[0146] Furthermore, it should be clear that not only can the program code read by the computer be executed, but also the operating system or other components operating on the computer can be instructed based on the program code to perform some or all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0147] Furthermore, it is understood that the program code read from the storage medium is written to the memory set in the expansion board inserted into the computer or to the memory set in the expansion module connected to the computer. Then, based on the instructions of the program code, the CPU or other components installed on the expansion board or expansion module execute some and all of the actual operations, thereby realizing the function of any of the above embodiments.

[0148] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0149] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as ROM, RAM, magnetic disk, or optical disk.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting small infrared targets, characterized in that, The method comprises the following steps: acquiring an infrared image of a target to be detected; inputting the infrared image into a pre-constructed basic network structure for feature extraction to obtain at least one low-level feature map and one high-level feature map; inputting the high-level feature map into a multi-receptive field model to obtain a first feature map; wherein the multi-receptive field model is used to expand the receptive field through a hollow convolution to obtain the first feature map rich in semantic information; the multi-receptive field model comprises at least two convolution modules with different receptive fields connected in parallel; the convolution module comprises hollow convolutions with different dilation rates connected in series; inputting the first feature map and the at least one low-level feature map into a fusion model according to a preset cascade rule to obtain a target detection result; the step of inputting the infrared image into the pre-constructed basic network structure for feature extraction to obtain at least one low-level feature map and one high-level feature map comprises: the basic network structure comprises a first convolution layer, a second convolution layer and a third convolution layer; performing convolution operation on the infrared image by using the first convolution layer to obtain a first low-level feature map; performing convolution operation on the first low-level feature map by using the second convolution layer to obtain a second low-level feature map; performing convolution operation on the second low-level feature map by using the third convolution layer to obtain a high-level feature map; the preset cascade rule comprises: inputting the second low-level feature map and the first feature map processed by upsampling into a fusion model to obtain a first fusion feature map; inputting the first low-level feature map and the first fusion feature map processed by upsampling into a fusion model to obtain a second fusion feature map; performing upsampling processing on the second fusion feature map to obtain the target detection result; the step of inputting the second low-level feature map and the first feature map processed by upsampling into a fusion model to obtain a first fusion feature map comprises: enhancing the first feature map processed by upsampling through a channel attention mechanism to obtain a first enhanced feature map; enhancing the second low-level feature map through a position attention mechanism to obtain a second enhanced feature map; fusing the first enhanced feature map and the second enhanced feature map to obtain the first fusion feature map; wherein the first fusion feature map is obtained through the following formula: wherein the X MSM for representing the first fusion feature map, the X l for representing the second low-level feature map, the X h for representing the first feature map, the G PA ( X l ) for representing the second enhanced feature map, the G CA ( X h ) for representing the first enhanced feature map.

2. The method of claim 1, wherein, the multi-receptive field model comprises a first convolution module, a second convolution module, a third convolution module, a fourth convolution module and a fifth convolution module connected in parallel; the first convolution module comprises four hollow convolutions with dilation rates of 1, 2, 3 and 6 connected in series, and the convolution kernels of the hollow convolutions are all 3*3; the second convolution module comprises four hollow convolutions with dilation rates of 1, 2, 5 and 7 connected in series, and the convolution kernels of the hollow convolutions are all 3*3; the third convolution module comprises four hollow convolutions with dilation rates of 1, 3, 5 and 9 connected in series, and the convolution kernels of the hollow convolutions are all 3*3; the fourth convolution module comprises four hollow convolutions with dilation rates of 1, 3, 7 and 11 connected in series, and the convolution kernels of the hollow convolutions are all 3*3; The fifth convolution module includes four hole convolutions with expansion rates of 1, 3, 8 and 13 in sequence, and the convolution kernels of the hole convolutions are all 3*3.

3. The method of claim 1, wherein, The first low-level feature map and the first fusion feature map after up-sampling are input into a fusion model to obtain a second fusion feature map. The first fusion feature map after up-sampling is enhanced through a channel attention mechanism to obtain a third enhanced feature map. The first low-level feature map is enhanced through a position attention mechanism to obtain a fourth enhanced feature map. The third enhanced feature map and the fourth enhanced feature map are fused to obtain the second fusion feature map; and the second fusion feature map is obtained through the following formula: wherein the for representing the second fusion feature map, the for representing the first low-level feature map, the for representing the first fusion feature map, the for representing the fourth enhanced feature map, the for representing the third enhanced feature map.

4. An infrared small target detection device, characterized in that, A method as claimed in any one of claims 1 to 3 is implemented by comprising: An acquisition module is configured to acquire an infrared image of a target to be detected. A first feature extraction module is configured to input the infrared image into a pre-constructed basic network structure for feature extraction to obtain at least one low-level feature map and one high-level feature map. A second feature extraction module is configured to input the high-level feature map into a multi-receptive field model to obtain a first feature map; the multi-receptive field model is configured to expand a receptive field through a hole convolution to obtain the first feature map with rich semantic information; the second feature extraction module includes at least two convolution modules with different receptive fields in parallel; and the convolution module includes hole convolutions with different expansion rates in series. A fusion module is configured to input the first feature map and the at least one low-level feature map into a fusion model according to a preset cascading rule to obtain a target detection result.

5. A computing device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method of any one of claims 1-3.

6. A computer readable storage medium, which stores a computer program, and the computer program, when executed in a computer, causes the computer to execute the method of any one of claims 1-3.

Images