An Infrared Small Target Detection Method Based on Double-Layer Feature Fusion Convolutional Network

Abstract

The present invention discloses an infrared small target detection method based on a double-layer feature fusion convolutional network. The steps of the detection method are as follows: perform target contrast enhancement preprocessing on the infrared image data training set and test set containing small targets; establish a double-layer feature fusion convolutional network, and adopt a fully convolutional network structure and a T-shaped fusion structure with double-layer parallel interaction and fusion between network layer N1 and upsampling network layer N2 to realize the information fusion of high-level features and low-level features; use the above network to learn the features of small targets in the training set, construct a loss function of the double-layer feature fusion convolutional network to calculate the loss of the network, and realize the training of network parameters; use the learned convolutional network to process the infrared image test set, and can adaptively judge the infrared image to detect small targets from different scene image data. The method can ensure the effective extraction of image features, reduce redundant background information, improve the detection rate and accuracy of the image, and has universality.

Description

Technical Field

[0001] The present invention relates to the field of infrared detection, and particularly to an infrared small target detection method based on a double-layer feature fusion convolutional network. Background Art

[0002] Different from visible light images, infrared images are images of the surface temperature distribution of objects. In a dark night or a harsh environment with thick smoke and dust, the target and the background can also be distinguished according to the radiation difference. Based on the above characteristics, infrared images have obvious advantages in target detection, tracking, recognition, etc., and have been widely used in modern military, space-based detection, security monitoring and other fields. However, in most applications, the infrared imaging distance from the target is far, and the target imaging is small, often only a few pixels to dozens of pixels, belonging to small targets. And because infrared images are obtained by "measuring" the heat radiated by objects, compared with visible light images, infrared images have the following deficiencies: 1. Low image resolution; 2. Low image contrast; 3. Low image signal-to-noise ratio; 4. There is no linear relationship between the gray distribution of the image and the reflection characteristics of the target. In the field of infrared small target detection, these deficiencies make the target lack texture, color and shape information, and are easily submerged in the background and noise. Especially in a complex background, false detections are extremely likely to occur. Therefore, infrared small target detection has become a very challenging task. At present, for the infrared small target detection task, traditional detection methods mostly use the differences between small targets and the background, such as features such as gray scale, gradient, and contrast in the image to achieve small target detection. However, most of the features selected by these algorithms are based on manual selection. Therefore, the feature parameters considered in the algorithm design are always limited, and better generality cannot be obtained. And the method based on the convolutional neural network (CNN) can extract the features in the data through training for the characterization of the target, and has good generality and stability. For example, the DNAnet network designs a densely nested and interactively fused network, achieving a better detection rate for infrared small targets in different scenarios; the MDvsFA network divides the small target detection task into two subtasks of suppressing false detections and reducing false alarms, and introduces adversarial learning to achieve a delicate balance between two competing criteria. However, most current CNN algorithms do not make full use of the information obtained by the network itself, resulting in waste of resources, and there are a large number of misjudgments and false detections for small and weak targets with low signal-to-noise ratio in a complex background. Summary of the Invention

[0003] Aiming at the above problems, the present invention proposes an infrared small target detection method based on a double-layer feature fusion convolutional network, which can ensure the extraction of image features, reduce redundant background information, effectively improve the detection rate and accuracy of the image, and has universality for the small target detection of infrared images.

[0004] To this end, the present invention adopts the following technical solutions:

[0005] An infrared small target detection method based on a double-layer feature fusion convolutional network, as Figure 1 shown, includes the following steps:

[0006] Step 1. Data preprocessing: perform target contrast enhancement preprocessing on the infrared image data training set and test set of different scenes containing small targets.

[0007] Step 2. Establish a double-layer feature fusion convolutional network, adopting a fully convolutional network structure with double-layer parallel interaction and fusion of network layer N1 and upsampling network layer N2. Share the information of the two network layers through an interactive connection method. Network layer N1 fuses the image feature extraction results into the network through convolution and upsampling, and upsampling network layer N2 fuses the image feature extraction results into the network through convolution and downsampling, realizing the interactive fusion of the detection information of network layer N1 and upsampling network layer N2; the information fusion method of network layer N1 and upsampling network layer N2 adopts a T-shaped fusion structure to realize the information fusion of high-level features and low-level features of the convolutional network, and ensure the utilization rate of the network model.

[0008] Step 3. Network training: Use the above double-layer feature fusion convolutional network to learn the features of small targets in the infrared image data training set of different scenes preprocessed in Step 1, construct a loss function of the double-layer feature fusion convolutional network to calculate the loss of the network, and realize the training of network parameters.

[0009] Step 4. Target detection: Use the learned double-layer feature fusion convolutional network to process the infrared image test set, and can make an adaptive judgment on the infrared image to detect small targets from the image data of different scenes.

[0010] Among them, the target contrast enhancement preprocessing method in Step 1 includes the following steps:

[0011] 1) Select Q original infrared images of different scenes as image data, with the image size of A×B, the pixel value distribution probability range of the image is 0-100%, and the pixel saturation brightness range is 0-1; the original infrared image scenes include different scenes such as clouds, seas, land, and buildings, and the number of training set images is not less than 500.

[0012] 2) By changing the image pixel values, set the pixels with a high pixel value distribution probability less than the 1% threshold in the image to high brightness 1, and set the pixels with a low pixel value distribution probability not less than the 20% threshold to low brightness 0; for the part of the original image excluding the above 1% high pixels and 20% low pixels, map it proportionally according to the change of pixel values, and calculate the pixel value I n '(x,y) of the nth image after target contrast enhancement according to the following formula:

[0013]

[0014] Among them, I n (x, y) is the pixel value of the nth original image at (x, y), where x = 1, 2, ……, A - 1, A, y = 1, 2, ……, B - 1, B, and n = 1, 2, ……, Q - 1, Q; ε is the brightness value at which the distribution probability of low pixel values in the original image is not less than the 20% threshold; θ is the brightness value at which the distribution probability of high pixel values in the original image is less than the 1% threshold; a is the proportionality coefficient; others is the part of the original image excluding the above 1% high pixels and 20% low pixels.

[0015] Among them, as Figure 2 、 3 、as shown in Figure 4, the convolution kernels of the network layer N1 and the upsampling network layer N2 are arranged in an axisymmetric form. The network layer N1 is composed of twelve cascaded dilated convolution modules and four T-shaped fusion convolution structures. The twelve dilated convolution modules all use 3×3 convolution kernels, and their dilation factors are 1, 3, 5, 7, 9, 11, 9, 7, 5, 3, 1, 1 in the transmission order; the four T-shaped fusion convolution structures also use 3×3 convolution kernels, and their dilation factors are 1, 3, 5, 7 respectively. The upsampling network layer N2 is composed of twelve cascaded dilated convolution modules and four T-shaped fusion convolution structures. The twelve dilated convolution modules all use 3×3 convolution kernels, and their dilation factors are 2, 5, 8, 11, 14, 17, 14, 11, 8, 5, 2, 1 in the transmission order; the four T-shaped fusion convolution structures also use 3×3 convolution kernels, and their dilation factors are 2, 5, 8, 11 respectively.

[0016] Among them, the network layer N1 and the upsampling network layer N2 perform dilated convolution according to the following formulas respectively:

[0017]

[0018] Among them, F j (g) is the result of dilated convolution of the discrete function g at the dilation factor j; g ∈ N, where N is a non-negative number; * is the convolution symbol; k is the discrete filter; s is the convolution range of the discrete function g, t is the convolution step of the filter k; p is the time value; let the discrete filter k: Ω r ∈R, with a size of (2r + 1) 2 , where Ω r is the passband bandwidth of the filter, r is the passband bandwidth radius, and R is the set of real numbers.

[0019] Among them, as Figure 2 、 5As shown, the information fusion of the network layer N1 and the upsampling network layer N2 adopts a T-shaped fusion structure, and the fusion convolution module of the network layer N1 and the upsampling network layer N2 fuses the shallow features and high-level features of the network according to the following formula:

[0020] y l = F j (γ l-2 ) + F j {F j' (γ l )}

[0021] where y l is the fusion feature of the l-th level of the network; F j (γ l-2 ) is the convolution feature of the (l - 2)-th level of the image when the dilation factor is j; F j {F j' (γ l )} is the feature after convolution of the l-th level of the image when the dilation factor is j, F j' (γ) is the convolution feature of the l-th level of the image when the dilation factor before fusion is j'; γ l is the information feature of the l-th level generated by the network, γ l-2 is the information feature of the (l - 2)-th level generated by the network, j' is the dilation factor before the l-th level of fusion, and j is the dilation factor during the l-th level of fusion.

[0022] Among them, the double-layer feature fusion convolution network maps the input image I from the first dilated convolution module of the network layer N1 and the upsampling network layer N2 to the twelfth dilated convolution module, forming the following network model:

[0023]

[0024] Then the output of the double-layer feature fusion convolution network is:

[0025] y(I) = F j {N1(I) + N2(I)}

[0026] where N1(I) is the result of the network layer N1 performing a convolution operation on the image I, and N2(I) is the result of the upsampling network layer N2 performing a convolution operation on the image I; F 1 j,1 , F 1 j,2 , ……, F 1 j,12 are respectively twelve cascaded dilated convolution modules of the network layer N1; F 2 j,1 , F 2 j,2 , ……, F 2j,12 They are twelve serially connected dilated convolution modules of the upsampling network layer N2 respectively.

[0027] Among them, the double-layer feature fusion convolutional network loss function DLFF_Loss constructed in step 3. consists of the D_Loss loss function and the F_Loss loss function, where

[0028] 1) The D_Loss loss function is:

[0029]

[0030] Among them, is the predicted target detection result, and y is the true target detection result;

[0031] 2) The F_Loss loss function is:

[0032] F_Loss = α t (1 - p t ) γ log(P t )

[0033] Among them, α t is the weight coefficient, γ is the focusing coefficient, where P t is the parameter p related to the binary cross-entropy loss function BCE_Loss, t = e -BCE_Loss , and the binary cross-entropy loss function BCE_Loss is:

[0034]

[0035] Among them, is the predicted target detection result, and y is the true target detection result;

[0036] 3) The double-layer feature fusion convolutional network loss function DLFF_Loss is:

[0037] DLFF_Loss = β0[D_Loss + (1 - β1)F_Loss]

[0038] Among them, β1 is the weight coefficient of the loss function, and β0 is the experimental coefficient.

[0039] The present invention proposes an infrared small target detection method based on a double-layer feature fusion convolutional network. By constructing a convolutional network with double-layer feature fusion, through the parallel interaction and fusion of the downsampling network layer N1 and the upsampling network layer N2, the target features can be effectively enhanced while the background is weakened, thus ensuring the detection rate and accuracy of small targets in infrared images. Moreover, the computational complexity is low and the automation degree is high. Through learning from the training set, the convolutional network can adaptively judge infrared images and detect small targets from different scene data. It is applicable to the detection of small and weak targets in complex backgrounds with low signal-to-noise ratios, and is not prone to false positives and missed detections, having universality. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Schematic diagram of the infrared small target detection process based on the double-layer feature fusion convolutional network of the present invention.

[0041] Figure 2 Schematic diagram of the structure of the double-layer feature fusion convolutional network of the present invention.

[0042] Figure 3 Schematic diagram of the structure of the network layer N1 of the double-layer feature fusion convolutional network of the present invention.

[0043] Figure 4 Schematic diagram of the structure of the upsampling network layer N2 of the double-layer feature fusion convolutional network of the present invention.

[0044] Figure 5 Schematic diagram of the T-shaped fusion structure of the information of the network layer N1 and the upsampling network layer N2 of the present invention.

[0045] Figure 6 Original example infrared images of different scenes of clouds, sea, land and buildings selected in the specific implementation manner of the present invention.

[0046] Figure 7 Schematic diagram of the comparison of the enhancement results of the original example infrared images of different scenes of clouds, sea, land and buildings selected in the specific implementation manner of the present invention.

[0047] Figure 8 The infrared image to be tested in the test set and its three-dimensional display in the specific implementation manner of the present invention.

[0048] Figure 9 Diagram and three-dimensional display of the small target in the infrared image to be tested obtained by the infrared small target detection method based on the double-layer feature fusion convolutional network of the present invention in the specific implementation manner of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0049] To make the objectives, features, and advantages of the present invention clearer, a more detailed description of a specific embodiment of the present invention is provided. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described, and thus, the present invention is not limited by the specific embodiments disclosed below.

[0050] Taking the original infrared images of different scenes such as clouds, the sea, land, and buildings as examples, a specific implementation manner is given according to the infrared small target detection method based on the double-layer feature fusion convolutional network of the present invention.

[0051] 1. Data preprocessing: Perform target contrast enhancement preprocessing on the infrared image data training set and test set of different scenes containing small targets.

[0052] First, the existing commonly used MDFA dataset was screened, images that did not meet the small targets were removed, and SIRST was added, improving the quality of the dataset. Finally, a dataset with 6000 training images and 100 test images was obtained. 500 images were selected for the experiment. As Figure 6 shown, the backgrounds of the images include clouds, the sea, land, and buildings. The image size is A×B, the set range of the pixel value distribution probability of the image is 0 - 100%, and the pixel saturation brightness range is 0 - 1;

[0053] By changing the pixel values of the image, pixels with a high pixel value distribution probability less than the 1% threshold are set to high brightness 1, and pixels with a low pixel value distribution probability not less than the 20% threshold are set to low brightness 0; for the part of the original image excluding the above 1% high pixels and 20% low pixels, it is mapped proportionally according to the change in pixel values. The pixel value I' n (x,y) of the nth image after target contrast enhancement is calculated according to the following formula:

[0054]

[0055] where, I n (x,y) is the pixel value of the nth original image at (x,y), x = 1, 2, ……, A - 1, A, y = 1, 2, ……, B - 1, B, n = 1, 2, ……, Q - 1, Q; ε is the brightness value of the low pixel value distribution probability not less than the 20% threshold in the original image; θ is the brightness value of the high pixel value distribution probability less than the 1% threshold in the original image; a is the proportionality coefficient; others is the part of the original image excluding the above 1% high pixels and 20% low pixels. The infrared image data of different scenes such as clouds, the sea, land, and buildings containing small targets are shown in terms of the enhancement effect as Figure 7 shown.

[0056] As Figure 1As shown, the image data of the training set passes through a conv1×1 network layer before being input into the double-layer feature fusion convolutional network, obtaining 32 target feature extraction images. The 1×1 convolution size is mainly to ensure the integrity of image feature detail extraction as much as possible, can extract rich features of the image as much as possible, and reduce redundant information.

[0057] 2. Establish a double-layer feature fusion convolutional network: Adopt a fully convolutional network structure with double-layer parallel interaction and fusion of network layer N1 and upsampling network layer N2. Share the information of the two network layers through an interactive connection method. Network layer N1 fuses the image feature extraction result into the network through convolutional upsampling, and upsampling network layer N2 fuses the image feature extraction result into the network through convolutional downsampling to achieve the interactive fusion of the detection information of network layer N1 and upsampling network layer N2; The information fusion method of network layer N1 and upsampling network layer N2 adopts a T-shaped fusion structure to achieve the information fusion of high-level features and shallow-level features of the convolutional network and ensure the utilization rate of the network model.

[0058] Among them, the convolutional kernels of network layer N1 and upsampling network layer N2 are arranged in an axisymmetric form. Network layer N1 consists of twelve cascaded dilated convolutional modules and four T-shaped fusion convolutional structures. All the twelve dilated convolutional modules adopt 3×3 convolutional kernels, and their dilation factors are 1, 3, 5, 7, 9, 11, 9, 7, 5, 3, 1, 1 in the transmission order; The four T-shaped fusion convolutional structures also adopt 3×3 convolutional kernels, and their dilation factors are 1, 3, 5, 7 respectively. The smaller the convolutional kernel size, the fewer parameters and the less computational amount required. CNN generally uses 3×3 as the convolutional kernel size. However, the smaller the convolutional kernel of the traditional convolutional network, the smaller the receptive field. To ensure that the information used is global rather than just local information, a larger receptive field should be ensured. The dilation factor of the dilated convolution determines the size of the receptive field. For this reason, the present invention has done a large number of experiments and finally determined the combination method of the network model. The upsampling network layer N2 consists of twelve cascaded dilated convolutional modules and four T-shaped fusion convolutional structures. All the twelve dilated convolutional modules adopt 3×3 convolutional kernels, and their dilation factors are 2, 5, 8, 11, 14, 17, 14, 1, 8, 5, 2, 1 in the transmission order; The four T-shaped fusion convolutional structures also adopt 3×3 convolutional kernels, and their dilation factors are 2, 5, 8, 11 respectively. The design idea of the upsampling network layer N2 is the same as that of the network layer N1. Higher spatial resolution is crucial. At low resolution, small objects may be lost or multiple targets may be over-segmented into single objects. Upsampling is the most effective method for improving resolution in image processing. For this reason, the present invention introduces a parallel-running upsampling network into the network. Use the transposed convolution network to perform upsampling processing on the image, and finally double the resolution of the original image.

[0059] Among them, the network layer N1 and the upsampling network layer N2 perform dilated convolution according to the following formulas respectively:

[0060]

[0061] Among them, F j (g) is the result of dilated convolution of the discrete function g at the dilation factor j; g ∈ N, and N is a non - negative number; * is the convolution symbol; k is the discrete filter; s is the convolution range of the discrete function g, t is the convolution step of the filter k; p is the time value; let the discrete filter k: Ω r ∈R, with a size of (2r + 1) 2 , where Ω r is the pass - band bandwidth of the filter, r is the pass - band bandwidth radius, and R is the set of real numbers.

[0062] Among them, the information fusion of the network layer N1 and the upsampling network layer N2 adopts a T - type fusion structure. Through the fusion convolution module of the network layer N1 and the upsampling network layer N2, the shallow - layer features and high - layer features of the network are fused according to the following formula:

[0063] y l = F j (γ l-2 ) + F j {F j' (γ l )}

[0064] Among them, y l is the fusion feature of the l - th level of the network; F j (γ l-2 ) is the convolution feature of the (l - 2) - th level of the image at the dilation factor j; F j {F j' (γ l )} is the feature after convolution of the l - th level of the image at the dilation factor j, F j' (γ) is the convolution feature of the l - th level of the image at the dilation factor j' before fusion; γ l is the information feature generated by the network at the l - th level, γ l-2 is the information feature generated by the network at the (l - 2) - th level, j' is the dilation factor before the l - th level of fusion, and j is the dilation factor during the l - th level of fusion.

[0065] Step 3. Network training: Use the above - mentioned double - layer feature fusion convolutional network to learn the features of small targets in the infrared image data training set of different scenarios pre - processed in Step 1, construct a loss function of the double - layer feature fusion convolutional network to calculate the loss of the network, and realize the training of network parameters;

[0066] The loss function DLFF_Loss of the constructed double-layer feature fusion convolutional network is composed of the D_Loss loss function and the F_Loss loss function, where

[0067] 1) The D_Loss loss function is:

[0068]

[0069] where, is the predicted target detection result, and y is the true target detection result;

[0070] 2) The F_Loss loss function is:

[0071] F_Loss = α t (1 - p t ) γ log(P t )

[0072] where, α t is the weight coefficient, γ is the focusing coefficient, where P t is the parameter p related to the binary cross-entropy loss function BCE_Loss t = e -BCE_Loss , and the binary cross-entropy loss function BCE_Loss is:

[0073]

[0074] where, is the predicted target detection result, and y is the true target detection result;

[0075] 3) The loss function DLFF_Loss of the double-layer feature fusion convolutional network is:

[0076] DLFF_Loss = β0[D_Loss + (1 - β1)F_Loss]

[0077] where, β1 is the weight coefficient of the loss function, and β0 is the experimental coefficient.

[0078] Step 4. Target detection. Using the learned convolutional network to process the infrared image test set, as Figure 8 shown, the infrared image can be adaptively judged, and small targets can be detected from different scene image data, as Figure 9 shown. It can be seen that by using the infrared small target detection method based on the double-layer feature fusion convolutional network of the present invention, the target features in the infrared image can be effectively enhanced, while the background will be weakened, thus ensuring the detection rate and accuracy of small targets in the infrared image. It is applicable to the detection of small and weak targets with complex backgrounds and low signal-to-noise ratios, and is not prone to misjudgment and missed detection, and has universality.

[0079] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. An infrared small target detection method based on a double-layer feature fusion convolutional network, characterized in that It includes the following steps: Step 1. Data preprocessing: perform target contrast enhancement preprocessing on the infrared image data training set and test set of different scenarios containing small targets; Step 2. Establish a double-layer feature fusion convolutional network, adopting a fully convolutional network structure with double-layer parallel interaction and fusion between network layer N1 and upsampling network layer N2. Share the information of the two network layers through an interactive connection method. Network layer N1 fuses the image feature extraction results into the network through convolution and upsampling, and upsampling network layer N2 fuses the image feature extraction results into the network through convolution and downsampling, realizing the interactive fusion of the detection information of network layer N1 and upsampling network layer N2; The information fusion method between network layer N1 and upsampling network layer N2 adopts a T-shaped fusion structure to realize the information fusion of the high-level features and low-level features of the convolutional network, ensuring the utilization rate of the network model; The convolutional kernels of network layer N1 and upsampling network layer N2 are arranged in an axisymmetric form. Network layer N1 consists of twelve cascaded dilated convolution modules and four T-shaped fusion convolution structures. All the twelve dilated convolution modules adopt 3×3 convolutional kernels, and their dilation factors are 1, 3, 5, 7, 9, 11, 9, 7, 5, 3, 1, 1 in the transmission order; The four T-shaped fusion convolution structures also adopt 3×3 convolutional kernels, and their dilation factors are 1, 3, 5, 7 respectively; Upsampling network layer N2 consists of twelve cascaded dilated convolution modules and four T-shaped fusion convolution structures. All the twelve dilated convolution modules adopt 3×3 convolutional kernels, and their dilation factors are 2, 5, 8, 11, 14, 17, 14, 11, 8, 5, 2, 1 in the transmission order; The four T-shaped fusion convolution structures also adopt 3×3 convolutional kernels, and their dilation factors are 2, 5, 8, 11 respectively; Step 3. Network training: use the above double-layer feature fusion convolutional network to learn the features of small targets in the infrared image data training set preprocessed in Step 1, construct a loss function for the double-layer feature fusion convolutional network to calculate the loss of the network, and realize the training of network parameters; Step 4. Target detection: use the learned double-layer feature fusion convolutional network to process the infrared image test set, and can make an adaptive judgment on the infrared image to detect small targets from the image data of different scenarios.

2. The infrared small target detection method based on a double-layer feature fusion convolutional network according to claim 1, wherein, The target contrast enhancement preprocessing method described in Step 1 includes the following steps: 1) Select Q original infrared images of different scenarios as image data, with the image size of A×B. The pixel value distribution probability range of the image is 0-100%, and the pixel saturation brightness range is 0-1; 2) By changing the image pixel values, pixels with a high pixel value distribution probability less than the 1% threshold in the image are set to high brightness 1, and pixels with a low pixel value distribution probability less than the 20% threshold are set to low brightness 0; for the part of the original image excluding the above 1% of high pixels and 20% of low pixels, according to the equal ratio mapping of pixel value changes, the pixel value I of the nth image after target contrast enhancement is calculated according to the following formula n '(x,y): Among them, I n (x, y) is the pixel value of the nth original image at (x, y), where x = 1, 2, ······, A - 1, A, y = 1, 2, ······, B - 1, B, and n = 1, 2, ······, Q - 1, Q; ε is the brightness value with the distribution probability of low pixel values in the original image not less than the 20% threshold; θ is the brightness value with the distribution probability of high pixel values in the original image less than the 1% threshold; a is the proportionality coefficient; others is the part of the original image excluding the above 1% high pixels and 20% low pixels.

3. The infrared small target detection method based on a double-layer feature fusion convolutional network according to claim 1, wherein The network layer N1 and the upsampling network layer N2 perform dilated convolution according to the following formulas respectively: Among them, F j (g) is the result of the hollow convolution of the discrete function g at the dilation factor j; g ∈ N, where N is a non - negative number; * is the convolution symbol; k is the discrete filter; s is the convolution range of the discrete function g, t is the convolution step of the filter k; p is the time value.

4. A method for infrared small target detection based on a double-layer feature fusion convolutional network according to claim 1 or 3, characterized in that, The information fusion between network layer N1 and upsampling network layer N2 adopts a T-shaped fusion structure, and the low-level features and high-level features of the network are fused according to the following formula through the fusion convolution modules of network layer N1 and upsampling network layer N2: y l = F j (γ l-2 ) + F j {F j' (γ l )} Among them, y l is the fusion feature of the l-th level of the network; F j (γ l-2 ) is the convolutional feature of the (l-2)-th level of the image when the dilation factor is j; F j {F j' (γ l )} is the feature after convolution of the l-th level of the image when the dilation factor is j, F j' (γ) is the convolutional feature of the l-th level of the image when the dilation factor before fusion is j'; γ l is the information feature of the l-th level generated by the network, γ l-2 is the information feature of the (l-2)-th level generated by the network, j' is the dilation factor before the l-th level of fusion, and j is the dilation factor during the l-th level of fusion.

5. The infrared small target detection method based on a double-layer feature fusion convolutional network according to claim 1 or 3, characterized in that The double-layer feature fusion convolutional network maps the input image I from the first dilated convolution module of network layer N1 and upsampling network layer N2 to the twelfth dilated convolution module, forming the following network model: The output of the double-layer feature fusion convolutional network is as follows: y(I) = F j {N1(I) + N2(I)} Among them, N1(I) is the convolution operation of the network layer N1 on the image I, and N2(I) is the convolution operation of the upsampling network layer N2 on the image I; F 1 j,1 , F 1 j,2 , ……, F 1 j,12 are respectively twelve cascaded dilated convolution modules of the network layer N1; F 2 j,1 , F 2 j,2 , ……, F 2 j,12 are respectively twelve cascaded dilated convolution modules of the upsampling network layer N2; F j,1 (I) is the convolution feature extraction result of the image I before inputting into the N1 and N2 networks.

6. The infrared small target detection method based on a double-layer feature fusion convolutional network according to claim 1, wherein, In step 3, the loss function DLFF_Loss of the constructed double-layer feature fusion convolutional network is composed of the D_Loss loss function and the F_Loss loss function, where 1) The D_Loss loss function is: Among them, is the predicted object detection result, and y is the true object detection result; 2) The F_Loss loss function is: F_Loss = α t (1 - p t ) γ log(P t ) Among them, α t is the weight coefficient, γ is the focusing coefficient, where P t is the parameter p related to the binary cross-entropy loss function BCE_Loss t = e -BCE_Loss , and the binary cross-entropy loss function BCE_Loss is: Among them, is the predicted object detection result, and y is the true object detection result; 3) The loss function DLFF_Loss of the double-layer feature fusion convolutional network is: DLFF_Loss = β0[D_Loss+(1-β1)F_Loss] where β1 is the weight coefficient of the loss function and β0 is the experimental coefficient.

Images