Floating object defogging image acquisition method and system in haze environment, and detection method

Abstract

The invention relates to a method and a system for acquiring a defogging image of a floating object in a haze environment, and a detection method. The method comprises the steps of obtaining a fog concentration coefficient based on a minimum fog concentration value, a maximum fog concentration value and an average fog concentration value; acquiring atmospheric light values corresponding to different pixel points; in combination with the fog concentration coefficient and the atmospheric light values corresponding to the different pixel points, rough transmittances corresponding to the different pixel points are obtained; compensating the rough transmittances corresponding to the different pixel points by using a Gaussian attenuation fitting method and combining the atmospheric light values corresponding to the different pixel points, so as to obtain Gaussian function transmittances corresponding to the different pixel points; a guided filtering method is adopted to carry out refining operation on the Gaussian function transmissivity corresponding to different pixel points so as to obtain refined transmissivity corresponding to different pixel points; and based on the atmospheric light values corresponding to the different pixel points and the refined transmittances corresponding to the different pixel points, obtaining a floating object defogging image.

Description

Technical Field

[0001] Several embodiments of this specification relate to the field of image recognition technology, specifically to methods, systems, and detection methods for acquiring and detecting defogging images of floating objects in hazy environments. Background Technology

[0002] Floating debris, as a significant apparent source of river pollution, seriously impacts river and lake environmental protection efforts. Reducing the amount and duration of floating debris accumulation, and strengthening floating debris monitoring, are crucial ways to improve the ecological quality of the aquatic environment. Current methods for monitoring and tracking floating debris include manual identification and machine vision identification; however, complex aquatic environments can easily create difficulties in monitoring, such as ambient light and fog / haze obstruction, leading to missed detections.

[0003] Fog and haze are natural phenomena caused by the scattering of light by water droplets and particles in the atmosphere. The presence of these phenomena significantly reduces the contrast and visibility of images obtained by some imaging systems. Specifically, images acquired in foggy or hazy conditions typically suffer from reduced signal-to-noise ratio, color distortion, and information loss, severely impacting the detection of floating objects. Particularly noteworthy is that water environments often have higher fog concentrations and humidity compared to typical land scenes. Furthermore, the light reflection effect of the water surface further amplifies the "whitening" interference caused by fog and haze, drastically compressing the contrast between floating objects and the water background. This makes even small floating targets more easily obscured in low-contrast, murky images, posing a more severe challenge to monitoring work than in land environments.

[0004] Therefore, there is an urgent need for a method that can effectively detect floating objects on the water surface in smoggy environments. Summary of the Invention

[0005] This specification provides a method, system, and detection method for acquiring and detecting floating objects in hazy environments, which can effectively detect floating objects on the water surface in hazy images.

[0006] The technical solution is as follows:

[0007] Firstly, embodiments of this specification provide a method for acquiring defogging images of floating objects in a hazy environment, including:

[0008] Acquire images of floating debris and haze on the water surface in a hazy environment;

[0009] Obtain the brightness entropy map and texture feature map corresponding to the haze image of floating objects on the water surface;

[0010] Based on the brightness entropy map and texture feature map, obtain the fog concentration distribution map;

[0011] Based on the minimum fog concentration value, maximum fog concentration value, and average fog concentration value of the fog concentration distribution map, the fog concentration coefficient is obtained.

[0012] Based on the haze image of floating objects on the water surface, obtain the atmospheric light value corresponding to each pixel;

[0013] By utilizing the dark channel prior and combining the haze image of floating objects on the water surface, the fog concentration coefficient, and the atmospheric light value corresponding to each pixel, the coarse transmittance corresponding to each pixel is obtained.

[0014] By using the Gaussian attenuation fitting method and combining the atmospheric light values ​​corresponding to different pixels, the coarse transmittance corresponding to different pixels is compensated to obtain the Gaussian function transmittance corresponding to different pixels.

[0015] A guided filtering method is used to refine the Gaussian function transmittance of different pixels to obtain the refined transmittance of each pixel.

[0016] Based on the haze image of floating objects on the water surface, the atmospheric light value corresponding to each pixel, and the refined transmittance corresponding to each pixel, a dehazing image of floating objects is obtained.

[0017] As a preferred embodiment, obtaining the brightness entropy map corresponding to the haze image of floating objects on the water surface includes:

[0018] Normalize the haze image of floating objects on the water surface to obtain a normalized haze image;

[0019] Based on the normalized haze image, the brightness channel image is obtained;

[0020] The brightness channel image is processed using the entropy formula, and a guided filtering operation is performed to obtain the brightness entropy image.

[0021] The step of obtaining the brightness channel image based on the normalized haze image includes:

[0022] ;

[0023] in, Indicates the position in the bright channel diagram The value corresponding to the pixel at that location. Indicates the location in the normalized haze image The pixel corresponding to Channel values, These represent the three channels of the image. Represents a position A preset-size pixel window centered on the pixel. This indicates taking the maximum value.

[0024] As a preferred embodiment, the step of obtaining the texture feature map corresponding to the haze image of floating objects on the water surface includes:

[0025] Convert images of floating debris and haze on the water surface to grayscale;

[0026] Gaussian blur and downsampling are applied to the grayscale image to obtain the first downsampled blurred image;

[0027] Gaussian blur and downsampling are applied to the first downsampled blur image to obtain the second downsampled blur image;

[0028] The second downsampled blurred image is then upsampled and normalized to obtain a texture feature map.

[0029] As a preferred embodiment, the fog concentration coefficient is obtained based on the minimum fog concentration value, the maximum fog concentration value, and the average fog concentration value, including:

[0030] ;

[0031] in, Indicates the fog concentration coefficient. This represents the minimum fog concentration value. This represents the average fog concentration value. This indicates the maximum fog concentration value.

[0032] As a preferred embodiment, the step of obtaining the atmospheric light values ​​corresponding to different pixels based on the haze image of floating objects on the water surface includes:

[0033] Convert the haze image of floating objects on the water surface from the RGB color space to the HSV color space to obtain a color image of the floating objects;

[0034] A morphological closing operation is performed on the color image of the floating object to obtain the processed image;

[0035] The processed image is then subjected to cross-bilateral filtering for smoothing and normalization to obtain the atmospheric light values ​​corresponding to different pixels.

[0036] As a preferred embodiment, the method of utilizing the dark channel prior and combining it with images of floating objects on the water surface, fog concentration coefficients, and atmospheric light values ​​corresponding to different pixels to obtain the approximate transmittance of each pixel includes:

[0037] ;

[0038] in, Indicates position The approximate transmittance corresponding to each pixel. Indicates the fog concentration coefficient. Indicates position The atmospheric light value corresponding to that pixel. This indicates taking the minimum value.

[0039] As a preferred embodiment, the step of using Gaussian attenuation fitting method and combining it with the atmospheric light value corresponding to each pixel to compensate for the coarse transmittance corresponding to each pixel, in order to obtain the Gaussian function transmittance corresponding to each pixel, includes:

[0040] ;

[0041] in, This represents the correction factor. Represents the Euler number. To correct the constant, Indicates position The Gaussian transmittance corresponding to the pixel at that location.

[0042] As a preferred approach, a dehazed image of floating objects is obtained based on the haze image of water surface debris, the atmospheric light value corresponding to each pixel, and the refined transmittance corresponding to each pixel, including:

[0043] ;

[0044] in, Indicates the position of floating objects in the dehazed image. The value corresponding to the pixel at that location. This indicates the location of floating debris in the normalized haze image obtained by normalizing the haze image. The value corresponding to the pixel at that location. Indicates position The atmospheric light value corresponding to that pixel. Indicates position The refined transmittance corresponding to each pixel. This indicates taking the maximum value.

[0045] Secondly, embodiments of this specification provide a system for acquiring images of floating objects in a hazy environment, including:

[0046] The first image acquisition module acquires images of floating objects and haze on the water surface.

[0047] The second image acquisition module acquires the brightness entropy map and texture feature map corresponding to the haze image of floating objects on the water surface.

[0048] The concentration distribution map acquisition module obtains the fog concentration distribution map based on the brightness entropy map and texture feature map;

[0049] The coefficient acquisition module obtains the fog concentration coefficient based on the minimum fog concentration value, maximum fog concentration value, and average fog concentration value of the fog concentration distribution map.

[0050] The local light value acquisition module acquires the atmospheric light values ​​corresponding to different pixels based on the haze image of floating objects on the water surface.

[0051] The transmittance acquisition module utilizes the dark channel prior and combines it with images of floating objects and haze on the water surface, fog concentration coefficient, and atmospheric light values ​​corresponding to different pixels to obtain the rough transmittance corresponding to each pixel.

[0052] The compensation module uses the Gaussian attenuation fitting method and combines the atmospheric light value corresponding to each pixel to compensate for the coarse transmittance of each pixel, so as to obtain the Gaussian function transmittance of each pixel.

[0053] The refinement module uses a guided filtering method to refine the Gaussian function transmittance corresponding to each pixel to obtain the refined transmittance corresponding to each pixel.

[0054] The third image acquisition module acquires a defogging image of the floating objects based on the haze image of the floating objects on the water surface, the atmospheric light value corresponding to each pixel, and the refined transmittance corresponding to each pixel.

[0055] Thirdly, embodiments of this specification provide a method for detecting floating objects, including:

[0056] A defogging image of floating objects in a hazy environment is obtained by using the method for acquiring defogging images of floating objects in a hazy environment as described in the first aspect of the embodiment;

[0057] An improved YOLOv5s model was used to identify floating objects in dehazed images.

[0058] The improved YOLOv5s model includes a backbone network, a neck network, and a detection network with a four-layer detection structure connected in sequence.

[0059] The backbone network includes a first CBS module, a second CBS module, a first C3 module, a third CBS module, a second C3 module, a fourth CBS module, a third C3 module, a fifth CBS module, a C3-MS-CBAM module, and an SPPF module connected in sequence; the second CBS module is connected to the neck network through the sixth CBS module, the third CBS module is connected to the neck network through the seventh CBS module, and the fourth CBS module is connected to the neck network through the eighth CBS module.

[0060] The C3-MS-CBAM module includes the ninth CBS module, the tenth CBS module, the eleventh CBS module, the merging module, the twelfth CBS module, and the MS-CBAM module connected in sequence. The C3-MS-CBAM module also includes the thirteenth CBS module. The ninth CBS module is also connected to the merging module through the thirteenth CBS module.

[0061] The MS-CBAM module includes a multi-scale channel attention module, a spatial attention module, and a multiplication module connected in sequence. The multi-scale channel attention module is also directly connected to the multiplication module.

[0062] In the multi-scale channel attention module, the initial feature map of the input is processed through the first image processing path and the second image processing path to obtain the first processed image and the second processed image, respectively. The first processed image and the second processed image are then merged and activated by a function to obtain the third processed image. Finally, the third processed image is multiplied with the initial feature map to obtain the output image.

[0063] In the first image processing path, global average pooling, pointwise convolution, function activation, and pointwise convolution are performed sequentially on the initial feature map.

[0064] In the second image processing path, pointwise convolution, function activation, and pointwise convolution operations are performed sequentially on the initial feature map.

[0065] Fourthly, embodiments of this specification provide an electronic device, including a processor and a memory; the processor is connected to the memory; the memory is used to store executable program code; the processor reads the executable program code stored in the memory to run a program corresponding to the executable program code, so as to perform the steps described in the first or third aspect of the above embodiments.

[0066] Fifthly, embodiments of this specification provide a computer storage medium storing a plurality of instructions adapted for loading by a processor and executing the steps described in the first or third aspect of the above embodiments.

[0067] The beneficial effects of the technical solutions provided in some embodiments of this specification include at least the following:

[0068] By combining the brightness entropy map and texture feature map, a fog distribution map is obtained. Based on the minimum fog concentration value, maximum fog concentration value, and average fog concentration value of the fog concentration distribution map, an adaptively adjusted fog concentration coefficient is obtained. Based on the fog image of floating objects on the water surface, the atmospheric light value corresponding to each pixel is obtained. Subsequently, the fog image of floating objects on the water surface is defogging based on the adaptively adjusted fog concentration coefficient and the local atmospheric light value to improve the defogging effect.

[0069] By combining the haze image of floating objects on the water surface, the fog concentration coefficient, and the atmospheric light value corresponding to each pixel, the coarse transmittance corresponding to each pixel is obtained. Gaussian attenuation fitting compensation and guided filtering refinement operations are then performed sequentially to make the edges of small targets on the water surface clearer and the details and textures restored more naturally, thereby significantly improving the overall visual quality of the dehazed image of floating objects.

[0070] Traditional CBAM attention mechanisms process features at a single scale, making them ill-suited to varying target scales. This leads to tiny target features being easily obscured by background noise. Spatial and channel attention often employ a fixed-order, sequential computation, lacking cross-layer feature interaction capabilities, which can cause feature response biases in large-scale scenarios. This invention proposes the MS-CBAM multi-scale attention mechanism. Based on the traditional CBAM architecture, the MS-CBAM multi-scale attention mechanism replaces the channel attention module with a multi-scale channel attention module, forming an attention mechanism with hierarchical perception capabilities. The C3 module in the original YOLOv5s model, based on the cross-stage local connectivity structure of CSPNet, while improving computational efficiency, suffers from a single-scale feature modeling mechanism that struggles to effectively capture key features of multi-granularity targets, especially at the C3 module location in shallow network stages, where it easily loses detailed texture information of tiny targets. The MS-CBAM module, by introducing a multi-scale mechanism, simultaneously establishes a progressive feature selection capability from local details to global semantics in both channel and spatial dimensions. It preserves the edge responses of small targets through multi-branch dilated convolutions while suppressing background noise interference through multi-scale processing. Attached Figure Description

[0071] 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0072] Figure 1 A flowchart illustrating a method for acquiring defogging images of floating objects in a hazy environment, according to some embodiments of this disclosure, is shown.

[0073] Figure 2 A schematic diagram of the overall structure of the improved YOLOv5s model is shown.

[0074] Figure 3 A schematic diagram of the structure of the first C3 module is shown.

[0075] Figure 4 A schematic diagram of the structure of the second C3 module is shown.

[0076] Figure 5 A schematic diagram of the third C3 module is shown.

[0077] Figure 6 A schematic diagram of the Bottleneck module is shown.

[0078] Figure 7 A schematic diagram of the CBS module is shown.

[0079] Figure 8 A schematic diagram of the C3-MS-CBAM module is shown.

[0080] Figure 9 A schematic diagram of the MS-CBAM module is shown.

[0081] Figure 10 A schematic diagram of the image processing logic in the multi-scale channel attention module is shown.

[0082] Figure 11 A schematic diagram illustrating the detection effect of the floating object detection method described in the embodiments of this specification is shown.

[0083] Figure 12 A schematic diagram of the structure of a system for acquiring images of floating objects in a hazy environment, according to some embodiments of the present disclosure, is shown.

[0084] Figure 13 A schematic block diagram of an electronic device according to some embodiments of the present disclosure is shown.

[0085] In the diagram: 11. First CBS module; 12. Second CBS module; 13. Third CBS module; 14. Fourth CBS module; 15. Fifth CBS module; 16. Sixth CBS module; 17. Seventh CBS module; 18. Eighth CBS module; 19. Ninth CBS module; 110. Tenth CBS module; 111. Eleventh CBS module; 112. Twelfth CBS module; 113. Thirteenth CBS module; 21. First C3 module; 22. Second C3 module; 23. Third C3 module; 3. C3-MS-CBAM module; 4. SPPF module; 51. Merge module; 52. Multiply module; 6. MS-CBAM module. Detailed Implementation

[0086] The technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings.

[0087] The terms "first," "second," "third," etc., in the description, claims, and accompanying drawings are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.

[0088] The following description provides examples and does not limit the scope, applicability, or examples set forth in the claims. Changes may be made to the function and arrangement of the described elements without departing from the scope of this specification. Various processes or components may be appropriately omitted, substituted, or added to the examples. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Furthermore, features described with respect to some examples may be combined into other examples.

[0089] Figure 1 A flowchart illustrating a method for acquiring defogging images of floating objects in a hazy environment, according to some embodiments of this disclosure, is shown. Figure 1 As shown, methods for acquiring dehazing images of floating objects in hazy environments can include at least the following:

[0090] Step 102: Obtain images of floating debris and haze on the water surface under hazy conditions;

[0091] Step 104: Obtain the brightness entropy map and texture feature map corresponding to the haze image of floating objects on the water surface;

[0092] Step 106: Obtain the fog concentration distribution map based on the brightness entropy map and texture feature map;

[0093] Step 108: Obtain the fog concentration coefficient based on the minimum fog concentration value, maximum fog concentration value, and average fog concentration value from the fog concentration distribution map;

[0094] Step 110: Based on the haze image of floating objects on the water surface, obtain the atmospheric light value corresponding to each pixel;

[0095] Step 112: Using the dark channel prior, and combining the haze image of floating objects on the water surface, the fog concentration coefficient, and the atmospheric light value corresponding to each pixel, obtain the rough transmittance corresponding to each pixel.

[0096] Step 114: Using the Gaussian attenuation fitting method and combining the atmospheric light values ​​corresponding to different pixels, the coarse transmittance corresponding to different pixels is compensated to obtain the Gaussian function transmittance corresponding to different pixels.

[0097] Step 116: Use the guided filtering method to refine the Gaussian function transmittance corresponding to each pixel to obtain the refined transmittance corresponding to each pixel.

[0098] Step 118: Obtain a defogging image of floating objects based on the haze image of floating objects on the water surface, the atmospheric light value corresponding to each pixel, and the refined transmittance corresponding to each pixel.

[0099] In foggy images, the distribution and concentration of fog are directly perceptible, primarily characterized by high brightness values. Therefore, fog distribution can be understood as the degree of disorder in brightness information. In most cases, the higher the degree of disorder, the denser and more diffuse the fog distribution. In physics, the concept that describes the degree of disorder is information entropy. Therefore, brightness entropy values ​​can be used to indirectly reflect fog characteristics; that is, the higher the entropy value, the higher the randomness of fog appearance. However, brightness entropy values ​​alone cannot fully reflect the fog distribution; they also need to be combined with the texture features of the original image. Therefore, this invention combines a brightness entropy map and a texture feature map to obtain a fog distribution map.

[0100] This application obtains an adaptively adjusted fog concentration coefficient based on the minimum fog concentration value, maximum fog concentration value, and average fog concentration value of the fog concentration distribution map. Based on the fog image of floating objects on the water surface, it obtains the atmospheric light value corresponding to each pixel. Subsequently, based on the adaptively adjusted fog concentration coefficient and the local atmospheric light value, the fog image of floating objects on the water surface is defogging to achieve a better defogging effect.

[0101] Understandably, the adaptive fog concentration coefficient can dynamically adjust the defogging intensity based on the specific fog distribution in the image. By using minimum, maximum, and average values, it can accurately determine the severity of fog in the current image (whether it's thin fog or dense fog), thus automatically matching the most suitable coefficient and avoiding insufficient defogging (not removing fog completely) or excessive defogging (image distortion) caused by fixed parameters. Due to the complex situation of uneven lighting on the water surface, strong reflections, or the presence of shadows, the local atmospheric light value is calculated individually for each pixel. This avoids overexposure and whitening in bright areas and color distortion in dark areas during defogging, ensuring the accuracy of color reproduction throughout the image.

[0102] This application combines images of floating objects in haze, fog concentration coefficients, and atmospheric light values ​​corresponding to different pixels to obtain the coarse transmittance of each pixel. Then, it performs Gaussian attenuation fitting compensation and guided filtering refinement operations to make the edges of small targets on the water surface clearer and the details and textures restored more naturally, thereby significantly improving the overall visual quality of the dehazed image of floating objects.

[0103] In some embodiments of this specification, obtaining the brightness entropy map corresponding to the haze image of floating objects on the water surface includes:

[0104] Normalize the haze image of floating objects on the water surface to obtain a normalized haze image;

[0105] Based on the normalized haze image, the brightness channel image is obtained;

[0106] The brightness channel image is processed using the entropy formula, and a guided filtering operation is performed to obtain the brightness entropy image.

[0107] The step of obtaining the brightness channel image based on the normalized haze image includes:

[0108] ;

[0109] in, Indicates the position in the bright channel diagram The value corresponding to the pixel at that location. Indicates the location in the normalized haze image The pixel corresponding to Channel values, These represent the three channels of the image. Represents a position A preset-size pixel window centered on the pixel. This indicates taking the maximum value.

[0110] The process involves processing the brightness channel image using the entropy formula and performing guided filtering to obtain a brightness entropy image, including:

[0111] ;

[0112] in, Represents the position in the brightness entropy map The luminance entropy value corresponding to the pixel at that location. Denotes the regularization constant, taking... , This indicates a guided filtering operation.

[0113] In some embodiments of this specification, obtaining the texture feature map corresponding to the water surface floating object haze image includes:

[0114] Convert the image of floating debris and haze on the water surface into a grayscale image;

[0115] Gaussian blur and downsampling are applied to the grayscale image to obtain the first downsampled blurred image;

[0116] Gaussian blur and downsampling are applied to the first downsampled blur image to obtain the second downsampled blur image;

[0117] The second downsampled blurred image is then upsampled and normalized to obtain a texture feature map.

[0118] The process of converting images of floating debris and haze on the water surface into grayscale images includes:

[0119] ;

[0120] in, Represents the position in a grayscale image The value corresponding to the pixel at that location. Indicates the location of floating objects in the haze image. The red channel value corresponding to that pixel. Indicates the location of floating objects in the haze image. The green channel value corresponding to that pixel. Indicates the location of floating objects in the haze image. The value of the blue channel corresponding to the pixel.

[0121] The process involves performing Gaussian blurring and downsampling on the grayscale image to obtain a first downsampled blurred image with a size reduced to half the original image size, including:

[0122] ;

[0123] (Note: This is the Gaussian blur formula);

[0124] in, Indicates the position in the first downsampled blur map The value corresponding to the pixel at that location. Indicates downsampling, This represents the standard deviation of the Gaussian filter. Represents pi (π). This indicates the coordinate position of a pixel.

[0125] Specifically, the first downsampled blurred image is subjected to Gaussian blurring and downsampling to obtain a second downsampled blurred image whose size is reduced to 1 / 4 of the original image size, including:

[0126] ;

[0127] in, Indicates the position in the second downsampled blur map The numerical value corresponding to the pixel at that location.

[0128] The second downsampled blurred image undergoes upsampling and normalization to obtain a texture feature map, including:

[0129] ;

[0130] in, Represents the position in the texture feature map The value corresponding to the pixel at that location. This indicates upsampling.

[0131] Among them, based on the brightness entropy map and texture feature map, the fog concentration distribution map is obtained, including:

[0132] ;

[0133] in, Indicates the location in the fog concentration distribution map The value corresponding to the pixel at that location. This indicates a guided filtering operation.

[0134] In some embodiments of this specification, a fog concentration coefficient is obtained based on the minimum fog concentration value, the maximum fog concentration value, and the average fog concentration value, including:

[0135] ;

[0136] in, Indicates the fog concentration coefficient. This represents the minimum fog concentration value. This represents the average fog concentration value. This indicates the maximum fog concentration value.

[0137] In some embodiments of this specification, obtaining the atmospheric light values ​​corresponding to different pixels based on the haze image of floating objects on the water surface includes:

[0138] Convert the haze image of floating objects on the water surface from the RGB color space to the HSV color space to obtain a color image of the floating objects;

[0139] A morphological closing operation is performed on the color image of the floating object to obtain the processed image;

[0140] The processed image is then subjected to cross-bilateral filtering for smoothing and normalization to obtain the atmospheric light values ​​corresponding to different pixels.

[0141] The process involves performing a morphological closing operation on the color image of the floating objects to obtain the processed image, including:

[0142] ;

[0143] in, Indicates the position of floating objects in the color image The value corresponding to the pixel at that location. Indicates the position in the processed image The value corresponding to the pixel at that location. This indicates a morphological closing operation. This indicates that a structuring element is used to indicate the closing operation.

[0144] The process involves performing cross-bilateral filtering smoothing and normalization on the processed image to obtain the atmospheric light values ​​corresponding to different pixels, including:

[0145] ;

[0146] in, This indicates cross-bilateral filtering for smoothing. Indicates the location of floating objects in the haze image. The numerical value corresponding to the pixel at that location.

[0147] In some embodiments of this specification, the method of utilizing dark channel priors and combining them with images of floating objects on the water surface, fog concentration coefficients, and atmospheric light values ​​corresponding to different pixels to obtain the approximate transmittance of each pixel includes:

[0148] ;

[0149] in, Indicates position The approximate transmittance corresponding to each pixel. Indicates the fog concentration coefficient. Indicates position The atmospheric light value corresponding to that pixel. This indicates taking the minimum value.

[0150] In some embodiments of this specification, the method of using Gaussian attenuation fitting and combining it with the atmospheric light values ​​corresponding to different pixels to compensate for the coarse transmittance of each pixel in order to obtain the Gaussian function transmittance of each pixel includes:

[0151] ;

[0152] in, This represents the correction factor, which can be 0.4. Represents the Euler number. As a correction constant, it can be taken as 0.5. Indicates position The Gaussian transmittance corresponding to the pixel at that location.

[0153] In some embodiments of this specification, a guided filtering method is used to refine the Gaussian transmittance corresponding to each pixel to obtain the refined transmittance corresponding to each pixel, including:

[0154] ;

[0155] in, Indicates position The refined transmittance corresponding to the pixel.

[0156] In some embodiments of this specification, a dehazed image of floating objects is obtained based on a haze image of water surface debris, the atmospheric light value corresponding to each pixel, and the refined transmittance corresponding to each pixel, including:

[0157] ;

[0158] in, Indicates the position of floating objects in the dehazed image. The value corresponding to the pixel at that location. This indicates the location of floating debris in the normalized haze image obtained by normalizing the haze image. The value corresponding to the pixel at that location. Indicates position The atmospheric light value corresponding to that pixel. Indicates position The refined transmittance corresponding to each pixel. This indicates taking the maximum value.

[0159] This specification also provides a method for detecting floating objects, which may include at least:

[0160] The above-described method for obtaining defogging images of floating objects in a hazy environment was used to obtain defogging images of floating objects.

[0161] An improved YOLOv5s model was used to identify floating objects in dehazed images.

[0162] The improved YOLOv5s model includes a backbone network, a neck network, and a detection network with a four-layer detection structure connected in sequence.

[0163] The backbone network includes a first CBS module 11, a second CBS module 12, a first C3 module 21, a third CBS module 13, a second C3 module 22, a fourth CBS module 14, a third C3 module 23, a fifth CBS module 15, a C3-MS-CBAM module 3, and an SPPF module 4 connected in sequence. The second CBS module 12 is connected to the neck network through a sixth CBS module 16, the third CBS module 13 is connected to the neck network through a seventh CBS module 17, and the fourth CBS module 14 is connected to the neck network through an eighth CBS module 18.

[0164] The C3-MS-CBAM module 3 includes a ninth CBS module 19, a tenth CBS module 110, an eleventh CBS module 111, a merging module 51, a twelfth CBS module 112, and an MS-CBAM module 6 connected in sequence. The C3-MS-CBAM module 3 also includes a thirteenth CBS module 113. The ninth CBS module 19 is also connected to the merging module 51 through the thirteenth CBS module 113.

[0165] MS-CBAM module 6 includes a multi-scale channel attention module, a spatial attention module, and a multiplication module 52 connected in sequence. The multi-scale channel attention module is also directly connected to the multiplication module 52.

[0166] In the multi-scale channel attention module, the initial feature map of the input is processed through the first image processing path and the second image processing path, respectively. The image is processed to obtain a first processed image and a second processed image. These two images are then merged and activated to obtain a third processed image. Finally, the third processed image is compared with the initial feature map. Perform a multiplication operation to obtain the output image;

[0167] In the first image processing path, the initial feature map is processed sequentially. Perform global average pooling, pointwise convolution, function activation, and pointwise convolution operations.

[0168] In the second image processing path, pointwise convolution, function activation, and pointwise convolution operations are performed sequentially on the initial feature map.

[0169] It is understandable that floating objects on water surfaces present the problem of extremely small targets. Therefore, to address the issue that extremely small targets occupy few pixels in images, easily leading to decreased accuracy and missed detections, this specification proposes an improved YOLOv5s model in its embodiments. This model employs a detection network with a four-layer detection structure to enhance the recognition ability of small targets. Furthermore, the traditional CBAM attention mechanism processes features on a single scale, making it difficult to adapt to changes in target scale. This results in extremely small target features being easily submerged by background noise. Spatial and channel attention often employ a fixed-order serial computation, lacking cross-level feature interaction capabilities, which can cause feature response bias in large-scale changing scenes. Therefore, this specification proposes the MS-CBAM multi-scale attention mechanism in its embodiments. The MS-CBAM multi-scale attention mechanism is based on the architecture of the traditional CBAM attention mechanism, replacing the channel attention module with a multi-scale channel attention module, forming an attention mechanism with hierarchical perception capabilities. The C3 module in the original YOLOv5s model is based on the cross-stage local connectivity structure of CSPNet. While this module can improve computational efficiency, its single-scale feature modeling mechanism struggles to effectively capture key features of multi-granularity targets, especially at the C3 module location in shallow network stages, where it easily loses detailed texture information of extremely small targets. The MS-CBAM module 6 introduces a multi-scale mechanism, simultaneously establishing a progressive feature selection capability from local details to global semantics in both channel and spatial dimensions. It preserves the edge responses of small targets through multi-branch dilated convolutions and suppresses background noise interference through multi-scale processing. The overall structure of the improved YOLOv5s model can be found by referring to... Figure 2 As shown.

[0170] The channel attention mechanism simultaneously performs max pooling and global average pooling on the input feature map, resulting in two 1×1 feature maps of the same dimension. These two feature maps are then fed into a shared layer to compress parameters. Finally, the two feature maps are added element-wise and passed through a sigmoid activation function to obtain the channel attention map. The shared layer here is a multilayer perceptron plus a hidden layer, with an activation size of [value missing]. , Let C represent the set of real numbers, C represent the number of channels, and r represent the reduction ratio.

[0171] ;

[0172] in, This represents the Sigmoid activation function. Indicates shared multi-layer sensing operations. Indicates average pooling. This indicates max pooling. This represents the input feature map. Max pooling and average pooling are applied to the input feature map, respectively. Max pooling and average pooling operations are performed to generate two 1x1 feature maps of different dimensions. These operations capture global information from different perspectives. A shared multilayer perceptron (MLP) operation inputs the two 1x1 feature maps obtained from max pooling and average pooling into a shared MLP, compressing parameters and outputting two feature maps. The shared MLP typically contains two fully connected layers to model channel features. Finally, the two feature maps output from the shared MLP are summed, and channel attention weights are generated using a sigmoid activation function. The shared layer here is a multilayer perceptron plus a hidden layer, with an activation size of [value missing]. .in, and The weights represent the weights of a multilayer perceptron, which are shared by the two inputs and are obtained using the ReLU function. , This represents the feature map after average pooling. This represents the feature map after max pooling.

[0173] The detection head is a key component in object detection models, responsible for predicting the location, size, and class of objects based on feature maps provided by the neck network. It typically consists of multiple convolutional layers and outputs the object's class probability and bounding box coordinates through classification and regression branches. Traditional YOLOv5s uses three detection heads of different scales, corresponding to different feature map sizes: 20×20, 40×40, and 80×80. The large 80×80 feature map focuses on small object detection, the medium 40×40 feature map is used for medium-sized objects, and the small 20×20 feature map is more suitable for large objects. Detection depth is influenced by the number of layers in the backbone and neck networks; deeper networks can extract richer features but also increase computational complexity.

[0174] Under fixed input size conditions, the YOLOv5s model obtains multi-level feature layers through stepwise sampling at 8x, 16x, and 32x, and performs detection tasks at three scales respectively. However, the feature layer generated by 8x downsampling has the most basic perceptual unit. When the actual size of the target is smaller than the basic unit, it is very easy to miss detections.

[0175] To address the issue of missed detection of small targets, this specification's embodiments optimize the structure of the YOLOv5s detection head module. In addition to the original three-layer detection, a 160×160 detection head is added for detecting extremely small targets. The new detection layer generates feature maps through a 4x downsampling operation, achieving a denser anchor box distribution by reducing the perception range of individual grids. This finer grid division mechanism significantly improves the grid's ability to capture extremely small targets.

[0176] in, Figure 3The structure of the first C3 module 21 is shown. Figure 4 The structure of the second C3 module 22 is shown. Figure 5 The structure of the third C3 module 23 is shown. Figure 6 The structure of the Bottleneck module is shown. Figure 7 The structure of the CBS module is shown.

[0177] in, Figure 8 The structure of C3-MS-CBAM module 3 is shown. Figure 9 A schematic diagram of the MS-CBAM module 6 is shown. Figure 10 The diagram illustrates the image processing logic within the multi-scale channel attention module. This module combines multi-scale feature extraction with channel attention. The multi-scale attention mechanism captures both local and global information and emphasizes the most discriminative features at different scales. The multi-scale attention mechanism is shown below.

[0178] ;

[0179] ;

[0180] ;

[0181] in, Channel attention calculation representing local features, Channel attention calculation representing global features This indicates that pointwise convolution will input features The number of channels is reduced to 1 / R of the original number. Represents the ReLU activation function. Pointwise convolution restores the number of channels to the same number as the original input channels, where R represents the channel scaling ratio. Indicates global average pooling. This represents the Sigmoid activation function.

[0182] Spatial attention mechanisms extract attention by processing the spatial representation of features. While distinct from, they complement multi-scale channel attention mechanisms. Spatial attention mechanisms first process the feature map... The channel is compressed using average pooling and max pooling operations to obtain two... The feature maps are concatenated along the channels, then dimensionality is reduced by a 7×7 convolution, and finally a spatial attention map is formed by the Sigmoid function.

[0183] Figure 11The diagram shows the detection effect of the floating object detection method described in the embodiments of this specification. As can be seen from the diagram, accurate identification can be achieved for small-sized image targets.

[0184] Figure 12 This document illustrates a schematic diagram of a system for acquiring images of floating objects in a hazy environment, representing some embodiments of this disclosure. The various embodiments in this specification are described in a progressive manner; similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on its differences from other embodiments. In particular, the embodiments for acquiring images of floating objects are largely similar to the embodiments for acquiring images of floating objects, and therefore are described simply; relevant details can be found in the descriptions of the embodiments for acquiring images of floating objects.

[0185] like Figure 12 As shown, a system for acquiring images of floating objects in a hazy environment can include at least:

[0186] The first image acquisition module acquires images of floating objects and haze on the water surface.

[0187] The second image acquisition module acquires the brightness entropy map and texture feature map corresponding to the haze image of floating objects on the water surface.

[0188] The concentration distribution map acquisition module obtains the fog concentration distribution map based on the brightness entropy map and texture feature map;

[0189] The coefficient acquisition module obtains the fog concentration coefficient based on the minimum fog concentration value, maximum fog concentration value, and average fog concentration value of the fog concentration distribution map.

[0190] The local light value acquisition module acquires the atmospheric light values ​​corresponding to different pixels based on the haze image of floating objects on the water surface.

[0191] The transmittance acquisition module utilizes the dark channel prior and combines it with images of floating objects and haze on the water surface, fog concentration coefficient, and atmospheric light values ​​corresponding to different pixels to obtain the rough transmittance corresponding to each pixel.

[0192] The compensation module uses the Gaussian attenuation fitting method and combines the atmospheric light value corresponding to each pixel to compensate for the coarse transmittance of each pixel, so as to obtain the Gaussian function transmittance of each pixel.

[0193] The refinement module uses a guided filtering method to refine the Gaussian function transmittance corresponding to each pixel to obtain the refined transmittance corresponding to each pixel.

[0194] The third image acquisition module acquires a defogging image of the floating objects based on the haze image of the floating objects on the water surface, the atmospheric light value corresponding to each pixel, and the refined transmittance corresponding to each pixel.

[0195] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this specification are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., Digital Versatile Discs (DVDs)), or semiconductor media (e.g., Solid State Disks (SSDs)).

[0196] Figure 13 A block diagram of an electronic device 1300 that can implement various embodiments of the present disclosure is shown. For example... Figure 13 As shown, the electronic device 1300 includes a processor 1310, a disk drive 1320, an input / output interface 1330, a network interface 1340, and a memory 1350. The processor 1310, disk drive 1320, input / output interface 1330, network interface 1340, and memory 1350 can communicate with each other via a communication bus 1360.

[0197] The processor 1310 can be implemented using a general-purpose CPU, microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits to execute relevant programs and implement the technical solution provided in this application.

[0198] The memory 1350 can be implemented in the form of ROM (Read Only Memory), RAM (Read Access Memory), static memory, dynamic storage devices, etc. The memory 1350 can store the operating system 1351 used to control the operation of the electronic device 1300, and the basic input / output system (BIOS) 1352 used to control the low-level operations of the electronic device 1300. Additionally, it can store a web browser 1353, a data storage management system 1354, etc. In summary, when implementing the technical solution provided in this application through software or firmware, the relevant program code is stored in the memory 1350 and is called and executed by the processor 1310.

[0199] The input / output interface 1330 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0200] Network interface 1340 is used to connect a communication module (not shown in the figure) to enable communication and interaction between the device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0201] Bus 1360 includes a pathway for transmitting information between various components of the device, such as processor 1310, disk drive 1320, input / output interface 1330, network interface 1340, and memory 1350.

[0202] It should be noted that although the above-described device only shows the processor 1310, disk drive 1320, input / output interface 1330, network interface 1340, memory 1350, and other components such as 1360, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the method of this application, and does not necessarily include all the components shown in the figures.

[0203] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0204] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing. Furthermore, although operations are depicted in a specific order, this should be understood as requiring that such operations be performed in the specific order shown or in sequential order, or requiring that all illustrated operations be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the foregoing discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations.

[0205] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A method for acquiring dehazing images of floating objects in a hazy environment, characterized in that, include: Acquire images of floating debris and haze on the water surface in a hazy environment; Obtain the brightness entropy map and texture feature map corresponding to the haze image of floating objects on the water surface; Based on the brightness entropy map and texture feature map, obtain the fog concentration distribution map; Based on the minimum fog concentration value, maximum fog concentration value, and average fog concentration value of the fog concentration distribution map, the fog concentration coefficient is obtained. Based on the haze image of floating objects on the water surface, obtain the atmospheric light value corresponding to each pixel; By utilizing the dark channel prior and combining the haze image of floating objects on the water surface, the fog concentration coefficient, and the atmospheric light value corresponding to each pixel, the coarse transmittance corresponding to each pixel is obtained. By using the Gaussian attenuation fitting method and combining the atmospheric light values ​​corresponding to different pixels, the coarse transmittance corresponding to different pixels is compensated to obtain the Gaussian function transmittance corresponding to different pixels. A guided filtering method is used to refine the Gaussian function transmittance of different pixels to obtain the refined transmittance of each pixel. Based on the haze image of floating objects on the water surface, the atmospheric light value corresponding to each pixel, and the refined transmittance corresponding to each pixel, a dehazing image of floating objects is obtained.

2. The method for acquiring defogging images of floating objects in a hazy environment according to claim 1, characterized in that, The acquisition of the brightness entropy map corresponding to the haze image of floating objects on the water surface includes: Normalize the haze image of floating objects on the water surface to obtain a normalized haze image; Based on the normalized haze image, the brightness channel image is obtained; The brightness channel image is processed using the entropy formula, and a guided filtering operation is performed to obtain the brightness entropy image. The step of obtaining the brightness channel image based on the normalized haze image includes: ； in, Indicates the position in the bright channel diagram The value corresponding to the pixel at that location. Indicates the location in the normalized haze image The pixel corresponding to Channel values, These represent the three channels of the image. Represents a position A preset-size pixel window centered on the pixel. This indicates taking the maximum value.

3. The method for acquiring defogging images of floating objects in a hazy environment according to claim 1, characterized in that, The process of acquiring the texture feature map corresponding to the haze image of floating objects on the water surface includes: Convert images of floating debris and haze on the water surface to grayscale; Gaussian blur and downsampling are applied to the grayscale image to obtain the first downsampled blurred image; Gaussian blur and downsampling are applied to the first downsampled blur image to obtain the second downsampled blur image; The second downsampled blurred image is then upsampled and normalized to obtain a texture feature map.

4. The method for acquiring defogging images of floating objects in a hazy environment according to claim 1, characterized in that, The fog concentration coefficient is obtained based on the minimum fog concentration value, the maximum fog concentration value, and the average fog concentration value, including: ； in, Indicates the fog concentration coefficient. This represents the minimum fog concentration value. This represents the average fog concentration value. This indicates the maximum fog concentration value.

5. The method for acquiring defogging images of floating objects in a hazy environment according to claim 1, characterized in that, The process of obtaining atmospheric light values ​​for different pixels based on haze images of floating objects on the water surface includes: Convert the haze image of floating objects on the water surface from the RGB color space to the HSV color space to obtain a color image of the floating objects; A morphological closing operation is performed on the color image of the floating object to obtain the processed image; The processed image is then subjected to cross-bilateral filtering for smoothing and normalization to obtain the atmospheric light values ​​corresponding to different pixels.

6. The method for acquiring defogging images of floating objects in a hazy environment according to claim 2, characterized in that, The method of utilizing dark channel priors and combining images of floating objects and haze on the water surface, fog concentration coefficients, and atmospheric light values ​​corresponding to different pixels to obtain the approximate transmittance of each pixel includes: ； in, Indicates position The approximate transmittance corresponding to each pixel. Indicates the fog concentration coefficient. Indicates position The atmospheric light value corresponding to that pixel. This indicates taking the minimum value.

7. The method for acquiring defogging images of floating objects in a hazy environment according to claim 6, characterized in that, The method of using Gaussian attenuation fitting and combining the atmospheric light values ​​corresponding to different pixels to compensate for the coarse transmittance of each pixel to obtain the Gaussian function transmittance of each pixel includes: ； in, This represents the correction factor. Represents the Euler number. To correct the constant, Indicates position The Gaussian transmittance corresponding to the pixel.

8. The method for acquiring defogging images of floating objects in a hazy environment according to claim 1, characterized in that, Based on the haze image of floating objects on the water surface, the atmospheric light value corresponding to each pixel, and the refined transmittance corresponding to each pixel, a dehazed image of the floating objects is obtained, including: ； in, Indicates the position of floating objects in the dehazed image. The value corresponding to the pixel at that location. This indicates the location of floating debris in the normalized haze image obtained by normalizing the haze image. The value corresponding to the pixel at that location. Indicates position The atmospheric light value corresponding to that pixel. Indicates position The refined transmittance corresponding to each pixel. This indicates taking the maximum value.

9. A system for acquiring images of floating objects in hazy environments, characterized in that, include: The first image acquisition module acquires images of floating objects and haze on the water surface. The second image acquisition module acquires the brightness entropy map and texture feature map corresponding to the haze image of floating objects on the water surface. The concentration distribution map acquisition module obtains the fog concentration distribution map based on the brightness entropy map and texture feature map; The coefficient acquisition module obtains the fog concentration coefficient based on the minimum fog concentration value, maximum fog concentration value, and average fog concentration value of the fog concentration distribution map. The local light value acquisition module acquires the atmospheric light values ​​corresponding to different pixels based on the haze image of floating objects on the water surface. The transmittance acquisition module utilizes the dark channel prior and combines it with images of floating objects and haze on the water surface, fog concentration coefficient, and atmospheric light values ​​corresponding to different pixels to obtain the rough transmittance corresponding to each pixel. The compensation module uses the Gaussian attenuation fitting method and combines the atmospheric light value corresponding to each pixel to compensate for the coarse transmittance of each pixel, so as to obtain the Gaussian function transmittance of each pixel. The refinement module uses a guided filtering method to refine the Gaussian function transmittance corresponding to each pixel to obtain the refined transmittance corresponding to each pixel. The third image acquisition module acquires a defogging image of the floating objects based on the haze image of the floating objects on the water surface, the atmospheric light value corresponding to each pixel, and the refined transmittance corresponding to each pixel.

10. A method for detecting floating objects, characterized in that, include: A defogging image of floating objects in a hazy environment is obtained by using the method for acquiring defogging images of floating objects in a hazy environment as described in any one of claims 1 to 8; An improved YOLOv5s model was used to identify floating objects in dehazed images. The improved YOLOv5s model includes a backbone network, a neck network, and a detection network with a four-layer detection structure connected in sequence. The backbone network includes a first CBS module, a second CBS module, a first C3 module, a third CBS module, a second C3 module, a fourth CBS module, a third C3 module, a fifth CBS module, a C3-MS-CBAM module, and an SPPF module connected in sequence; the second CBS module is connected to the neck network through the sixth CBS module, the third CBS module is connected to the neck network through the seventh CBS module, and the fourth CBS module is connected to the neck network through the eighth CBS module. The C3-MS-CBAM module includes the ninth CBS module, the tenth CBS module, the eleventh CBS module, the merging module, the twelfth CBS module, and the MS-CBAM module connected in sequence. The C3-MS-CBAM module also includes the thirteenth CBS module. The ninth CBS module is also connected to the merging module through the thirteenth CBS module. The MS-CBAM module includes a multi-scale channel attention module, a spatial attention module, and a multiplication module connected in sequence. The multi-scale channel attention module is also directly connected to the multiplication module. In the multi-scale channel attention module, the initial feature map of the input is processed through the first image processing path and the second image processing path to obtain the first processed image and the second processed image, respectively. The first processed image and the second processed image are then merged and activated by a function to obtain the third processed image. Finally, the third processed image is multiplied with the initial feature map to obtain the output image. In the first image processing path, global average pooling, pointwise convolution, function activation, and pointwise convolution are performed sequentially on the initial feature map. In the second image processing path, pointwise convolution, function activation, and pointwise convolution operations are performed sequentially on the initial feature map.

Images