WO2023065665A1 - 图像处理方法、装置、设备、存储介质及计算机程序产品 - Google Patents

图像处理方法、装置、设备、存储介质及计算机程序产品 Download PDF

Info

Publication number
WO2023065665A1
WO2023065665A1 PCT/CN2022/095025 CN2022095025W WO2023065665A1 WO 2023065665 A1 WO2023065665 A1 WO 2023065665A1 CN 2022095025 W CN2022095025 W CN 2022095025W WO 2023065665 A1 WO2023065665 A1 WO 2023065665A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
noise reduction
pixel
processed
area
Prior art date
Application number
PCT/CN2022/095025
Other languages
English (en)
French (fr)
Inventor
史超超
Original Assignee
深圳市慧鲤科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳市慧鲤科技有限公司 filed Critical 深圳市慧鲤科技有限公司
Publication of WO2023065665A1 publication Critical patent/WO2023065665A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/50—Image enhancement or restoration by the use of more than one image, e.g. averaging, subtraction
    • G06T5/70—
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20212—Image combination
    • G06T2207/20221—Image fusion; Image merging

Definitions

  • Embodiments of the present disclosure relate to but are not limited to the field of computer vision, and in particular, relate to an image processing method, device, device, storage medium, and computer program product.
  • Image noise reduction has always been one of the hotspots in image processing research.
  • Image noise reduction can improve the image with noise, which is beneficial to reduce the image quality degradation caused by image noise interference.
  • Noise reduction can effectively improve the image quality, increase the signal-to-noise ratio, and better reflect the information carried by the original image.
  • Embodiments of the present disclosure provide an image processing method, device, device, storage medium, and computer program product.
  • the embodiment of the present disclosure provides an image processing method, including: determining the semantic segmentation image of the image to be processed, and determining the texture image of the image to be processed; fusing the semantic segmentation image and the texture image , to obtain a fused image; based on the fused image, determine a first noise reduction parameter corresponding to the image to be processed; and based on the first noise reduction parameter, perform noise reduction processing on the image to be processed.
  • the fused image since the fused image is obtained by fusing the semantic segmentation image and the texture image, the fused image represents the result of the fusion of the texture of the image to be processed and the different semantic regions in the image to be processed, so that the image corresponding to the image to be processed can be determined based on the fused image
  • the first noise reduction parameter so that the determined first noise reduction parameter can be based on the texture of the image to be processed and the fusion result of different semantic regions, and then the image to be processed can be denoised through the first noise reduction parameter, and the image to be processed can be denoised
  • the details in the image to be processed will not be lost, which improves the noise reduction ability of the image to be processed.
  • the determining the first noise reduction parameters corresponding to the image to be processed based on the fused image includes: determining the first noise reduction parameters corresponding to the pixels in the image to be processed based on the fused image.
  • Noise reduction parameters; wherein, pixels with different pixel values correspond to different first noise reduction parameters; performing noise reduction processing on the image to be processed based on the first noise reduction parameters includes: based on the pixel corresponding to The first noise reduction parameter is to perform noise reduction processing on the pixel.
  • pixels with different pixel values correspond to different first noise reduction parameters
  • the pixels are subjected to noise reduction processing based on the first noise reduction parameters corresponding to the pixels, so that according to the fusion of texture features and semantic features of pixels in the image to be processed,
  • the pixel-level noise reduction of the image to be processed is realized, and then the noise of the pixels of the image to be processed can be accurately reduced, and the noise reduction capability of the image to be processed is improved.
  • the first noise reduction parameter includes at least two sub-noise reduction parameters, and the noise reduction strengths corresponding to the at least two sub-noise reduction parameters are different; the determination of the image to be processed is based on the fused image
  • the corresponding first noise reduction parameters include: dividing the image to be processed into at least two regions based on the fused image; respectively determining sub-noise reduction parameters corresponding to the regions according to the at least two sub-noise reduction parameters
  • the step of performing noise reduction processing on the image to be processed based on the first noise reduction parameter includes: performing noise reduction processing on the regions based on sub-noise reduction parameters corresponding to the regions.
  • the image to be processed is divided into at least two regions, the sub-noise reduction parameters corresponding to the regions are respectively determined, and the region is denoised based on the sub-noise reduction parameters corresponding to the regions, so that the same region uses the same Noise reduction parameters, different regions use different noise reduction parameters, which can not only improve the noise reduction ability of the image to be processed, but also quickly denoise the image to be processed.
  • the determining the texture image of the image to be processed includes: determining a gradient image sequence corresponding to the image to be processed; the gradient image sequence includes at least two normalized gradient images of different scales ; Based on the gradient image sequence, determine the texture image.
  • the texture image is determined based on the gradient image sequence, so that the texture image is obtained by combining gradient images of different scales, so that the determined texture image can accurately reflect the image to be processed. texture.
  • the determining the gradient image sequence corresponding to the image to be processed includes: determining a first image sequence; the first image sequence includes N first images, and among the N first images The i-th first image is obtained by down-sampling the (i-1)-th first image, and the first first image is the image to be processed; N is an integer greater than or equal to 2, and i is greater than or equal to is an integer of 2; performing image gradient processing on the first images in the first image sequence respectively to obtain the gradient image sequence.
  • the first image sequence is obtained by down-sampling the image to be processed to different degrees, and then image gradient processing is performed on the first image in the first image sequence to obtain a gradient image sequence, so that the obtained gradient image sequence can reflect the first image sequence.
  • Gradient information of an image sequence, and then based on the gradient image sequence, the texture image can be accurately determined.
  • performing image gradient processing on the first images in the first image sequence to obtain the gradient image sequence includes: respectively performing image gradient processing on the first images in the first image sequence Each image is subjected to noise reduction processing using a second noise reduction parameter to obtain a second image sequence; image gradient processing is performed on the second images in the second image sequence respectively to obtain the gradient image sequence.
  • the first image in the first image sequence is denoised first, and then the second image sequence obtained by denoising processing is subjected to image gradient processing to obtain a gradient image sequence, the impact of image noise on gradient calculation can be reduced. Influence, so that the obtained gradient image sequence is accurate.
  • the determination of the texture image based on the gradient image sequence includes: the first upsampled image obtained by upsampling the Nth gradient image, and the (N-1)th gradient image
  • the images are combined to obtain a combined image; N is an integer greater than or equal to 2; when N is 2, the combined image is determined as the texture image; when N is greater than 2, the obtained jth
  • the merged image is upsampled to obtain the (j+1)th upsampled image; the (j+1)th upsampled image is merged with the (N-1-j)th gradient image to obtain the (j+1)th upsampled image and the (N-1-j) gradient image to obtain the ( j+1) merged images; j is an integer greater than or equal to 1; determine the merged image obtained last time as the texture image.
  • the obtained j-th merged image is up-sampled to obtain the (j+1)-th up-sampled image, and the (j+1)-th up-sampled image is combined with the (N-1-j)-th gradient image Merge to obtain the (j+1)th merged image, determine the last merged image as the texture image, the texture image can be obtained by merging each gradient image in the gradient image sequence, so that the texture image can accurately reflect the Handles the texture of the image.
  • the merging the semantically segmented image and the texture image to obtain a fused image includes: determining a weight value corresponding to at least one region in the semantically segmented image; Modifying the pixel value of at least one area of the at least one area to a weight value corresponding to the at least one area to obtain a weight image; fusing the weight image and the texture image to obtain the fusion image.
  • the pixel value of at least one region in the semantic segmentation image can be modified to the weight value corresponding to at least one region to obtain a weight image; the weight image and the texture image are fused to obtain a fusion image, so as to provide different Different weight values are set for the semantic regions, and then the noise reduction strength can be determined based on the weight values, which improves the noise reduction ability of the image to be processed.
  • the merging the weight image and the texture image to obtain the fused image includes: correspondingly subtracting pixel values in the texture image from pixel values in the weight image , to obtain the target image; modifying the pixel values in the target image greater than the first threshold to the first threshold, and modifying the pixel values in the target image smaller than the second threshold to the second threshold, The fused image is obtained; the first threshold is greater than the second threshold.
  • the fused image can be determined based on the target image obtained by correspondingly subtracting the pixel values in the texture image from the pixel values in the weight image, thereby providing a way to realize the fusion of the weight image and the texture image, so that the fused image can be accurately expresses the information carried by the image to be processed; and, by modifying the pixel values in the target image greater than the first threshold to the first threshold, and modifying the pixel values in the target image smaller than the second threshold to the second threshold, thus Based on the fused image, the texture area and flat area of the image to be processed can be easily distinguished, so that differential noise reduction can be carried out for the texture area and flat area in the image to be processed, and the noise of the image to be processed can be reduced without losing the noise of the image to be processed.
  • the details in the image are processed, and the noise reduction ability of the image to be processed is improved.
  • the determining the first noise reduction parameters respectively corresponding to the pixels in the image to be processed based on the fused image includes: obtaining the pixel value of the first pixel in the fused image; according to the The size of the pixel value of the first pixel, setting a first noise reduction parameter for the second pixel in the image to be processed respectively corresponding to the first pixel; wherein, the first noise reduction parameter includes a standard deviation and a window size , if the pixel value of the mth first pixel is greater than the pixel value of the nth first pixel, then the standard deviation of the mth second pixel is smaller than the standard deviation of the nth second pixel, and the mth second pixel
  • the window size of the second pixel is smaller than the window size of the nth second pixel, m and n are different, and both m and n are integers greater than or equal to 1.
  • the first noise reduction parameters are set for the second pixels in the image to be processed respectively corresponding to the first pixels, so that the first pixels corresponding to the first pixels with different pixel values Different first noise reduction parameters are set for the second pixel, so that noise reduction can be accurately performed on each pixel of the image to be processed, and the noise reduction capability of the image to be processed is improved.
  • the pixel whose pixel value in the fused image is the minimum value indicates that the pixel in the first area in the corresponding image to be processed is a pixel in the flat area, so that the pixel is denoised with a greater noise reduction strength, so that the pixel can be denoised.
  • the pixels in the flat area are effectively denoised; the pixels with the maximum pixel value in the fusion image represent the pixels in the third area of the corresponding image to be processed as pixels in the texture area, so a smaller noise reduction strength is used to This pixel is denoised, so that when denoising the pixels in the texture area, less texture information is lost; and, for the pixels in the second area except the first area and the third area, a medium denoising Noise reduction can be performed according to the noise strength, which can make the obtained image smooth after noise reduction.
  • the at least two regions include a fourth region and a fifth region; said dividing the image to be processed into at least two regions based on the fused image includes: obtaining the fused image The pixel value of the first pixel; according to the size of the pixel value of the first pixel, the image to be processed is divided into the fourth area and the fifth area; wherein, the same as the second pixel in the fourth area The pixel values of the first pixels in the fused image respectively corresponding to the third threshold; the pixel values of the first pixels in the fused image respectively corresponding to the second pixels in the fifth area are less than or equal to the A third threshold; wherein, the noise reduction strength of the sub-noise reduction parameters corresponding to the fourth area is smaller than the noise reduction strength of the sub-noise reduction parameters corresponding to the fifth area.
  • the pixel value in the fused image is greater than the third threshold, it is determined that the pixels in the fourth area in the corresponding image to be processed have more texture, so that the pixels in the fourth area are denoised with a smaller noise reduction strength , so that when denoising pixels with more textures, less texture information is lost; in the case where the pixel value in the fused image is less than or equal to the third threshold, determine the pixel value of the fifth region in the corresponding image to be processed There are less textures, so that the pixels in the fifth area are denoised with a greater noise reduction strength, so that the pixels with less textures can be effectively denoised, and the noise reduction capability of the image to be processed is improved.
  • an embodiment of the present disclosure provides an image processing device, including: a determining unit, configured to determine a semantically segmented image of an image to be processed, and determine a texture image of the image to be processed; a fusion unit, configured to process the The semantic segmentation image is fused with the texture image to obtain a fused image; the determining unit is further configured to determine a first noise reduction parameter corresponding to the image to be processed based on the fused image; the noise reduction unit is configured to The first noise reduction parameter is to perform noise reduction processing on the image to be processed.
  • an embodiment of the present disclosure provides an image processing device, including: a memory and a processor, the memory stores a computer program that can run on the processor, and the processor implements the computer program when executing the computer program The image processing method described above.
  • an embodiment of the present disclosure provides a computer-readable storage medium, the computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to Implement the image processing method described above.
  • an embodiment of the present disclosure further provides a computer program product, including computer readable codes, or a computer readable storage medium bearing computer readable codes, when the computer readable codes are stored in a processor of an electronic device When running in the electronic device, the processor in the electronic device executes the above-mentioned image processing method.
  • FIG. 1 is a schematic diagram of an implementation flow of an image processing method provided by an embodiment of the present disclosure
  • FIG. 2 is a schematic diagram of an implementation of an image processing method provided by an embodiment of the present disclosure
  • FIG. 3 is a schematic diagram of an implementation flow of another image processing method provided by an embodiment of the present disclosure.
  • FIG. 4 is a schematic diagram of an implementation flow of another image processing method provided by an embodiment of the present disclosure.
  • FIG. 5 is a schematic diagram of an implementation flow of another image processing method provided by an embodiment of the present disclosure.
  • FIG. 6 is a schematic diagram of an implementation of determining a texture image based on a gradient image sequence provided by an embodiment of the present disclosure
  • FIG. 7 is a schematic diagram of an implementation flow of an image processing method provided by another embodiment of the present disclosure.
  • FIG. 8 is a schematic diagram of the composition and structure of an image processing device provided by an embodiment of the present disclosure.
  • FIG. 9 is a schematic diagram of hardware entities of an image processing device provided by an embodiment of the present disclosure.
  • first, second, etc. are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.
  • the traditional method of deep learning or three-dimensional block matching algorithm can effectively remove part of the noise under dark light.
  • the method based on deep learning generally builds a model for noise, simulates the noise distribution in dark and light scenes, and generates data pairs.
  • the ability to reduce noise depends on the difference between the noise model and the real noise. The difference, the noise reduction ability is good and bad, the reusability is low, and the robustness is poor.
  • the traditional method of BM3D performs global or specific frequency band noise reduction through spatial or frequency domain characteristics. Although it can reduce the noise in the image, there are a lot of cases where image details are lost.
  • any image processing apparatus mentioned in the embodiments of the present application may be a processor or a chip, and the processor or chip may be applied to an image processing device.
  • any image processing apparatus mentioned in the embodiments of the present application may be an image processing device.
  • the image processing device may include an image processing component, for example, the image processing component may include a camera component.
  • the image processing device may include at least one or a combination of at least two of the following: camera, server, mobile phone (Mobile Phone), tablet computer (Pad), computer with wireless transceiver function, palmtop computer, desktop Computers, personal digital assistants, portable media players, smart speakers, navigation devices, smart watches, smart glasses, smart necklaces and other wearable devices, pedometers, digital TVs, virtual reality (Virtual Reality, VR) terminal equipment, enhanced Reality (Augmented Reality, AR) terminal equipment, wireless terminals in Industrial Control, wireless terminals in Self Driving, wireless terminals in Remote Medical Surgery, Smart Grid ), wireless terminals in Transportation Safety, wireless terminals in Smart City, wireless terminals in Smart Home, vehicles in the Internet of Vehicles system, vehicle-mounted equipment, and vehicle-mounted modules and more.
  • FIG. 1 is a schematic diagram of an implementation flow of an image processing method provided by an embodiment of the present disclosure. As shown in FIG. 1 , the method is applied to an image processing device, and the method includes:
  • S101 Determine a semantically segmented image of an image to be processed, and determine a texture image of the image to be processed.
  • the image to be processed may be a raw image.
  • the original image may be an image obtained by image capture.
  • the original image may be an image frame in a video.
  • the original image may be an image read locally, downloaded, or read from other devices (for example, a hard disk, a USB flash drive, or other terminals, etc.).
  • the image to be processed may be an image obtained by performing at least one of the following processing on the original image: scaling processing, cropping processing, denoising processing, adding noise processing, grayscale processing, rotation processing, normalization deal with.
  • the original image may be scaled and then rotated to obtain the image to be processed.
  • Determining the semantically segmented image of the image to be processed may include: performing semantically segmented image on the image to be processed to obtain a semantically segmented image.
  • Determining the texture image of the image to be processed may include: performing texture detection on the image to be processed to obtain a texture image.
  • Performing semantic segmentation on the image to be processed to obtain a semantically segmented image may include: inputting the image to be processed into a semantic segmentation network (or called a semantic segmentation model), and performing semantic segmentation on the image to be processed through the semantic segmentation network to obtain a semantically segmented image.
  • the semantic segmentation network may be obtained by training a plurality of labeled first training images.
  • Semantic segmentation networks can include one of the following: Fully Convolution Networks (FCN), SegNet, U-Net, DeepLab v1, DeepLab v2, DeepLab v3, DenseNet, E-Net, Link-Net, masked area volumes Mask R-CNN, Pyramid Scene Parseing Network (PSPNet), RefineNet, Gated Feedback Refinement Network (G-FRNet), and the evolution of these networks.
  • Performing texture detection on the image to be processed to obtain a texture image may include: inputting the image to be processed into a texture detection network (or texture detection model), and performing texture detection on the image to be processed through the texture detection network to obtain a texture image.
  • the texture detection network may include: Deep Texture Encoding Network (Deep-TEN) and the like.
  • the texture detection network can also be called an edge segmentation network
  • the edge segmentation network can include one of the following: Richer Convolutional Features for Edge Detection (RCF) network based on richer features, an overall nested edge Detection (holistically-nested edge detection, HED) network, Canny edge detection network, and the evolution of these networks.
  • RCF Richer Convolutional Features for Edge Detection
  • HED overall nested edge Detection
  • Canny edge detection network Canny edge detection network
  • the pixel size of the semantic segmentation image and the texture image can be the same as the pixel size of the image to be processed.
  • the pixel size of the image to be processed is 800 ⁇ 600 or 800 ⁇ 600 ⁇ 3
  • 800 is the image to be processed
  • 600 is the pixel size of the image to be processed in the height direction
  • 3 is the number of channels of the image to be processed
  • the channels include red, green and blue (RGB) three channels
  • the pixel size of the semantic segmentation image and the texture image Both are 800 ⁇ 600.
  • the semantic segmentation network can judge the category to which each location belongs from the empirical knowledge learned from a large number of labeled training sets (including multiple labeled first training images) and the local features of each location.
  • Semantic segmentation of the image to be processed is to perform semantic classification of each pixel in the image to be processed, and to determine that each pixel in the image to be processed belongs to a certain object or background.
  • the identification value of the corresponding pixel in the semantic segmentation image can be determined as the first value
  • the identification value of the corresponding pixel in the semantic segmentation image can be determined as the value corresponding to the target object, N is an integer greater than or equal to 1, and there are N types of values corresponding to the target object, and the first value different.
  • the identification value of each pixel in the semantic segmentation image can be N+1 kinds of values, and N is the total number of object categories, so that based on the positions of different types of values in the semantic segmentation image, the background part and the The position of each type of object.
  • the semantically segmented image may be called a semantic mask (Segm mask).
  • the texture detection of the image to be processed is to determine whether each pixel in the image to be processed is a texture pixel or an edge pixel. When the value of a certain pixel in the texture image is larger, it indicates that the pixel corresponding to the pixel value is The higher the probability of a texture pixel; the smaller the value of a certain pixel in the texture image, the lower the probability that the pixel corresponding to the pixel value is a texture pixel.
  • S102 may be implemented in the following manner: obtain the pixel value corresponding to each pixel in the semantic segmentation image, obtain the pixel value corresponding to each pixel in the texture image, and obtain the pixel value corresponding to each pixel in the semantic segmentation image , and perform mathematical calculations corresponding to the pixel values corresponding to each pixel in the texture image to obtain a fused image.
  • the pixel value corresponding to each pixel in the semantic segmentation image may be the real pixel value of each pixel in the semantic segmentation image, or the pixel value corresponding to each pixel in the semantic segmentation image may be the same as each pixel in the semantic segmentation image The pixel values of the real pixel value map.
  • the pixel value corresponding to each pixel in the texture image may be the real pixel value of each pixel in the texture image, or the pixel value corresponding to each pixel in the texture image may be the real pixel mapping of each pixel in the texture image pixel value.
  • the mapped pixel value may have a mapping relationship with the real pixel value, for example, the mapped pixel value may have a one-to-one mapping relationship or a one-to-many mapping relationship with the real pixel value.
  • the mapped pixel value can be obtained by calculating the real pixel value, or the mapped pixel value can be obtained by using the real pixel value and a mapping relationship, and the mapping relationship includes the correspondence between the real pixel value and the mapped pixel value relation.
  • Performing mathematical calculations includes, but is not limited to, at least one of the following: multiplication, addition, subtraction, division, exponential operations, logarithmic operations, and the like.
  • the mathematical calculation is performed on the pixel value corresponding to each pixel in the semantic segmentation image and the pixel value corresponding to each pixel in the texture image, which may include: corresponding to each pixel in the semantic segmentation image Multiply the pixel value corresponding to each pixel in the texture image, or subtract the pixel value corresponding to each pixel in the texture image from the pixel value corresponding to each pixel in the semantic segmentation image.
  • the pixel value corresponding to the pixel at row a and column b in the texture image is subtracted from the pixel value corresponding to the pixel at row a and column b in the semantic segmentation image.
  • Both a and b are integers greater than or equal to 1, and a and b may be the same or different.
  • the pixel values of the pixels in the fused image may be within a preset range.
  • the preset range may be [0,1], thus, the pixel value of the pixel in the fused image may be any value in [0,1].
  • the pixel value of a certain pixel in the fused image may be 0, 0.128, 0.75 or 1, etc.
  • the pixel values of the pixels in the fused image may be data in a data set, and the data set may be a preset set including at least two values.
  • the data set may include ⁇ 0,1 ⁇ , ⁇ 0,0.5,1 ⁇ or ⁇ 0,0.2,0.4,0.6,0.8,1 ⁇ and so on.
  • the first noise reduction parameter and/or the second noise reduction parameter described below can be the noise reduction parameter of the filter
  • the filter can be a Gaussian filter or other filters
  • the Gaussian filter is a 2-dimensional convolution using a Gaussian kernel Product operator
  • other filters can include one of the following: median filter, mean filter, bilateral filter, maximum and minimum filter, guide filter, Sobel (Sobel) filter, Prewitt filter, Laplac Laplacian filter, etc.
  • the embodiments of the present disclosure do not limit the filter and the noise reduction parameters of the filter.
  • the first noise reduction parameter and/or the following second noise reduction parameter may include: standard deviation and/or window size, and the difference between the first noise reduction parameter and/or the following second noise reduction parameter may include the difference in standard deviation and /or the window size is different.
  • the filter may include a Gaussian kernel
  • the element values in the Gaussian kernel may be determined based on the standard deviation
  • the size of the Gaussian kernel may be determined based on the window size.
  • the fused image since the fused image is obtained by fusing the semantic segmentation image and the texture image, the fused image represents the fusion result of the texture of the image to be processed and different semantic regions in the image to be processed, so based on the fused image Determining a first noise reduction parameter corresponding to the image to be processed, so that the determined first noise reduction parameter can be based on the fusion result of the texture of the image to be processed and different semantic regions, and then perform noise reduction on the image to be processed by the first noise reduction parameter, While denoising the image to be processed, details in the image to be processed will not be lost, and the noise reduction capability of the image to be processed is improved.
  • FIG. 2 is a schematic diagram of the implementation of the image processing method provided by the embodiment of the present disclosure, which can be applied to an image processing device.
  • the image processing device can first obtain the image 21 to be processed, and then input the image 21 to be processed into Semantic segmentation network and texture detection network, the semantic segmentation image 22 is obtained through the semantic segmentation network, and the texture image 23 is obtained through the texture detection network.
  • the semantic segmentation image 22 and the texture image 23 can be fused (or called mask fusion) to obtain the fused image 24 .
  • the fusion image 24 and the image to be processed 21 are input to the noise reduction module, and the noise reduction module can determine the corresponding first noise reduction parameter in the image to be processed based on the fusion image; Noise processing, after the noise reduction module completes the noise reduction of the image to be processed, the processed image 25 is output.
  • determining the semantically segmented image of the image to be processed may include: downsampling the image to be processed using a first downsampling factor to obtain a downsampled image; performing semantic segmentation on the downsampled image to obtain a target segmented image; The target segmented image is up-sampled to obtain a semantically segmented image with the same size as the image to be processed.
  • the first downsampling factor may be an integer greater than or equal to 2.
  • the first downsampling factor may be 2, 4, 8 or 10, etc.
  • the size of the image used for semantic segmentation can be reduced by downsampling the image to be processed by using the first downsampling factor to obtain the downsampled image.
  • the size of the image to be processed is M ⁇ N and the first downsampling factor is 10
  • the downsampled image may be M/10 ⁇ N/10.
  • the image to be processed may be downsampled multiple times by using the first downsampling factor to obtain a downsampled image.
  • the size of the image to be processed is M ⁇ N
  • the first downsampling factor is 2
  • the M/2 ⁇ N/2 image can be obtained by downsampling first, and then the M/2 ⁇ N/2 image
  • it is down-sampled again to obtain an M/8 ⁇ N/8 image
  • the M/4 ⁇ N/4 image is obtained by down-sampling again.
  • An image of 8 ⁇ N/8 is determined as a downsampled image.
  • the down-sampled image can be input into the semantic segmentation network, and the target segmentation image can be output through the semantic segmentation network.
  • the pixel value of each pixel in the target segmented image represents the object to which the pixel belongs, so that the pixel area corresponding to each different object (including the background) in the downsampled image can be determined based on the target segmented image.
  • the first upsampling factor may be used to upsample the target segmented image to obtain a semantically segmented image.
  • the first upsampling factor may be the same as the first downsampling factor described above.
  • the target segmented image may be up-sampled once using the first up-sampling factor to obtain the semantically segmented image. For example, when the size of the target segmented image is M/10 ⁇ N/10 and the first upsampling factor is 10, the size of the semantically segmented image may be M ⁇ N.
  • the target segmented image may be upsampled multiple times by using the first upsampling factor to obtain the semantically segmented image.
  • the first upsampling factor is 2, it can be upsampled first to obtain an image of M/4 ⁇ N/4, and at M/4 ⁇ N/ On the basis of the image of 4, upsample again to obtain an image of M/2 ⁇ N/2, and then on the basis of an image of M/2 ⁇ N/2, upsample again to obtain an image of M ⁇ N, and the obtained by upsampling
  • the M ⁇ N image is the semantically segmented image.
  • the semantic segmentation is performed on the down-sampled image obtained by down-sampling the image to be processed, and then the target segmentation image obtained by the semantic segmentation is up-sampled to obtain the semantic segmentation image, thereby reducing the amount of calculation for obtaining the semantic segmentation image , which reduces the time-consuming to obtain semantically segmented images.
  • FIG. 3 is a schematic diagram of an implementation flow of another image processing method provided by an embodiment of the present disclosure. As shown in FIG. 3 , the method is applied to an image processing device, and the method includes:
  • S301 Determine a semantically segmented image of an image to be processed, and determine a texture image of the image to be processed.
  • S303 may be implemented in the following manner: acquire the pixel value of the first pixel in the fused image; A second pixel in the image is processed to set a first noise reduction parameter.
  • the noise reduction strength of the mth second pixel is smaller than that of the nth second pixel. If the pixel value of the mth first pixel is equal to the pixel value of the nth first pixel, then the noise reduction strength of the mth second pixel is equal to the noise reduction strength of the nth second pixel. If the pixel value of the mth first pixel is smaller than the pixel value of the nth first pixel, the noise reduction strength of the mth second pixel is greater than the noise reduction strength of the nth second pixel.
  • the first noise reduction parameter includes a standard deviation and a window size, if the pixel value of the mth first pixel is greater than the pixel value of the nth first pixel, then the standard deviation of the mth second pixel is less than The standard deviation of the nth second pixel, and the window size of the mth second pixel is smaller than the window size of the nth second pixel, m and n are different, and both m and n are greater than or equal to 1 integer.
  • the standard deviation of the mth second pixel is equal to the standard deviation of the nth second pixel
  • the mth second The pixel window size is equal to the window size of the nth second pixel. If the pixel value of the mth first pixel is smaller than the pixel value of the nth first pixel, the standard deviation of the mth second pixel is greater than the standard deviation of the nth second pixel, and the mth second The pixel window size is larger than the nth second pixel window size.
  • the m first pixel and the n first pixel may be any two first pixels in the fused image.
  • the mth second pixel is a pixel corresponding to the mth first pixel in the image to be processed
  • the nth second pixel is a pixel corresponding to the nth first pixel in the image to be processed.
  • the first noise reduction parameter is set for the second pixel in the image to be processed corresponding to the first pixel, so that the first pixel with different pixel value Different first noise reduction parameters are set for the corresponding second pixels, so the noise reduction can be accurately performed on each pixel of the image to be processed, and the noise reduction capability of the image to be processed is improved.
  • the first noise reduction parameters corresponding to the second pixels in the image to be processed can be respectively determined based on the pixel value of the first pixel in the fused image; the second pixel in the image to be processed can be compared with the The first pixels are in one-to-one correspondence.
  • the second pixels in the image to be processed are mapped to the first pixels with different pixel values in the fused image, and the corresponding first noise reduction parameters are different.
  • two first pixels in the fusion image whose pixel values are 1 and 0.8 map to the second pixel in the image to be processed have different first noise reduction parameters.
  • pixels with different pixel values correspond to different first noise reduction parameters, and based on the first noise reduction parameters corresponding to the pixels, the pixels are subjected to noise reduction processing, so that the texture characteristics and The fusion of semantic features can achieve pixel-level noise reduction of the image to be processed, and then can accurately denoise the pixels of the image to be processed, improving the noise reduction ability of the image to be processed.
  • Fig. 4 is a schematic diagram of the implementation flow of another image processing method provided by the embodiment of the present disclosure. As shown in Fig. 4, the method is applied to an image processing device.
  • the first noise reduction parameters include at least Two sub-noise reduction parameters, the noise reduction strengths corresponding to the at least two sub-noise reduction parameters are different, the method includes:
  • S401 Determine a semantically segmented image of an image to be processed, and determine a texture image of the image to be processed.
  • the image to be processed may be divided into at least two regions according to the pixel value of the first pixel in the fused image.
  • Embodiments of the present disclosure do not limit the number of at least two regions.
  • the number of at least two regions may be two, three, five or ten, etc.
  • the pixel values of the pixels in the fused image are in the range [0,1]
  • at least two pixel value ranges can be predetermined
  • at least two pixel value ranges are continuous and disjoint
  • the union of at least two pixel value ranges is [ 0,1]
  • at least two pixel value ranges correspond to at least two regions one-to-one.
  • the minimum value of the pixel value is 0, and the maximum value of the pixel value is 1.
  • the second pixel in the image to be processed corresponding to the first pixel with a pixel value of 0 is a pixel in the flat area, requiring higher noise reduction strength
  • the first pixel with a pixel value of 1 corresponds to the second pixel in the image to be processed is a pixel in the texture area, requiring a lower noise reduction strength
  • the second pixel in the image to be processed corresponding to the first pixel whose pixel value is greater than 0 and less than 1 is a pixel in the texture area, and requires an intermediate noise reduction strength.
  • the noise reduction strength of sub-noise reduction parameters can be characterized by standard deviation and window size.
  • the noise reduction strength of the sub-noise reduction parameter is greater, the larger the standard deviation is, the larger the window size is; when the noise reduction strength of the sub-noise reduction parameter is smaller, the smaller the standard deviation is, the smaller the window size is.
  • An implementation manner in which the noise reduction strengths of the sub-noise reduction parameters corresponding to the first area, the second area, and the third area are sequentially reduced may be: the corresponding sub-noise reduction parameters of the first area include The standard deviation is greater than the standard deviation included in the corresponding sub-noise reduction parameters of the second area, and/or, the window size included in the corresponding sub-noise reduction parameters of the first area is greater than the corresponding sub-noise reduction parameters included in the second area and the standard deviation included in the corresponding sub-noise reduction parameters of the second area is greater than the standard deviation included in the corresponding sub-noise reduction parameters of the third area, and/or, the corresponding sub-noise reduction parameters of the second area include The window size included in the parameter is larger than the window size included in the corresponding sub-noise reduction parameter of the third area.
  • the pixel whose pixel value in the fused image is the minimum value represents that the pixel in the first area in the corresponding image to be processed is a pixel in the flat area, so that the pixel is denoised with a greater denoising force,
  • the pixels in the flat area can be effectively denoised;
  • the pixel with the maximum pixel value in the fusion image represents the pixel in the third area in the corresponding image to be processed as the pixel in the texture area, so that a smaller denoising method is used.
  • the at least two regions include a fourth region and a fifth region; said dividing the image to be processed into at least two regions based on the fused image includes: obtaining the fused image The pixel value of the first pixel; according to the size of the pixel value of the first pixel, the image to be processed is divided into the fourth area and the fifth area; wherein, the same as the second pixel in the fourth area The pixel values of the first pixels in the fused image respectively corresponding to the third threshold; the pixel values of the first pixels in the fused image respectively corresponding to the second pixels in the fifth area are less than or equal to the The third threshold: the noise reduction strength of the sub-noise reduction parameters corresponding to the fourth area is smaller than the noise reduction strength of the sub-noise reduction parameters corresponding to the fifth area.
  • the pixel value in the fused image is greater than the third threshold, it is determined that the pixels in the fourth region in the corresponding image to be processed have more texture, so that the pixels in the fourth region are treated with a smaller noise reduction strength.
  • Perform denoising so that less texture information is lost when denoising pixels with more texture;
  • determine the fifth area in the corresponding image to be processed The pixels with less texture have less texture, so the pixels in the fifth area are denoised with a greater denoising strength, so that the pixels with less texture can be effectively denoised, and the denoising capability of the image to be processed is improved.
  • the image to be processed is divided into at least two regions, the sub-noise reduction parameters corresponding to the regions are determined respectively, and the noise reduction processing is performed on the region based on the sub-noise reduction parameters corresponding to the regions, so that the same time
  • One area uses the same noise reduction parameter, and different areas use different noise reduction parameters, which can not only improve the noise reduction ability of the image to be processed, but also quickly denoise the image to be processed.
  • FIG. 5 is a schematic diagram of an implementation flow of another image processing method provided by an embodiment of the present disclosure. As shown in FIG. 5 , the method is applied to an image processing device, and the method includes:
  • S501 may be implemented in the following manner: using a multi-scale Canny detection algorithm for the image to be processed to determine a gradient image sequence corresponding to the image to be processed.
  • S501 may be implemented in the following manner: determine a first image sequence; the first image sequence includes N first images, and the i-th first image in the N first images is (i-1) first images are obtained by down-sampling, the first first image is the image to be processed; N is an integer greater than or equal to 2, and i is an integer greater than or equal to 2; The first image in an image sequence is subjected to image gradient processing to obtain the gradient image sequence.
  • the image to be processed is down-sampled to different degrees to obtain the first image sequence, and then the image gradient processing is performed on the first image in the first image sequence to obtain the gradient image sequence, and the obtained gradient image sequence
  • the gradient information of the first image sequence can be reflected, and the texture image can be accurately determined based on the gradient image sequence.
  • the i-th first image among the N first images may be obtained by down-sampling the (i-1)-th first image by using the second down-sampling factor.
  • the first first image in the N first images is the image to be processed
  • the second first image is obtained by downsampling the first first image using the second downsampling factor
  • the third first image An image is obtained by downsampling the second first image by using the second downsampling factor, and so on until N first images are obtained.
  • the second downsampling factor may be an integer greater than or equal to 2.
  • the second downsampling factor may be 2, 4, 8 or 10, etc.
  • the number of first images included in the first image sequence may be determined based on actual requirements (such as the computing capability of the image processing device). In the case of high computing power of the image processing device, the number of first images included in the first image sequence can be set larger; in the case of low computing power of the image processing device, the first image can be set to The number of first images included in the sequence is set to be small.
  • the first image sequence may respectively include: the first image of M/8 ⁇ N/8, the first image of M/4 ⁇ N/4 first image, M/2 ⁇ N/2 first image, M ⁇ N first image.
  • the first image sequence may be referred to as an image pyramid
  • the image pyramid includes at least two first images arranged from small to large in size from top to bottom.
  • the pixel value of the pixel in the gradient image in the gradient image sequence may be obtained by performing gradient calculation on the corresponding pixel value of the pixel in the first image.
  • dx(i, j) can be the gradient of the i-th row and j-th column pixel in the x direction in the first image, and I(i+1, j) can be the (i+1)-th row in the first image
  • the pixel value of the j column pixel, I (i, j) can be the pixel value of the i row and j column pixel in the first image;
  • dy (i, j) can be the i row j column pixel in the first image
  • I(i,j+1) may be the pixel value of the pixel in row i and column (j+1) in the first image.
  • I(i-1, j) can be the pixel value of the j-th column pixel in the (i-1) row in the first image;
  • I(i, j-1) can be the i-th row ( j-1)
  • the gradient value of the pixel in row i and column j can be determined based on dx(i,j) and dy(i,j).
  • the pixel value of the i-th row and j-column pixel in the gradient image corresponding to the first image is the gradient value of the i-th row and j-column pixel.
  • the pixel value in row i and column j may be determined as the gradient value of the pixel in row i and column j.
  • elements of 0 or other values can be added outside the edge pixel, such as 1 or 0.5, etc., so that through the above calculation method, first determine the gradient dx(i,j) of the pixel in the i-th row and the jth column in the x direction and the gradient dy(i,j) in the y direction, and then pass dx(i,j) and dy(i ,j), determine the gradient value of the i-th row and j-th column pixel.
  • the gradient value of the i-th row and j-column pixel is set to a specified value, and the specified value can be in the range of [0,1] Any value of , for example, the specified value can be 0, 0.5, or 1, etc.
  • performing image gradient processing on the first images in the first image sequence to obtain the gradient image sequence may include: respectively performing image gradient processing on the first images in the first image sequence The first images are subjected to noise reduction processing using the second noise reduction parameters to obtain a second image sequence; image gradient processing is respectively performed on the second images in the second image sequence to obtain the gradient image sequence.
  • the noise reduction processing can be performed on the first image in the first image sequence first, and then the image gradient processing is performed on the second image sequence obtained by the noise reduction processing to obtain a gradient image sequence, thereby reducing the impact of image noise on The effect of gradient calculation, so that the resulting gradient image sequence is accurate.
  • the second noise reduction parameter may be a preset noise reduction parameter, and the same or different second noise reduction parameters may be used for noise reduction of different first images.
  • the selection of the second noise reduction parameter can also be related to the parameters when the image is captured. For example, when the image is captured with a higher sensitivity value and/or a higher brightness value, you can use The second noise reduction parameter with stronger noise reduction strength; when the image is captured with a lower sensitivity value and/or lower brightness value, the second noise reduction parameter with weaker noise reduction strength can be used, wherein , the International Organization for Standardization stipulates that the photosensitive quantification is ISO (International Organization for Standardization).
  • the gradient image sequence includes at least two gradient images of different scales (that is, different pixel sizes) (that is, N gradient images, and the sizes of the N gradient images are arranged from large to small), it is possible to make at least two gradient images of different scales Combine to get texture image.
  • S502 can be implemented in the following manner: the first upsampled image obtained by upsampling the Nth gradient image is merged with the (N-1)th gradient image to obtain a merged image (That is, the first merged image); determine the merged image as the texture image.
  • N gradient images include the first gradient image and the second gradient image
  • the size of the first gradient image is larger than the size of the second gradient image
  • the second gradient image can be
  • the image obtained by upsampling is combined with the first gradient image to obtain a combined image, and the combined image is determined as the texture image.
  • S502 can be implemented in the following manner: the first upsampled image obtained by upsampling the Nth gradient image is merged with the (N-1)th gradient image to obtain a merged image; Perform up-sampling on the obtained j-th combined image to obtain the (j+1)-th up-sampled image; perform the (j+1)-th up-sampled image with the (N-1-j)-th gradient image Merge to obtain the (j+1)th merged image; j is an integer greater than or equal to 1; determine the merged image obtained last time as the texture image.
  • the first upsampled image obtained by upsampling the fourth gradient image can be merged with the third gradient image to obtain a merged image (ie, the first merged image); then the obtained The first merged image is upsampled to obtain the second upsampled image, and the second upsampled image is merged with the second gradient image to obtain the second merged image; then the obtained second merged image Perform upsampling to obtain the third upsampled image, merge the third upsampled image with the first gradient image to obtain the third merged image; after obtaining the third merged image, since there are no other gradient images Therefore, the third merged image is the merged image obtained last time, and the merged image obtained last time is determined as the texture image.
  • a merged image ie, the first merged image
  • the obtained The first merged image is upsampled to obtain the second upsampled image
  • the second upsampled image is merged with the second gradient image to obtain the second merged image
  • the obtained second merged image Perform ups
  • the size of the first upsampled image can be the same as that of the (N-1)th gradient image.
  • the (j+1)th upsampled image may have the same size as the (N-1-j)th gradient image.
  • the manner of merging each upsampled image and the corresponding gradient image may include: correspondingly performing the pixel value of each pixel in each upsampled image with the pixel value of each pixel in the corresponding gradient image A mathematical operation (for example, at least one of the following operations: multiplication, addition, subtraction, division, exponential operation, logarithmic operation, etc.), the obtained result is determined to be a merged image.
  • a mathematical operation for example, at least one of the following operations: multiplication, addition, subtraction, division, exponential operation, logarithmic operation, etc.
  • merging the first upsampled image and the (N-1)th gradient image to obtain the first merged image may include: multiplying each pixel value in the first upsampled image by the target Coefficient, get the first weighted image, multiply the pixel value of each pixel in the first weighted image with the pixel value of each pixel in the (N-1)th gradient image to get the first merged image .
  • Merging the (j+1)th upsampled image with the (N-1-j)th gradient image to obtain the (j+1)th merged image may include: combining the (j+1th) Each pixel value in the (j+1)th upsampled image is multiplied by the target coefficient to obtain the (j+1)th weighted image, and the pixel value of each pixel in the (j+1)th weighted image is compared with the (N- The pixel values of each pixel in the 1-j) gradient images are multiplied correspondingly to obtain the (j+1)th merged image.
  • the target coefficient can be a number greater than 1 or less than 1.
  • the target coefficient can be 0.5, 0.8, 0.9, 1.1 or 1.2, etc.
  • the jth merged image obtained is upsampled to obtain the (j+1)th upsampled image, and the (j+1)th upsampled image is combined with the (N-1-j)th Gradient images are merged to obtain the (j+1)th merged image, and the last merged image is determined as a texture image.
  • the texture image can be obtained by merging each gradient image in the gradient image sequence, so that the texture image can be accurately Reflects the texture of the image being processed.
  • the texture image is determined based on the gradient image sequence, so that the texture image is obtained by combining gradient images of different scales, so that the texture image can be determined accurately Reflect the texture of the image to be processed.
  • Fig. 6 is a schematic diagram of an implementation of determining a texture image based on a gradient image sequence provided by an embodiment of the present disclosure, which can be applied to an image processing device.
  • the image processing device can obtain a gradient image sequence, and the gradient image sequence includes 4 Gradient image, respectively gradient image 61 (can be the 4th gradient image of above-mentioned description), gradient image 62 (can be the 3rd gradient image of above-mentioned description), gradient image 63 (can be the second of above-mentioned description gradient image) and gradient image 64 (which may be the first gradient image described above), the four gradient images are determined by the image to be processed.
  • gradient image 61 can be the 4th gradient image of above-mentioned description
  • gradient image 62 can be the 3rd gradient image of above-mentioned description
  • gradient image 63 can be the second of above-mentioned description gradient image
  • gradient image 64 which may be the first gradient image described above
  • the gradient image 61 is up-sampled to obtain an image 65 , and the corresponding pixels in the image 65 and the gradient image 62 are multiplied to obtain a merged image 66 .
  • the merged image 66 is up-sampled to obtain an image 67 , and the corresponding pixels in the image 67 and the gradient image 63 are multiplied to obtain a merged image 68 .
  • the merged image 68 is up-sampled to obtain an image 69 , and the corresponding pixels in the image 69 and the gradient image 64 are multiplied to obtain a merged image 70 .
  • the merged image 70 is the merged image obtained last time, so the merged image 70 can be determined as the texture image of the image to be processed.
  • Fig. 6 only gives an example of a process for determining a texture image based on a gradient image sequence.
  • the thickness, texture and grayscale of the lines in each image in Figure 6 do not represent the real thickness, texture and grayscale, and the actual thickness, texture and grayscale can be changed accordingly according to the actual situation.
  • the scales of different parts of the various images in Figure 6 do not represent true scales.
  • FIG. 7 is a schematic diagram of an implementation flow of an image processing method provided by another embodiment of the present disclosure. As shown in FIG. 7, the method is applied to an image processing device, and the method includes:
  • Semantic segmentation images can include multiple different regions, different regions correspond to different objects, and different regions are identified by different labels.
  • a semantically segmented image may include label categories such as 0, 1, 2, 3, and 4.
  • the sky area is 0, which is displayed as blue
  • the green leaf area is 1, which is displayed as green
  • the building area is 2.
  • the ground area is 3, which is displayed in purple
  • other types of areas are represented by 4, which is displayed in red.
  • each pixel in the semantically segmented image can be a value among 0, 1, 2, 3, 4.
  • the value range of the weight value may be [-1,1], and the embodiment of the present disclosure does not limit the value range of the weight value, for example, the value range of the weight value may also be [-2,2] , [0,1] or [-255,255], etc.
  • the sky is relatively clear, so the weight value corresponding to the sky area may be larger, and the content carried by the green leaves is relatively rich, so the weight value corresponding to the green leaf area may be smaller.
  • the corresponding weight values can be 1, 0.2, -0.3, 0.2, and 0 respectively.
  • the pixel values of 1, 2, 3, and 4 are replaced by 1, 0.2, -0.3, 0.2, and 0, respectively, to obtain the weight image.
  • S704 may be implemented in the following manner: determine a weight image corresponding to the texture image, and fuse the weight image and the weight image to obtain a fusion image.
  • the pixel value of each pixel in the texture image can be replaced by the deciles corresponding to the pixel value to obtain the weight image.
  • the pixel value of the kth pixel in the texture image is 0.512
  • the pixel value of the kth pixel is modified by 0.5.
  • the pixel value of at least one region in the semantic segmentation image can be modified to the weight value corresponding to at least one region to obtain a weight image; the weight image and texture image are fused to obtain a fused image, so as to be processed
  • Different weight values are set for different semantic regions in the image, and then the noise reduction strength can be determined based on the weight value, which improves the noise reduction ability of the image to be processed.
  • S704 may be implemented in the following manner: correspondingly subtracting the pixel values in the texture image from the pixel values in the weight image to obtain a target image; based on the target image, determining the fusion image.
  • S704 may be implemented in the following manner: correspondingly multiply the pixel values in the texture image by the pixel values in the weight image to obtain a target image; based on the target image, determine the fusion image.
  • the obtained target image can be directly determined as the fused image.
  • the pixel values in the target image that are larger than the first threshold are modified to the first threshold, and the pixel values in the target image that are smaller than the second threshold are modified to the second threshold. Threshold to obtain the fused image; the first threshold is greater than the second threshold.
  • the first threshold may be 1, and the second threshold may be 0.
  • the fused image can be determined based on the target image obtained by subtracting the pixel values in the texture image from the pixel values in the weight image, thus providing a way to realize the fusion of the weight image and the texture image, so that the fusion
  • the image can accurately express the information carried by the image to be processed; and, by modifying the pixel value greater than the first threshold in the target image to the first threshold, and modifying the pixel value in the target image smaller than the second threshold to the second Threshold, so that based on the fused image, the textured area and the flat area of the image to be processed can be easily distinguished, so that differential noise reduction can be performed on the textured area and the flat area in the image to be processed.
  • the details in the image to be processed will be lost, and the noise reduction ability of the image to be processed will be improved.
  • based on the deep learning segmentation algorithm different regions of the scene are classified to obtain a semantically segmented image; the image texture boundary is calculated using the image gradient information, and a mask (corresponding to the above texture image) is generated by combining the classification information; using Generate masks for details and flat areas (corresponding to the above-mentioned fused image), and use different noise reduction strengths for noise reduction.
  • fast guided filtering or fast bilateral filtering is used as the basic noise reduction module
  • the mask (corresponding to the above-mentioned fused image) is used as the noise reduction input parameter to control the noise reduction range (corresponding to the above-mentioned window size) and Gaussian standard deviation size (corresponding to the above standard deviation); for example: when the mask value is smaller, it means that this area is a flat area, and the noise reduction radius and Gaussian standard deviation function need to be increased accordingly, and the reduction On the contrary, when the mask value is larger, it means that the details of the texture area are rich, and the noise reduction intensity needs to be reduced.
  • the details of the leaves can be preserved while effectively removing the noise in the sky area.
  • the embodiment of the present disclosure proposes a noise reduction and enhancement method for dark-light scenes based on image segmentation and texture detection. Compared with the related technologies, it is difficult to balance the noise reduction and detail preservation.
  • the embodiment of the present disclosure is based on cleaner noise reduction. Preserve more image details.
  • the embodiment of the present disclosure can be used as a basic module nested in various types of terminals or platforms.
  • the embodiment of the present disclosure uses a combination of a lightweight deep learning network and a gradient calculation mask. Compared with using a deep learning network or a traditional method to calculate a texture mask, the effect is better, and the obtained mask is more refined and more accurate. , stronger anti-noise ability, suitable for extreme scenes.
  • the embodiments of the present disclosure may have the following application scenarios: when the user takes pictures indoors and outdoors in dark scenes, and finds that the noise is too large or the details are seriously smeared, or taking too long to take pictures, the method can effectively remove the noise and retain the details.
  • the method can be used to effectively and quickly process it into a high-quality video.
  • the image processing device when the user takes a photo, after the user presses the shutter button, the image processing device takes the captured image as the image to be processed, and then uses the image processing method in the embodiment of the present disclosure to perform noise reduction processing on the image to be processed .
  • the user when the user determines that it is necessary to denoise the locally stored image, the user can select the image, and the image processing device determines the image as the image to be processed, and the user can select the displayed denoising button, so that The image processing device uses the image processing method in the embodiments of the present disclosure to perform noise reduction processing on the image to be processed.
  • the embodiments of the present disclosure provide an image processing device, which includes each unit included in the device, and each module included in each unit, which can be implemented by a processor in the image processing device; of course, it can also Realized by specific logic circuit.
  • FIG. 8 is a schematic diagram of the composition and structure of an image processing device provided by an embodiment of the present disclosure. As shown in FIG. 8 , the image processing device 800 includes:
  • a determining unit 801 configured to determine a semantically segmented image of the image to be processed, and determine a texture image of the image to be processed; a fusion unit 802, configured to fuse the semantically segmented image and the texture image to obtain a fused image;
  • the determining unit 801 is further configured to determine a first noise reduction parameter corresponding to the image to be processed based on the fused image;
  • the noise reduction unit 803 is configured to perform a noise reduction on the image to be processed based on the first noise reduction parameter The image is denoised.
  • the determining unit 801 is further configured to determine first noise reduction parameters respectively corresponding to pixels in the image to be processed based on the fused image; wherein, pixels with different pixel values correspond to different first noise reduction parameters.
  • Noise parameter: the noise reduction unit 803 is further configured to perform noise reduction processing on the pixel based on the first noise reduction parameter corresponding to the pixel.
  • the determining unit 801 is further configured to determine a gradient image sequence corresponding to the image to be processed; the gradient image sequence includes at least two normalized gradient images of different scales; based on the gradient image sequence , to determine the texture image.
  • the determining unit 801 is further configured to determine a first image sequence; the first image sequence includes N first images, and the i-th first image in the N first images is for the ( i-1) first images are obtained by down-sampling, the first first image is the image to be processed; N is an integer greater than or equal to 2, and i is an integer greater than or equal to 2; The first image in the image sequence is subjected to image gradient processing to obtain the gradient image sequence.
  • the determining unit 801 is further configured to respectively perform noise reduction processing on the first images in the first image sequence using a second noise reduction parameter to obtain a second image sequence;
  • the second image in the second image sequence is subjected to image gradient processing to obtain the gradient image sequence.
  • the determination unit 801 is further configured to merge the first upsampled image obtained by upsampling the Nth gradient image with the (N-1)th gradient image to obtain a merged image; N is An integer greater than or equal to 2; when N is 2, the combined image is determined as the texture image; when N is greater than 2, the jth combined image obtained is up-sampled to obtain the ( j+1) upsampled image; the (j+1)th upsampled image is merged with the (N-1-j) gradient image to obtain the (j+1)th merged image; j is An integer greater than or equal to 1; determine the merged image obtained last time as the texture image.
  • the fusion unit 802 is further configured to determine a weight value corresponding to at least one region in the semantically segmented image; modify the pixel value of at least one region in the semantically segmented image to the at least one region A weight image corresponding to the weight value is obtained; and the weight image is fused with the texture image to obtain the fused image.
  • the fusion unit 802 is further configured to correspondingly subtract pixel values in the texture image from pixel values in the weight image to obtain a target image; The pixel value is modified to the first threshold, and the pixel value in the target image smaller than the second threshold is modified to the second threshold to obtain the fusion image; the first threshold is greater than the second threshold.
  • the determining unit 801 is further configured to acquire the pixel value of the first pixel in the fused image; according to the pixel value of the first pixel, the to-be Processing the second pixel in the image to set the first noise reduction parameter; wherein, the first noise reduction parameter includes standard deviation and window size, if the pixel value of the m first pixel is greater than the pixel value of the n first pixel , then the standard deviation of the mth second pixel is smaller than the standard deviation of the nth second pixel, and the window size of the mth second pixel is smaller than the window size of the nth second pixel, m and n Different, both m and n are integers greater than or equal to 1.
  • the at least two areas include a first area, a second area, and a third area; the determination unit 801 is further configured to acquire the pixel value of the first pixel in the fused image; according to the first The size of the pixel value of the pixel, the image to be processed is divided into the first area, the second area and the third area; wherein, the fusion corresponding to the second pixel in the first area respectively The pixel value of the first pixel in the image is the minimum value of the pixel values in the fused image; the pixel value of the first pixel in the fused image respectively corresponding to the second pixel in the third area is the fused The maximum value of the pixel value in the image; the second area is an area other than the first area and the third area in the image to be processed; the sub-noise reduction parameters corresponding to the first area, the The noise reduction strengths of the sub-noise reduction parameters corresponding to the second area and the sub-noise reduction parameters corresponding to the third area decrease sequentially.
  • the at least two areas include a fourth area and a fifth area; the determination unit 801 is further configured to obtain the pixel value of the first pixel in the fused image; according to the pixel value of the first pixel size, dividing the image to be processed into the fourth area and the fifth area; wherein, the pixel values of the first pixels in the fused image respectively corresponding to the second pixels in the fourth area are greater than The third threshold; the pixel value of the first pixel in the fused image respectively corresponding to the second pixel in the fifth area is less than or equal to the third threshold; the sub-noise reduction parameter corresponding to the fourth area The noise reduction strength is smaller than the noise reduction strength of the sub-noise reduction parameters corresponding to the fifth area.
  • the above image processing method is realized in the form of software function modules and sold or used as an independent product, it can also be stored in a computer-readable storage medium.
  • the essence of the technical solutions of the embodiments of the present disclosure or the part that contributes to the related technologies can be embodied in the form of software products, the computer software products are stored in a storage medium, and include several instructions to make An image processing device executes all or part of the methods described in various embodiments of the present disclosure.
  • FIG. 9 is a schematic diagram of a hardware entity of an image processing device provided by an embodiment of the present disclosure. As shown in FIG. A computer program running on the processor 901, when the processor 901 executes the program, implements the image processing method of any one of the above embodiments.
  • the memory 902 stores computer programs that can run on the processor, the memory 902 is configured to store instructions and applications executable by the processor 901, and can also cache the processor 901 and each module in the image processing device 900 to be processed or processed
  • the data for example, image data, audio data, voice communication data and video communication data
  • FLASH flash memory
  • RAM random access memory
  • the processor 901 executes the program, any one of the above image processing methods is realized.
  • the processor 901 generally controls the overall operation of the image processing device 900 .
  • An embodiment of the present disclosure provides a computer-readable storage medium, where one or more programs are stored in the computer-readable storage medium, and the one or more programs can be executed by one or more processors, so as to implement the above-mentioned any embodiment. image processing method.
  • a computer readable storage medium may be a volatile storage medium or a nonvolatile storage medium.
  • the embodiment of the present disclosure may also provide a chip, the chip includes a processor, and the processor can call and run a computer program from the memory, so as to implement the image processing method in the embodiment of the present disclosure.
  • Chips may also include memory.
  • the processor can call and run the computer program from the memory, so as to realize the image processing method in the embodiment of the present disclosure.
  • the memory may be an independent device independent of the processor, or may be integrated in the processor.
  • the chip may also include an input interface.
  • the processor can control the input interface to communicate with other devices or chips, specifically, can obtain information or data sent by other devices or chips.
  • the chip may also include an output interface.
  • the processor can control the output interface to communicate with other devices or chips, specifically, can output information or data to other devices or chips.
  • the chip can be applied to the image processing device in the embodiments of the present disclosure, and the chip can implement the corresponding processes implemented by the image processing device in the various methods of the embodiments of the present disclosure. For the sake of brevity, no more repeat.
  • chips mentioned in the embodiments of the present disclosure may also be referred to as system-on-chip, system-on-chip, system-on-a-chip, or system-on-chip.
  • An embodiment of the present disclosure also provides a computer program product, the computer program product includes a computer-readable storage medium, the computer-readable storage medium stores a computer program, and the computer program includes instructions executable by at least one processor , implementing the image processing method in the embodiment of the present disclosure when the instructions are executed by the at least one processor.
  • the embodiment of the present disclosure also provides a computer program.
  • the computer program enables a computer to execute the image processing method in the embodiment of the present disclosure.
  • the above-mentioned image processing device, chip or processor may include the integration of any one or more of the following: application specific integrated circuit (Application Specific Integrated Circuit, ASIC), digital signal processor (Digital Signal Processor, DSP), digital signal processing device ( Digital Signal Processing Device, DSPD), Programmable Logic Device (Programmable Logic Device, PLD), Field Programmable Gate Array (Field Programmable Gate Array, FPGA), Central Processing Unit (Central Processing Unit, CPU), Graphics Processor (Graphics Processing Unit (GPU), embedded neural-network processing units (NPU), controllers, microcontrollers, microprocessors, programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. Understandably, the electronic device that implements the above processor function may also be other, which is not specifically limited in this embodiment of the present disclosure.
  • the above-mentioned computer-readable storage medium/memory can be a read-only memory (Read Only Memory, ROM), a programmable read-only memory (Programmable Read-Only Memory, PROM), an erasable programmable read-only memory (Erasable Programmable Read-Only Memory, EPROM), Electrically Erasable Programmable Read-Only Memory (Electrically Erasable Programmable Read-Only Memory, EEPROM), Magnetic Random Access Memory (Ferromagnetic Random Access Memory, FRAM), Flash Memory (Flash Memory), magnetic surface Memory, CD, or CD-ROM (Compact Disc Read-Only Memory, CD-ROM) and other storage; it can also be a variety of terminals including one or any combination of the above storage, such as mobile phones, computers, tablet devices, personal digital Assistant etc.
  • references throughout the specification to "one embodiment” or “an embodiment” or “an embodiment of the present disclosure” or “the foregoing embodiments” or “some implementations” or “some embodiments” mean the same as implementing A specific feature, structure, or characteristic related to an example is included in at least one embodiment of the present disclosure.
  • appearances of "in one embodiment” or “in an embodiment” or “embodiments of the present disclosure” or “the foregoing embodiments” or “some implementations” or “some embodiments” throughout the specification do not necessarily mean Must refer to the same embodiment.
  • the particular features, structures or characteristics may be combined in any suitable manner in one or more embodiments.
  • sequence numbers of the above-mentioned processes do not mean the order of execution, and the execution order of the processes should be determined by their functions and internal logic, rather than by the embodiments of the present disclosure.
  • the implementation process constitutes any limitation.
  • the serial numbers of the above-mentioned embodiments of the present disclosure are for description only, and do not represent the advantages and disadvantages of the embodiments.
  • the image processing device executes any step in the embodiments of the present disclosure, and may be a processor of the image processing device executes the step.
  • the embodiments of the present disclosure do not limit the order in which the image processing apparatus executes the following steps.
  • the methods for processing data in different embodiments may be the same method or different methods. It should also be noted that any step in the embodiments of the present disclosure can be executed independently by the image processing apparatus, that is, when the image processing apparatus executes any step in the foregoing embodiments, it may not depend on the execution of other steps.
  • first and second are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance or implicitly indicating the quantity of indicated technical features.
  • a feature defined as first or second may explicitly or implicitly include one or more of said features.
  • the disclosed devices and methods may be implemented in other ways.
  • the device embodiments described above are only illustrative.
  • the division of the units is only a logical function division.
  • the coupling, or direct coupling, or communication connection between the components shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or units may be electrical, mechanical or other forms of.
  • the units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed to multiple network units; Part or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.
  • each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, or each unit may be used as a single unit, or two or more units may be integrated into one unit; the above-mentioned integration
  • the unit can be realized in the form of hardware or in the form of hardware plus software functional unit.
  • the above-mentioned integrated units of the present disclosure are implemented in the form of software function modules and sold or used as independent products, they can also be stored in a computer-readable storage medium.
  • the computer software products are stored in a storage medium and include several instructions to make A computer device (which may be a personal computer, a server, or a network device, etc.) executes all or part of the methods described in various embodiments of the present disclosure.
  • the aforementioned storage medium includes various media capable of storing program codes such as removable storage devices, ROMs, magnetic disks or optical disks.
  • the term "and" does not affect the order of the steps. For example, if the image processing device executes A and then B, the image processing device may first execute A and then B, or the image processing device first executes Execute B, then execute A, or the image processing device executes A while executing B.
  • the image processing method described in the embodiments of the present disclosure is applicable to the field of computer vision, and is particularly applicable to the video and/or image data to be processed, such as the video and/or image captured by the image processing device, and stored in the storage space of the image processing device.
  • the stored video and/or image, the image processing device receives the video and/or image sent by other equipment, etc., and performs effective noise reduction processing, so that the processed image retains image details while removing noise, greatly improving the image quality. quality.
  • this method can effectively remove the noise and preserve the details.
  • the method can be used to effectively and quickly process it into a high-quality video.
  • the image can be selected, and the image processing device determines the image as the image to be processed, and the user can select the displayed denoising button, and then apply this method to obtain a high-quality image. image.
  • This disclosure does not limit the industrial applicability.
  • the method described in this disclosure can be used. image processing method.

Abstract

本公开实施例公开了一种图像处理方法、装置、设备、存储介质及计算机程序产品，其中，该方法包括：确定待处理图像的语义分割图像，以及确定所述待处理图像的纹理图像；对所述语义分割图像和所述纹理图像进行融合，得到融合图像；基于所述融合图像，确定所述待处理图像对应的第一降噪参数；基于所述第一降噪参数，对所述待处理图像进行降噪处理。

Description

图像处理方法、装置、设备、存储介质及计算机程序产品
本申请要求在2021年10月21日提交中国专利局、申请号为202111226036.8、申请名称为“图像处理方法、装置、设备及计算机可读存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本公开实施例涉及但不限于计算机视觉领域，尤其涉及一种图像处理方法、装置、设备、存储介质及计算机程序产品。
背景技术
图像降噪一直是图像处理研究的热点之一。图像降噪可以改善具有噪声的图像，有利于减少图像由于噪声干扰而导致的图像质量下降。降噪可以有效地提高图像质量，增大信噪比，更好的体现原始图像所携带的信息。
如何在图像降噪的情况下减少细节丢失越来越受到关注。
发明内容
本公开实施例提供一种图像处理方法、装置、设备、存储介质及计算机程序产品。
第一方面，本公开实施例提供一种图像处理方法，包括：确定待处理图像的语义分割图像，以及确定所述待处理图像的纹理图像；对所述语义分割图像和所述纹理图像进行融合，得到融合图像；基于所述融合图像，确定所述待处理图像对应的第一降噪参数；基于所述第一降噪参数，对所述待处理图像进行降噪处理。
这样，由于融合图像是对语义分割图像和纹理图像进行融合得到的，融合图像表征的是待处理图像的纹理和待处理图像中不同的语义区域融合的结果，从而基于融合图像确定待处理图像对应的第一降噪参数，使得确定的第一降噪参数能够基于待处理图像的纹理和不同的语义区域的融合结果，进而通过第一降噪参数对待处理图像进行降噪，能够在对待处理图像降噪的同时，不会丢失待处理图像中的细节，提升了待处理图像的降噪能力。
在一些实施例中，所述基于所述融合图像，确定所述待处理图像对应的第一降噪参数，包括：基于所述融合图像，确定所述待处理图像中的像素分别对应的第一降噪参数；其中，不同像素值的像素对应不同的第一降噪参数；所述基于所述第一降噪参数，对所述待处理图像进行降噪处理，包括：基于所述像素对应的第一降噪参数，对所述像素进行降噪处理。
这样，不同像素值的像素对应不同的第一降噪参数，基于像素对应的第一降噪参数，对像素进行降噪处理，从而能够根据待处理图像中像素的纹理特征和语义特征的融合，对待处理图像实现像素级别的降噪，进而能够精确地对待处理图像的像素进行降噪，提升了待处理图像的降噪能力。
在一些实施例中，所述第一降噪参数包括至少两个子降噪参数，所述至少两个子降噪参数对应的降噪力度不同；所述基于所述融合图像，确定所述待处理图像对应的第一降噪参数，包括：基于所述融合图像，将所述待处理图像分为至少两个区域；根据所述至少两个子降噪参数，分别确定所述区域对应的子降噪参数；所述基于所述第一降噪参数，对所述待处理图像进行降噪处理，包括：基于所述区域对应的子降噪参数，分别对所述区域进行降噪处理。
这样，基于融合图像，将待处理图像分为至少两个区域，分别确定区域对应的子降噪参数，基于区域对应的子降噪参数，对区域进行降噪处理，从而同一个区域使用同一个降噪参数，不同的区域使用不同的降噪参数，不仅能够提升待处理图像的降噪能力，还能够快速地对待处理图像进行降噪。
在一些实施例中，所述确定所述待处理图像的纹理图像，包括：确定与所述待处理图像对应的梯度图像序列；所述梯度图像序列包括至少两个不同尺度的归一化梯度图像；基于所述梯度图像序列，确定 所述纹理图像。
这样，通过确定与待处理图像对应的梯度图像序列，基于梯度图像序列，确定纹理图像，从而纹理图像是通过各个不同尺度的梯度图像结合得到的，使得确定纹理图像能够准确地反映待处理图像的纹理。
在一些实施例中，所述确定与所述待处理图像对应的梯度图像序列，包括：确定第一图像序列；所述第一图像序列包括N个第一图像，所述N个第一图像中第i个第一图像是对第(i-1)个第一图像进行下采样得到，第一个第一图像为所述待处理图像；N为大于或等于2的整数，i为大于或等于2的整数；分别对所述第一图像序列中的所述第一图像进行图像梯度处理，得到所述梯度图像序列。
这样，通过先对待处理图像进行不同程度的下采样，得到第一图像序列，再分别对第一图像序列中第一图像进行图像梯度处理，得到梯度图像序列，从而得到的梯度图像序列能够反映第一图像序列的梯度信息，进而基于梯度图像序列，能够准确地确定纹理图像。
在一些实施例中，所述分别对所述第一图像序列中的所述第一图像进行图像梯度处理，得到所述梯度图像序列，包括：分别对所述第一图像序列中的所述第一图像均采用第二降噪参数进行降噪处理，得到第二图像序列；分别对所述第二图像序列中的第二图像进行图像梯度处理，得到所述梯度图像序列。
这样，由于先分别对第一图像序列中第一图像进行降噪处理，然后再对降噪处理得到的第二图像序列进行图像梯度处理，得到梯度图像序列，从而能够减少图像噪声对梯度计算的影响，使得到的梯度图像序列准确。
在一些实施例中，所述基于所述梯度图像序列，确定所述纹理图像，包括：对第N个梯度图像进行上采样得到的第一个上采样图像，与第(N-1)个梯度图像进行合并，得到合并图像；N为大于或等于2的整数；在N为2的情况下，将所述合并图像确定为所述纹理图像；在N大于2的情况下，对得到的第j个合并图像进行上采样，得到第(j+1)个上采样图像；将所述第(j+1)个上采样图像与第(N-1-j)个梯度图像进行合并，得到第(j+1)个合并图像；j为大于或等于1的整数；将最后一次得到的合并图像确定为所述纹理图像。
这样，对得到的第j个合并图像进行上采样，得到第(j+1)个上采样图像，将第(j+1)个上采样图像与第(N-1-j)个梯度图像进行合并，得到第(j+1)个合并图像，将最后一次得到的合并图像确定为纹理图像，纹理图像能够通过对梯度图像序列中每个梯度图像合并得到，从而纹理图像能够准确地反映出待处理图像的纹理。
在一些实施例中，所述对所述语义分割图像和所述纹理图像进行融合，得到融合图像，包括：确定所述语义分割图像中至少一个区域对应的权重值；将所述语义分割图像中的至少一个区域的像素值，修改为所述至少一个区域对应的权重值，得到权重图像；将所述权重图像和所述纹理图像进行融合，得到所述融合图像。
这样，可将语义分割图像中的至少一个区域的像素值，修改为至少一个区域对应的权重值，得到权重图像；将权重图像和纹理图像进行融合，得到融合图像，从而为待处理图像中不同的语义区域设置不同的权重值，进而可以基于权重值来确定降噪力度，提升了待处理图像的降噪能力。
在一些实施例中，所述将所述权重图像和所述纹理图像进行融合，得到所述融合图像，包括：将所述纹理图像中的像素值与所述权重图像中的像素值对应相减，得到目标图像；将所述目标图像中大于第一阈值的像素值，修改为所述第一阈值，且将所述目标图像中小于第二阈值的像素值，修改为所述第二阈值，得到所述融合图像；所述第一阈值大于所述第二阈值。
这样，可基于将纹理图像中的像素值与权重图像中的像素值对应相减得到的目标图像，确定融合图像，从而提供了一种权重图像和纹理图像融合的实现方式，使得融合图像能够准确的表达待处理图像携带的信息；并且，通过将目标图像中大于第一阈值的像素值，修改为第一阈值，且将目标图像中小于第二阈值的像素值，修改为第二阈值，从而基于融合图像能够容易地区分出待处理图像的纹理区域和平坦区域，从而针对待处理图像中纹理区域和平坦区域进行差异化的降噪，能够在对待处理图像降噪的同时，不会丢失待处理图像中的细节，提升了待处理图像的降噪能力。
在一些实施例中，所述基于所述融合图像，确定所述待处理图像中的像素分别对应的第一降噪参数，包括：获取所述融合图像中第一像素的像素值；根据所述第一像素的像素值大小，为与所述第一像素分别对应的所述待处理图像中的第二像素设置第一降噪参数；其中，所述第一降噪参数包括标准差和窗口大小，若第m个第一像素的像素值大于第n个第一像素的像素值，则第m个第二像素的标准差小于第n个第二像素的标准差，且所述第m个第二像素的窗口大小小于所述第n个第二像素的窗口大小，m和n不同，m和n均为大于或等于1的整数。
这样，根据融合图像中第一像素的像素值大小，为与第一像素分别对应的待处理图像中的第二像素设置第一降噪参数，从而为不同像素值大小的第一像素所对应的第二像素设置不同的第一降噪参数，因此能够精确地对待处理图像的各个像素进行降噪，提升了待处理图像的降噪能力。
在一些实施例中，所述至少两个区域包括第一区域、第二区域以及第三区域；所述基于所述融合图像，将所述待处理图像分为至少两个区域，包括：获取所述融合图像中第一像素的像素值；根据所述第一像素的像素值大小，将所述待处理图像分为所述第一区域、所述第二区域以及所述第三区域；其中，与所述第一区域中第二像素分别对应的所述融合图像中第一像素的像素值，为所述融合图像中像素值的最小值；与所述第三区域中第二像素分别对应的所述融合图像中第一像素的像素值，为所述融合图像中像素值的最大值；所述第二区域为所述待处理图像中除所述第一区域和所述第三区域之外的区域；所述第一区域对应的子降噪参数、所述第二区域对应的子降噪参数以及所述第三区域对应的子降噪参数的降噪力度依次减小。
这样，融合图像中像素值为最小值的像素，表征对应的待处理图像中第一区域的像素为平坦区中的像素，从而采用较大的降噪力度对该像素进行降噪，从而能够对平坦区中的像素进行有效的降噪；融合图像中像素值为最大值的像素，表征对应的待处理图像中第三区域的像素为纹理区中的像素，从而采用较小的降噪力度对该像素进行降噪，进而能够对纹理区中的像素进行降噪时，较少地丢失纹理信息；并且，对除第一区域和第三区域之外的第二区域的像素，采用中等的降噪力度进行降噪，能够使得到的降噪后的图像平滑。
在一些实施例中，所述至少两个区域包括第四区域和第五区域；所述基于所述融合图像，将所述待处理图像分为至少两个区域，包括：获取所述融合图像中第一像素的像素值；根据所述第一像素的像素值大小，将所述待处理图像分为所述第四区域和所述第五区域；其中，与所述第四区域中第二像素分别对应的所述融合图像中第一像素的像素值，大于第三阈值；与所述第五区域中第二像素分别对应的所述融合图像中第一像素的像素值，小于或等于所述第三阈值；其中，所述第四区域对应的子降噪参数的降噪力度，小于所述第五区域对应的子降噪参数的降噪力度。
这样，在融合图像中像素值大于第三阈值的情况下，确定对应的待处理图像中第四区域的像素的纹理较多，从而采用较小的降噪力度对第四区域的像素进行降噪，从而在对纹理较多的像素进行降噪时，较少的丢失纹理信息；在融合图像中像素值小于或等于第三阈值的情况下，确定对应的待处理图像中第五区域的像素的纹理较少，从而采用较大的降噪力度对第五区域的像素进行降噪，从而能够对纹理较少中的像素进行有效的降噪，提升了待处理图像的降噪能力。
第二方面，本公开实施例提供一种图像处理装置，包括：确定单元，用于确定待处理图像的语义分割图像，以及确定所述待处理图像的纹理图像；融合单元，用于对所述语义分割图像和所述纹理图像进行融合，得到融合图像；所述确定单元，还用于基于所述融合图像，确定所述待处理图像对应的第一降噪参数；降噪单元，用于基于所述第一降噪参数，对所述待处理图像进行降噪处理。
第三方面，本公开实施例提供一种图像处理设备，包括：存储器和处理器，所述存储器存储有可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现上述的图像处理方法。
第四方面，本公开实施例提供一种计算机可读存储介质，所述计算机可读存储介质存储有一个或者多个程序，所述一个或者多个程序可被一个或者多个处理器执行，以实现上述的图像处理方法。
第五方法，本公开实施例还提供了一种计算机程序产品，包括计算机可读代码，或者承载有计算机可读代码的计算机可读存储介质，当所述计算机可读代码在电子设备的处理器中运行时，所述电子设备中的处理器执行用于实现上述的图像处理方法。
附图说明
此处所说明的附图用来提供对本公开的进一步理解，构成本公开的一部分，本公开的示意性实施例及其说明用于解释本公开，并不构成对本公开的不当限定。
图1为本公开实施例提供的一种图像处理方法的实现流程示意图；
图2为本公开实施例提供的图像处理方法的实现方式示意图；
图3为本公开实施例提供的另一种图像处理方法的实现流程示意图；
图4为本公开实施例提供的又一种图像处理方法的实现流程示意图；
图5为本公开实施例提供的再一种图像处理方法的实现流程示意图；
图6为本公开实施例提供的一种基于梯度图像序列确定纹理图像的实现方式示意图；
图7为本公开另一实施例提供的一种图像处理方法的实现流程示意图；
图8为本公开实施例提供的一种图像处理装置的组成结构示意图；
图9为本公开实施例提供的一种图像处理设备的硬件实体示意图。
具体实施方式
下面将通过实施例并结合附图具体地对本公开的技术方案进行详细说明。下面这几个具体的实施例可以相互结合，对于相同或相似的概念或过程可能在某些实施例中不再赘述。
需要说明的是：在本公开实例中，第一、第二等是用于区别类似的对象，而不必用于描述特定的顺序或先后次序。
另外，本公开实施例所记载的技术方案之间，在不冲突的情况下，可以任意组合。在本公开的描述中，多个的含义是两个或两个以上，除非另有明确具体的限定。
暗光场景噪声在图像处理中比较常见，大部分相机由于硬件原因在暗光下光子感应较差，而高感光度通常会导致生成的图像带有明显的噪声。
相关技术中通过深度学习或三维块匹配算法(Block Matching 3D，BM3D)传统方法可以有效的去除部分暗光下噪声。然而，基于深度学习的方法，一般是针对噪声建立模型，模拟暗光场景下的噪声分布，生成数据对，降噪的能力取决噪声模型与真实噪声的差异，但由于模拟噪声与实际噪声往往存在差异，降噪能力时好时坏，复用性低，鲁棒性差。BM3D传统方法通过空域或频域特性进行全局或特定频段的降噪，虽然能够减少图像中的噪声，但存在大量丢失图像细节的情况。
本申请实施例中提到的任一图像处理装置可以是处理器或者芯片，处理器或者芯片可以应用于图像处理设备中。或者，本申请实施例中提到的任一图像处理装置可以是图像处理设备。在一些实施例中，图像处理设备可以包括图像处理组件，例如，图像处理组件可以包括摄像头组件。在另一些实施例中，图像处理设备可以包括以下至少之一或者至少两者的组合：相机、服务器、手机(Mobile Phone)、平板电脑(Pad)、带无线收发功能的电脑、掌上电脑、台式计算机、个人数字助理、便捷式媒体播放器、智能音箱、导航装置、智能手表、智能眼镜、智能项链等可穿戴设备、计步器、数字TV、虚拟现实(Virtual Reality，VR)终端设备、增强现实(Augmented Reality，AR)终端设备、工业控制(Industrial Control)中的无线终端、无人驾驶(Self Driving)中的无线终端、远程手术(Remote Medical Surgery)中的无线终端、智能电网(Smart Grid)中的无线终端、运输安全(Transportation Safety)中的无线终端、智慧城市(Smart City)中的无线终端、智慧家庭(Smart Home)中的无线终端、车联网系统中的车、车载设备、车载模块等等。
图1为本公开实施例提供的一种图像处理方法的实现流程示意图，如图1所示，该方法应用于图像处理装置，该方法包括：
S101、确定待处理图像的语义分割图像，以及确定所述待处理图像的纹理图像。
在一些实施例中，待处理图像可以是原始图像。例如，原始图像可以通过图像拍摄得到的图像。又例如，原始图像可以是视频中的图像帧。再例如，原始图像可以是从本地中读取的图像、下载的图像或者从其它设备(例如，硬盘、U盘或其它终端等)读取到的图像。
在另一些实施例中，待处理图像可以是通过对原始图像进行以下至少之一处理得到的图像：缩放处理、裁剪处理、去噪处理、添加噪声处理、灰度处理、旋转处理、归一化处理。例如，可以对原始图像进行缩放处理后，再进行旋转处理，得到待处理图像。
确定待处理图像的语义分割图像，可以包括：对待处理图像进行语义分割，得到语义分割图像。确定所述待处理图像的纹理图像，可以包括：对待处理图像进行纹理检测，得到纹理图像。
对待处理图像进行语义分割，得到语义分割图像，可以包括：将待处理图像输入至语义分割网络(或者称语义分割模型)中，通过语义分割网络对待处理图像进行语义分割，得到语义分割图像。其中，语义分割网络可以是通过对多个带标签的第一训练图像进行训练得到的。
语义分割网络可以包括以下之一：全卷积网络(Fully Convolution Networks，FCN)、SegNet、U-Net、DeepLab v1、DeepLab v2、DeepLab v3、DenseNet、E-Net、Link-Net、掩膜区域卷积神经网络(Mask R-CNN)、金字塔场景解析网络(Pyramid Scene Parseing Network，PSPNet)、RefineNet、门控反馈优化网络(Gated Feedback Refinement Network，G-FRNet)，以及这些网络的演进网络等。
对待处理图像进行纹理检测，得到纹理图像，可以包括：将待处理图像输入至纹理检测网络(或者称纹理检测模型)中，通过纹理检测网络对待处理图像进行纹理检测，得到纹理图像。纹理检测网络可以包括：深度纹理编码网络(Deep Texture Encoding Network，Deep-TEN)等。
在一些实施例中，纹理检测网络还可以称为边缘分割网络，边缘分割网络可以包括以下之一：基于更丰富特征的边缘检测(Richer Convolutional Features for Edge Detection，RCF)网络、整体嵌套的边缘检测(holistically-nested edge detection，HED)网络、Canny边缘检测网络，以及这些网络的演进网络等。
语义分割图像和纹理图像的像素尺寸可以均与待处理图像的像素尺寸相同，例如，在待处理图像的像素尺寸为800×600或者800×600×3的情况下，其中，800为待处理图像在宽度方向上的像素尺寸， 600为待处理图像在高度方向上的像素尺寸，3为待处理图像的通道数，通道包括红绿蓝(RGB)三通道，语义分割图像和纹理图像的像素尺寸均为800×600。
语义分割网络可以从大量标注的训练集(包括多个带标签的第一训练图像)中学习的经验知识以及每个位置的局部特征，判断各个位置属于的类别。对待处理图像进行语义分割，是为了对待处理图像中的每个像素进行语义分类，确定待处理图像中每个像素属于某一种物体或者背景。在待处理图像中某一个像素属于背景的情况下，可以将语义分割图像中对应像素的标识值确定为第一值，在待处理图像中某一个像素属于N个类别中目标类物体的情况下，可以将语义分割图像中对应像素的标识值确定为与目标类物体对应的值，N为大于或等于1的整数，与目标类物体对应的值的取值也有N种，且与第一值不同。这样，语义分割图像中每一像素的标识值可以是N+1种数值，N为物体类别的总数，从而可以基于语义分割图像中不同种类的值的位置，确定待处理图像中的背景部分和每一类物体的位置。在一些实施方式中，语义分割图像可以称为语义掩膜(Segm mask)。
对待处理图像进行纹理检测的，是为了确定待处理图像中的每个像素是否为纹理像素或边缘像素，在纹理图像中的某一个像素值越大的情况下，表明该像素值对应的像素为纹理像素的概率越高；在纹理图像中的某一个像素值越小的情况下，表明该像素值对应的像素为纹理像素的概率越低。
S102、对所述语义分割图像和所述纹理图像进行融合，得到融合图像。
在一些实施例中，S102可以通过以下方式实现：获取语义分割图像中每个像素对应的像素值，获取纹理图像中每个像素对应的像素值，将语义分割图像中每个像素对应的像素值，和纹理图像中每个像素对应的像素值对应进行数学计算，得到融合图像。其中，语义分割图像中每个像素对应的像素值可以是语义分割图像中每个像素的真实像素值，或者，语义分割图像中每个像素对应的像素值可以是与语义分割图像中每个像素的真实像素值映射的像素值。其中，纹理图像中每个像素对应的像素值可以是纹理图像中每个像素的真实像素值，或者，纹理图像中每个像素对应的像素值可以是与纹理图像中每个像素的真实像素映射的像素值。示例性地，映射的像素值可以与真实像素值具有映射关系，例如，映射的像素值可以与真实像素值具有一对一的映射关系，或者，一对多的映射关系。示例性地，映射的像素值可以是通过真实像素值计算得到的，或者，映射的像素值可以是通过真实像素值和映射关系得到，映射关系包括真实像素值和映射的像素值之间的对应关系。
进行数学计算包括但不限于以下至少之一：相乘、相加、相减、相除、指数运算、对数运算等。
以数学计算为相减运算举例，将语义分割图像中每个像素对应的像素值，和纹理图像中每个像素对应的像素值对应进行数学计算，可以包括：将语义分割图像中每个像素对应的像素值，和纹理图像中每个像素对应的像素值对应相乘，或者，将纹理图像中每个像素对应的像素值，与语义分割图像中每个像素对应的像素值对应相减。例如，将纹理图像中第a行第b列的像素对应的像素值，与语义分割图像中第a行第b列的像素对应的像素值对应相减。a、b均为大于或等于1的整数，a、b可以相同或不同。
通过将语义分割图像和纹理图像进行融合，能够充分利用各自优点，获取更加精细的掩码。例如语义分割的天空区域的分割，纹理检测在天空区域可能无法检测干净。
S103、基于所述融合图像，确定所述待处理图像对应的第一降噪参数。
在一些实施例中，融合图像中像素的像素值可以处于预设范围。例如，预设范围可以是[0,1]，这样，融合图像中像素的像素值可以是[0,1]中的任一个值。例如，融合图像中某个像素的像素值可以为0、0.128、0.75或者1等。在另一些实施例中，融合图像中像素点的像素值可以是数据集合中的数据，数据集合可以是预先设定的包括至少两个数值的集合。例如，数据集合可以包括{0,1}、{0,0.5,1}或者{0,0.2,0.4,0.6,0.8,1}等。
第一降噪参数和/或下述的第二降噪参数可以是滤波器的降噪参数，滤波器可以是高斯滤波器或其它滤波器，高斯滤波器是利用高斯核的一个2维的卷积算子，其它滤波器可以包括以下之一：中值滤波器、均值滤波器、双边滤波器、最大最小值滤波、引导滤波器、索贝尔(Sobel)滤波器、Prewitt滤波器、拉普拉斯(Laplacian)滤波器等等，本公开实施例对滤波器和滤波器的降噪参数不作限制。第一降噪参数和/或下述的第二降噪参数可以包括：标准差和/或窗口大小，第一降噪参数和/或下述的第二降噪参数不同可以包括标准差不同和/或窗口大小不同。
本公开实施例中滤波器可以包括高斯核，基于标准差可以确定高斯核中的元素值，基于窗口大小可以确定高斯核的尺寸。
S104、基于所述第一降噪参数，对所述待处理图像进行降噪处理。
在本公开实施例中，由于融合图像是对语义分割图像和纹理图像进行融合得到的，融合图像表征的是待处理图像的纹理和待处理图像中不同的语义区域融合的结果，从而基于融合图像确定待处理图像对应的第一降噪参数，使得确定的第一降噪参数能够基于待处理图像的纹理和不同的语义区域的融合结果，进而通过第一降噪参数对待处理图像进行降噪，能够在对待处理图像降噪的同时，不会丢失待处理图像 中的细节，提升了待处理图像的降噪能力。
图2为本公开实施例提供的图像处理方法的实现方式示意图，可以应用于图像处理装置，如图2所示，图像处理装置可以先得到待处理图像21，然后将待处理图像21分别输入至语义分割网络和纹理检测网络，通过语义分割网络的得到语义分割图像22，通过纹理检测网络得到纹理图像23。
在得到语义分割图像22和纹理图像23之后，可以对语义分割图像22和纹理图像23进行融合(或者称掩码融合)，得到融合图像24。接着将融合图像24和待处理图像21输入至降噪模块，降噪模块可以基于融合图像，确定待处理图像中对应的第一降噪参数；基于第一降噪参数，对待处理图像中进行降噪处理，在降噪模块对待处理图像降噪完成之后，输出处理后的图像25。
在一些实施例中，确定待处理图像的语义分割图像，可以包括：对待处理图像采用第一下采样因子进行下采样，得到下采样图像；对下采样图像进行语义分割，得到目标分割图像；对目标分割图像进行上采样，得到与待处理图像的尺寸相同的语义分割图像。
第一下采样因子可以为大于或等于2的整数。例如，第一下采样因子可以为2、4、8或10等。
通过对待处理图像采用第一下采样因子进行下采样得到下采样图像，可以减小用于语义分割的图像的尺寸。例如，在待处理图像的尺寸为M×N，第一下采样因子为10的情况下，下采样图像可以为M/10×N/10。
待处理图像可以采用第一下采样因子进行多次下采样，得到下采样图像。例如，在待处理图像的尺寸为M×N，第一下采样因子为2的情况下，可以先下采样得到M/2×N/2的图像，再在M/2×N/2的图像的基础上，再次下采样得到M/4×N/4的图像，在M/4×N/4的图像的基础上，再次下采样得到M/8×N/8的图像，将该M/8×N/8的图像确定为下采样图像。
可以将下采样图像输入至语义分割网络中，通过语义分割网络输出目标分割图像。目标分割图像中每个像素的像素值表征该像素所属的对象，从而基于目标分割图像可以确定下采样图像中各个不同的对象(包括背景)所对应的像素区域。
可以采用第一上采样因子对目标分割图像进行上采样，得到语义分割图像。第一上采样因子可以与上述的第一下采样因子相同。
在一些实施例中，可以对目标分割图像采用第一上采样因子进行一次上采样，得到语义分割图像。例如，在目标分割图像的尺寸为M/10×N/10，第一上采样因子为10的情况下，语义分割图像的尺寸可以为M×N。
在另一些实施例中，可以对目标分割图像采用第一上采样因子进行多次上采样，得到语义分割图像。例如，在目标分割图像的尺寸为M/8×N/8，第一上采样因子为2的情况下，可以先上采样得到M/4×N/4的图像，在M/4×N/4的图像的基础上，再次上采样得到M/2×N/2的图像，再在M/2×N/2的图像的基础上，再次上采样得到M×N的图像，上采样得到的M×N的图像即为语义分割图像。
在本公开实施例中，通过对待处理图像进行下采样得到的下采样图像进行语义分割，然后将语义分割得到的目标分割图像进行上采样得到语义分割图像，从而能够减少得到语义分割图像的计算量，降低了得到语义分割图像的耗时。
图3为本公开实施例提供的另一种图像处理方法的实现流程示意图，如图3所示，该方法应用于图像处理装置，该方法包括：
S301、确定待处理图像的语义分割图像，以及确定所述待处理图像的纹理图像。
S302、对所述语义分割图像和所述纹理图像进行融合，得到融合图像。
S303、基于所述融合图像，确定所述待处理图像中的像素分别对应的第一降噪参数；其中，不同像素值的像素对应不同的第一降噪参数。
在一些实施例中，S303可以通过以下方式实现：获取所述融合图像中第一像素的像素值；根据所述第一像素的像素值大小，为与所述第一像素分别对应的所述待处理图像中的第二像素设置第一降噪参数。
示例性地，若第m个第一像素的像素值大于第n个第一像素的像素值，则第m个第二像素的降噪力度小于第n个第二像素的降噪力度。若第m个第一像素的像素值等于第n个第一像素的像素值，则第m个第二像素的降噪力度等于第n个第二像素的降噪力度。若第m个第一像素的像素值小于第n个第一像素的像素值，则第m个第二像素的降噪力度大于第n个第二像素的降噪力度。
示例性地，所述第一降噪参数包括标准差和窗口大小，若第m个第一像素的像素值大于第n个第一像素的像素值，则第m个第二像素的标准差小于第n个第二像素的标准差，且所述第m个第二像素的窗口大小小于所述第n个第二像素的窗口大小，m和n不同，m和n均为大于或等于1的整数。若第m个第一像素的像素值等于第n个第一像素的像素值，则第m个第二像素的标准差等于第n个第二像素的标准差，且所述第m个第二像素的窗口大小等于所述第n个第二像素的窗口大小。若第m个第一像素的像素值小于第n个第一像素的像素值，则第m个第二像素的标准差大于第n个第二像素的标准差，且所述 第m个第二像素的窗口大小大于所述第n个第二像素的窗口大小。
第m个第一像素、第n个第一像素可以是融合图像中的任两个第一像素。第m个第二像素为待处理图像中与第m个第一像素对应的像素，第n个第二像素为待处理图像中与第n个第一像素对应的像素。
通过这种方式，根据融合图像中第一像素的像素值大小，为与第一像素分别对应的待处理图像中的第二像素设置第一降噪参数，从而为不同像素值大小的第一像素所对应的第二像素设置不同的第一降噪参数，因此能够精确地对待处理图像的各个像素进行降噪，提升了待处理图像的降噪能力。
S304、基于所述像素对应的第一降噪参数，对所述像素进行降噪处理。
可以基于所述融合图像中第一像素的像素值，分别确定所述待处理图像中的第二像素对应的第一降噪参数；待处理图像中的第二像素可以与所述融合图像中的第一像素一一对应。
在一些实施例中，融合图像中像素值不同的第一像素映射的待处理图像中的第二像素，对应的第一降噪参数不同。例如，融合图像中像素值为1和0.8的两个第一像素映射的待处理图像中的第二像素，对应的第一降噪参数不同。
在本公开实施例中，不同像素值的像素对应不同的第一降噪参数，基于像素对应的第一降噪参数，对像素进行降噪处理，从而能够根据待处理图像中像素的纹理特征和语义特征的融合，对待处理图像实现像素级别的降噪，进而能够精确地对待处理图像的像素进行降噪，提升了待处理图像的降噪能力。
图4为本公开实施例提供的又一种图像处理方法的实现流程示意图，如图4所示，该方法应用于图像处理装置，在本公开实施例中，所述第一降噪参数包括至少两个子降噪参数，所述至少两个子降噪参数对应的降噪力度不同，该方法包括：
S401、确定待处理图像的语义分割图像，以及确定所述待处理图像的纹理图像。
S402、对所述语义分割图像和所述纹理图像进行融合，得到融合图像。
S403、基于所述融合图像，将所述待处理图像分为至少两个区域。
可以根据融合图像中第一像素的像素值大小，将所述待处理图像分为至少两个区域。本公开实施例不限定至少两个区域的数量。至少两个区域的数量可以为两个、三个、五个或十个等等。例如，融合图像中像素的像素值在[0,1]范围内，可以预定至少两个像素值范围，至少两个像素值范围连续且不相交，且至少两个像素值范围的并集为[0,1]，根据融合图像中第一像素的像素值为至少两个像素值范围中某一个像素值范围，将待处理图像分为至少两个区域。其中，至少两个像素值范围和至少两个区域一一对应。
在一些实施例中，所述至少两个区域包括第一区域、第二区域以及第三区域；所述基于所述融合图像，将所述待处理图像分为至少两个区域，包括：获取所述融合图像中第一像素的像素值；根据所述第一像素的像素值大小，将所述待处理图像分为所述第一区域、所述第二区域以及所述第三区域；其中，与所述第一区域中第二像素分别对应的所述融合图像中第一像素的像素值，为所述融合图像中像素值的最小值；与所述第三区域中第二像素分别对应的所述融合图像中第一像素的像素值，为所述融合图像中像素值的最大值；所述第二区域为所述待处理图像中除所述第一区域和所述第三区域之外的区域；所述第一区域对应的子降噪参数、所述第二区域对应的子降噪参数以及所述第三区域对应的子降噪参数的降噪力度依次减小。
在融合图像中像素的像素值在[0,1]范围内的情况下，像素值中的最小值为0，像素值中的最大值为1。像素值为0的第一像素对应的待处理图像中的第二像素为平坦区的像素，需要较高的降噪力度，像素值为1的第一像素对应的待处理图像中的第二像素为纹理区的像素，需要较低的降噪力度，像素值大于0且小于1的第一像素对应的待处理图像中的第二像素为纹理区的像素，需要中间的降噪力度。
子降噪参数的降噪力度可以用标准差和窗口大小来表征。在子降噪参数的降噪力度越大的情况下，标准差越大，窗口大小越大；在子降噪参数的降噪力度越小的情况下，标准差越小，窗口大小越小。
所述第一区域、所述第二区域、所述第三区域对应的子降噪参数的降噪力度依次减小的一种实施方式可以为：第一区域的对应的子降噪参数包括的标准差，大于第二区域的对应的子降噪参数包括的标准差，和/或，第一区域的对应的子降噪参数包括的窗口大小，大于第二区域的对应的子降噪参数包括的窗口大小；以及，第二区域的对应的子降噪参数包括的标准差，大于第三区域的对应的子降噪参数包括的标准差，和/或，第二区域的对应的子降噪参数包括的窗口大小，大于第三区域的对应的子降噪参数包括的窗口大小。
通过这种方式，融合图像中像素值为最小值的像素，表征对应的待处理图像中第一区域的像素为平坦区中的像素，从而采用较大的降噪力度对该像素进行降噪，从而能够对平坦区中的像素进行有效的降噪；融合图像中像素值为最大值的像素，表征对应的待处理图像中第三区域的像素为纹理区中的像素，从而采用较小的降噪力度对该像素进行降噪，进而能够对纹理区中的像素进行降噪时，较少地丢失纹理信息；并且，对除第一区域和第三区域之外的第二区域的像素，采用中等的降噪力度进行降噪，能够使得到的降噪后的图像平滑。
在一些实施例中，所述至少两个区域包括第四区域和第五区域；所述基于所述融合图像，将所述待处理图像分为至少两个区域，包括：获取所述融合图像中第一像素的像素值；根据所述第一像素的像素值大小，将所述待处理图像分为所述第四区域和所述第五区域；其中，与所述第四区域中第二像素分别对应的所述融合图像中第一像素的像素值，大于第三阈值；与所述第五区域中第二像素分别对应的所述融合图像中第一像素的像素值，小于或等于所述第三阈值；所述第四区域对应的子降噪参数的降噪力度，小于所述第五区域对应的子降噪参数的降噪力度。
通过这种方式，在融合图像中像素值大于第三阈值的情况下，确定对应的待处理图像中第四区域的像素的纹理较多，从而采用较小的降噪力度对第四区域的像素进行降噪，从而在对纹理较多的像素进行降噪时，较少的丢失纹理信息；在融合图像中像素值小于或等于第三阈值的情况下，确定对应的待处理图像中第五区域的像素的纹理较少，从而采用较大的降噪力度对第五区域的像素进行降噪，从而能够对纹理较少中的像素进行有效的降噪，提升了待处理图像的降噪能力。
S404、根据所述至少两个子降噪参数，分别确定所述区域对应的子降噪参数。
S405、基于所述区域对应的子降噪参数，分别对所述区域进行降噪处理。
在本公开实施例中，基于融合图像，将待处理图像分为至少两个区域，分别确定区域对应的子降噪参数，基于区域对应的子降噪参数，对区域进行降噪处理，从而同一个区域使用同一个降噪参数，不同的区域使用不同的降噪参数，不仅能够提升待处理图像的降噪能力，还能够快速地对待处理图像进行降噪。
图5为本公开实施例提供的再一种图像处理方法的实现流程示意图，如图5所示，该方法应用于图像处理装置，该方法包括：
S501、确定与所述待处理图像对应的梯度图像序列；所述梯度图像序列包括至少两个不同尺度的归一化梯度图像。
在一些实施例中，S501可以通过以下方式实现：对待处理图像采用多尺度Canny检测算法，确定待处理图像对应的梯度图像序列。
在另一些实施例中，S501可以通过以下方式实现：确定第一图像序列；所述第一图像序列包括N个第一图像，所述N个第一图像中第i个第一图像是对第(i-1)个第一图像进行下采样得到，第一个第一图像为所述待处理图像；N为大于或等于2的整数，i为大于或等于2的整数；分别对所述第一图像序列中的所述第一图像进行图像梯度处理，得到所述梯度图像序列。
通过这种方式，通过先对待处理图像进行不同程度的下采样，得到第一图像序列，再分别对第一图像序列中第一图像进行图像梯度处理，得到梯度图像序列，从而得到的梯度图像序列能够反映第一图像序列的梯度信息，进而基于梯度图像序列，能够准确地确定纹理图像。
N个第一图像中第i个第一图像可以是对第(i-1)个第一图像，采用第二下采样因子进行下采样得到。这样，N个第一图像中第一个第一图像为所述待处理图像，第二个第一图像是对第一个第一图像采用第二下采样因子进行下采样得到，第三个第一图像是对第二个第一图像采用第二下采样因子进行下采样得到，以此类推，直到得到N个第一图像。
第二下采样因子可以为大于或等于2的整数。例如，第二下采样因子可以为2、4、8或10等。第一图像序列中包括的第一图像的数量可以基于实际需求(例如图像处理装置的计算能力)确定。在图像处理装置的计算能力较高的情况下，可以将第一图像序列中包括的第一图像的数量设置的较大；在图像处理装置的计算能力较低的情况下，可以将第一图像序列中包括的第一图像的数量设置的较小。
以第一图像序列中包括4张第一图像，且第二下采样因子为2的情况下，第一图像序列中可以分别包括：M/8×N/8的第一图像、M/4×N/4的第一图像、M/2×N/2的第一图像、M×N的第一图像。
在一些实施例中，第一图像序列可以称为图像金字塔，图像金字塔从上到下包括尺寸从小到大排列的至少两张第一图像。
梯度图像序列中梯度图像中像素的像素值，可以是通过对应的第一图像中像素的像素值进行梯度计算得到的。在一些实施例中，通过一个第一图像，确定梯度图像中第i行第j列像素的像素值的计算方式可以为：dx(i,j)＝I(i+1,j)-I(i,j)，dy(i,j)＝I(i,j+1)-I(i,j)。其中，dx(i,j)可以为第一图像中第i行第j列像素在x方向上的梯度，I(i+1,j)可以为第一图像中第(i+1)行第j列像素的像素值，I(i,j)可以为第一图像中第i行第j列像素的像素值；dy(i,j)可以为第一图像中第i行第j列像素在y方向上的梯度，I(i,j+1)可以为第一图像中第i行第(j+1)列像素的像素值。在另一些实施例中，通过一个第一图像，确定梯度图像中第i行第j列像素的像素值的计算方式可以为：dx(i,j)＝[I(i+1,j)-I(i-1,j)]/2；dy(i,j)＝[I(i,j+1)-I(i,j-1)]/2。其中，I(i-1,j)可以为第一图像中第(i-1)行第j列像素的像素值；I(i,j-1)可以为第一图像中第i行第(j-1)列像素的像素值。
可以基于dx(i,j)和dy(i,j)确定第i行第j列像素的梯度值。例如可以通过公式G(x,y)＝sqrt{[dx(i,j)]^2+[dy(i,j)]^2}确定第i行第j列像素的梯度值，其中，G(x,y)第i行第j列像素的梯度值，其中，sqrt指平方 根计算，[dx(i,j)]^2是对[dx(i,j)]作平方运算，[dy(i,j)]^2是对[dy(i,j)]作平方运算。这样，一个第一图像对应的梯度图像中第i行第j列像素的像素值，为第i行第j列像素的梯度值。
在一些实施例中，如果第i行第j列像素为边缘像素的情况下，可以将第i行第j列像素值确定为第i行第j列像素的梯度值。在另一些实施例中，如果第i行第j列像素为边缘像素的情况下，可以在边缘像素的外侧添加0或者其它数值的元素，其它数值例如是1或0.5等，从而通过上述的计算方法，先确定第i行第j列像素在x方向上的梯度dx(i,j)和在y方向上的梯度dy(i,j)，然后再通过dx(i,j)和dy(i,j)，确定第i行第j列像素的梯度值。在又一些实施例中，如果第i行第j列像素为边缘像素的情况下，则将第i行第j列像素的梯度值设置为指定值，指定值可以是[0,1]范围内的任一值，例如，指定值可以为0、0.5或者1等。
在一些实施例中，所述分别对所述第一图像序列中的所述第一图像进行图像梯度处理，得到所述梯度图像序列，可以包括：分别对所述第一图像序列中的所述第一图像均采用第二降噪参数进行降噪处理，得到第二图像序列；分别对所述第二图像序列中的第二图像进行图像梯度处理，得到所述梯度图像序列。
在这种方式下，可先对第一图像序列中第一图像进行降噪处理，然后再对降噪处理得到的第二图像序列进行图像梯度处理，得到梯度图像序列，从而能够减少图像噪声对梯度计算的影响，使得到的梯度图像序列准确。
第二降噪参数可以是预先设置的降噪参数，对不同的第一图像在降噪时可以采用相同或不同的第二降噪参数。在一些实施例中，第二降噪参数的选用还可以和图像拍摄时的参数有关，例如，在图像拍摄时采用较高的感光度值和/或较高的亮度值的情况下，可以使用较强降噪力度的第二降噪参数；在图像拍摄时采用较低的感光度值和/或较低的亮度值的情况下，可以使用较弱降噪力度的第二降噪参数，其中，国际标准组织对感光度量化规定为ISO(International Organization for Standardization)。
通过这种方式，由于先对第一图像序列中的第一图像进行降噪处理，然后再得到梯度图像序列，从而能够减少图像噪声对梯度计算的影响，使得到的梯度图像序列更加准确。
S502、基于所述梯度图像序列，确定所述纹理图像。
由于梯度图像序列中包括至少两个不同尺度(即不同像素尺寸)的梯度图像(即N个梯度图像，N个梯度图像的尺寸从大到小排列)，可以对至少两个不同尺度的梯度图像进行合并，得到纹理图像。
在N为2的情况下，S502可以通过以下方式实现：对第N个梯度图像进行上采样得到的第一个上采样图像，与第(N-1)个梯度图像进行合并，得到合并图像(即第一个合并图像)；将所述合并图像确定为所述纹理图像。
这样，在N为2的情况下，N个梯度图像包括第一个梯度图像和第二个梯度图像，第一个梯度图像的尺寸大于第二个梯度图像的尺寸，可以对第二个梯度图像进行上采样得到的图像，与第一个梯度图像进行合并，得到合并图像，将所述合并图像确定为所述纹理图像。
在N大于2的情况下，S502可以通过以下方式实现：对第N个梯度图像进行上采样得到的第一个上采样图像，与第(N-1)个梯度图像进行合并，得到合并图像；对得到的第j个合并图像进行上采样，得到第(j+1)个上采样图像；将所述第(j+1)个上采样图像与第(N-1-j)个梯度图像进行合并，得到第(j+1)个合并图像；j为大于或等于1的整数；将最后一次得到的合并图像确定为所述纹理图像。
以N等于4进行举例，可以对第四个梯度图像进行上采样得到的第一个上采样图像，与第三个梯度图像进行合并，得到合并图像(即第一个合并图像)；然后对得到的第一个合并图像进行上采样，得到第二个上采样图像，将第二个上采样图像与第二个梯度图像进行合并，得到第二个合并图像；然后对得到的第二个合并图像进行上采样，得到第三个上采样图像，将第三个上采样图像与第一个梯度图像进行合并，得到第三个合并图像；在得到第三个合并图像之后，由于没有其它的梯度图像来进行合并，因此，第三个合并图像就是最后一次得到的合并图像，将最后一次得到的合并图像确定为所述纹理图像。
第一个上采样图像的尺寸可以与第(N-1)个梯度图像的尺寸相同。第(j+1)个上采样图像的尺寸可以与第(N-1-j)个梯度图像的尺寸相同。
示例性地，对每个上采样图像和对应的梯度图像进行合并的方式可以包括：将每个上采样图像中每个像素的像素值，与对应的梯度图像中每个像素的像素值对应进行数学运算(例如以下至少之一的运算：相乘、相加、相减、相除、指数运算、对数运算等)，得到的结果确定为合并图像。
示例性地，对第一个上采样图像与第(N-1)个梯度图像进行合并，得到第一个合并图像，可以包括：对第一个上采样图像中的每个像素值乘以目标系数，得到第一个加权图像，将第一个加权图像中每个像素的像素值，与第(N-1)个梯度图像中每个像素的像素值对应相乘，得到第一个合并图像。将所述第(j+1)个上采样图像与第(N-1-j)个梯度图像进行合并，得到第(j+1)个合并图像，可以包括：将所述第(j+1)个上采样图像中的每个像素值乘以目标系数，得到第(j+1)个加权图像，将第(j+1)个加权图像中每个像素的像素值，与第(N-1-j)个梯度图像中每个像素的像素值对应相乘，得到第(j+1)个合并图像。
目标系数可以是大于1或者小于1的数。例如，目标系数可以为0.5、0.8、0.9、1.1或1.2等等。
通过这种方式，对得到的第j个合并图像进行上采样，得到第(j+1)个上采样图像，将第(j+1)个上采样图像与第(N-1-j)个梯度图像进行合并，得到第(j+1)个合并图像，将最后一次得到的合并图像确定为纹理图像，纹理图像能够通过对梯度图像序列中每个梯度图像合并得到，从而纹理图像能够准确地反映出待处理图像的纹理。
S503、确定待处理图像的语义分割图像。
S504、对所述语义分割图像和所述纹理图像进行融合，得到融合图像。
S505、基于所述融合图像，确定所述待处理图像对应的第一降噪参数。
S506、基于所述第一降噪参数，对所述待处理图像进行降噪处理。
在本公开实施例中，通过确定与待处理图像对应的梯度图像序列，基于梯度图像序列，确定纹理图像，从而纹理图像是通过各个不同尺度的梯度图像结合得到的，使得确定纹理图像能够准确地反映待处理图像的纹理。
图6为本公开实施例提供的一种基于梯度图像序列确定纹理图像的实现方式示意图，可以应用于图像处理装置，如图6所示，图像处理装置可以得到梯度图像序列，梯度图像序列包括4个梯度图像，分别为梯度图像61(可以是上述说明的第四个梯度图像)、梯度图像62(可以是上述说明的第三个梯度图像)、梯度图像63(可以是上述说明的第二个梯度图像)以及梯度图像64(可以是上述说明的第一个梯度图像)，4个梯度图像是通过待处理图像确定的。然后对梯度图像61进行上采样，得到图像65，将图像65和梯度图像62中对应像素点相乘，得到合并图像66。接着对合并图像66进行上采样得到图像67，将图像67和梯度图像63中对应像素点相乘，得到合并图像68。接着对合并图像68进行上采样得到图像69，将图像69和梯度图像64中对应像素点相乘，得到合并图像70。
其中，合并图像70是最后一次得到的合并图像，因此可以将合并图像70确定为待处理图像的纹理图像。
需要说明的是，图6只是给出了一种基于梯度图像序列确定纹理图像的流程示例，图6中梯度图像序列中的各个梯度图像，以及各个合并图像、各个上采样图像只是给出了一种参考性的示例，并不代表真实的各个梯度图像、各个合并图像、各个上采样图像。图6中各个图像中线条的粗细、纹理和灰度，并不代表真实的粗细、纹理和灰度，实际的粗细、纹理和灰度可以根据实际情况进行相应地更改。图6中中各个图像的不同部分的比例并不代表真实的比例。
图7为本公开另一实施例提供的一种图像处理方法的实现流程示意图，如图7所示，该方法应用于图像处理装置，该方法包括：
S701、确定待处理图像的语义分割图像，以及确定所述待处理图像的纹理图像。
S702、确定所述语义分割图像中至少一个区域对应的权重值。
语义分割图像可以包括多个不同的区域，不同的区域对应不同的对象，不同的区域采用不同的标签进行标识。例如，语义分割图像可以包括0、1、2、3、4这样的标签类别，示例性地，天空区域为0，显示为蓝色，绿植树叶区域为1，显示为绿色，建筑区域为2，显示为黄色，地面区域为3，显示为紫色，其它类别区域用4表示，显示为红色。在实施过程中，语义分割图像中的每个像素可以是0、1、2、3、4中的一个值。
可以基于语义分割图像设置不同的权重值，权重值越大代表所需的降噪力度越大，权重值越小代表所需的降噪力度越小。在一些实施例中，权重值的取值范围可以为[-1,1]，本公开实施例不限定权重值的取值范围，例如权重值的取值范围还可以为[-2,2]、[0,1]或者[-255,255]等。
示例性地，天空较为清澈，因此天空区域对应的权重值可以较大，而绿植树叶所携带的内容较为丰富，因此绿植树叶区域对应的权重值可以较小。
S703、将所述语义分割图像中的至少一个区域的像素值，修改为所述至少一个区域对应的权重值，得到权重图像。
例如，语义分割图像中像素值为0、1、2、3、4的区域，对应的权重值可以分别为1、0.2、-0.3、0.2、0，这样，可以将语义分割图像中值0、1、2、3、4的像素值，分别替换为1、0.2、-0.3、0.2、0，得到权重图像。
S704、将所述权重图像和所述纹理图像进行融合，得到所述融合图像。
在一些实施例中，S704可以通过以下方式实现：确定纹理图像对应的权值图像，将权重图像和权值图像进行融合，得到融合图像。例如，可以将纹理图像中每个像素的像素值替换为该像素值对应的十分位数，得到权值图像。例如，在纹理图像中第k个像素的像素值为0.512的情况下，将第k个像素的像素值修改0.5。
通过这种方式，可将语义分割图像中的至少一个区域的像素值，修改为至少一个区域对应的权重值， 得到权重图像；将权重图像和纹理图像进行融合，得到融合图像，从而为待处理图像中不同的语义区域设置不同的权重值，进而可以基于权重值来确定降噪力度，提升了待处理图像的降噪能力。
在另一些实施例中，S704可以通过以下方式实现：将所述纹理图像中的像素值与所述权重图像中的像素值对应相减，得到目标图像；基于所述目标图像，确定所述融合图像。
在又一些实施例中，S704可以通过以下方式实现：将所述纹理图像中的像素值与所述权重图像中的像素值对应相乘，得到目标图像；基于所述目标图像，确定所述融合图像。
在一些实施例中，可以直接将得到的目标图像确定为融合图像。
在另一些实施例中，将所述目标图像中大于第一阈值的像素值，修改为所述第一阈值，且将所述目标图像中小于第二阈值的像素值，修改为所述第二阈值，得到所述融合图像；所述第一阈值大于所述第二阈值。
在一些实施例方式中，第一阈值可以为1，第二阈值可以为0。
通过这种方式，可基于将纹理图像中的像素值与权重图像中的像素值对应相减得到的目标图像，确定融合图像，从而提供了一种权重图像和纹理图像融合的实现方式，使得融合图像能够准确的表达待处理图像携带的信息；并且，通过将目标图像中大于第一阈值的像素值，修改为第一阈值，且将目标图像中小于第二阈值的像素值，修改为第二阈值，从而基于融合图像能够容易地区分出待处理图像的纹理区域和平坦区域，从而针对待处理图像中纹理区域和平坦区域进行差异化的降噪，能够在对待处理图像降噪的同时，不会丢失待处理图像中的细节，提升了待处理图像的降噪能力。
S705、基于所述融合图像，确定所述待处理图像对应的第一降噪参数。
S706、基于所述第一降噪参数，对所述待处理图像进行降噪处理。
在本公开实施例中，基于深度学习分割算法，将场景不同区域分类，得到语义分割图像；使用图像梯度信息计算出图像纹理边界，结合分类信息并生成掩码(对应上述的纹理图像)；使用生成细节及平坦区的掩码(对应上述的融合图像)，使用不同的降噪力度进行降噪。
在本公开实施例中，考虑算法效果及性能平衡，采用快速引导滤波或快速双边滤波作为基础降噪模块，将掩码(对应上述的融合图像)作为降噪输入参数控制降噪范围(对应上述的窗口大小)及高斯标准差大小(对应上述的标准差)；例如：当掩码数值越小的时，说明此区域为平坦区，降噪半径及高斯标准差函数需要相应的增大，降噪力度增强；反之，当掩码数值越大，说明纹理区细节丰富，需减小降噪力度。
通过采用本公开实施例提供的图像处理方法，能够在有效去除天空区域噪声的同时，保留树叶的细节。
本公开实施例提出了一种基于图像分割及纹理检测的暗光场景降噪与增强方法，对比相关技术中存在的降噪与细节保留难以平衡，本公开实施例在降噪更干净的基础上保留更多图像细节。本公开实施例可作为基础模块嵌套于各类中终端或者平台使用。本公开实施例使用轻量级深度学习网络和梯度计算掩码相结合的方式，相较于使用深度学习网络或者传统方法计算纹理掩码，效果更好，得到的掩码更加精细化，更加准确，抗噪声能力更强，适用于极端场景。
本公开实施例可以具有以下应用场景：当用户在暗光场景室内室外拍照时，发现噪声太大或细节涂抹严重或拍照耗时太久时，采用本方法可以有效的去除噪声且保留细节。当用户在夜晚或暗光场景下拍摄视频后保存下来进行后处理时，应用本方法，可以有效快速的处理成高质量的视频。
在一些实施例中，在用户进行拍照时，用户在按下快门键之后，图像处理装置将拍摄的图像作为待处理图像，然后采用本公开实施例中的图像处理方法对待处理图像进行降噪处理。在另一些实施例中，用户在确定需要对本地保存的图像进行去噪的时候，可以选择该图像，图像处理装置将该图像确定为待处理图像，用户可以选择显示的去噪按钮，以使图像处理装置采用本公开实施例中的图像处理方法对待处理图像进行降噪处理。
基于前述的实施例，本公开实施例提供一种图像处理装置，该装置包括所包括的各单元、以及各单元所包括的各模块，可以通过图像处理装置中的处理器来实现；当然也可通过具体的逻辑电路实现。
图8为本公开实施例提供的一种图像处理装置的组成结构示意图，如图8所示，图像处理装置800包括：
确定单元801，用于确定待处理图像的语义分割图像，以及确定所述待处理图像的纹理图像；融合单元802，用于对所述语义分割图像和所述纹理图像进行融合，得到融合图像；所述确定单元801，还用于基于所述融合图像，确定所述待处理图像对应的第一降噪参数；降噪单元803，用于基于所述第一降噪参数，对所述待处理图像进行降噪处理。
在一些实施例中，确定单元801，还用于基于所述融合图像，确定所述待处理图像中的像素分别对应的第一降噪参数；其中，不同像素值的像素对应不同的第一降噪参数；降噪单元803，还用于基于所述像 素对应的第一降噪参数，对所述像素进行降噪处理。
在一些实施例中，所述第一降噪参数包括至少两个子降噪参数，所述至少两个子降噪参数对应的降噪力度不同；确定单元801，还用于基于所述融合图像，将所述待处理图像分为至少两个区域；根据所述至少两个子降噪参数，分别确定所述区域对应的子降噪参数；降噪单元803，还用于基于所述区域对应的子降噪参数，分别对所述区域进行降噪处理。
在一些实施例中，确定单元801，还用于确定与所述待处理图像对应的梯度图像序列；所述梯度图像序列包括至少两个不同尺度的归一化梯度图像；基于所述梯度图像序列，确定所述纹理图像。
在一些实施例中，确定单元801，还用于确定第一图像序列；所述第一图像序列包括N个第一图像，所述N个第一图像中第i个第一图像是对第(i-1)个第一图像进行下采样得到，第一个第一图像为所述待处理图像；N为大于或等于2的整数，i为大于或等于2的整数；分别对所述第一图像序列中的所述第一图像进行图像梯度处理，得到所述梯度图像序列。
在一些实施例中，确定单元801，还用于分别对所述第一图像序列中的所述第一图像均采用第二降噪参数进行降噪处理，得到第二图像序列；分别对所述第二图像序列中的第二图像进行图像梯度处理，得到所述梯度图像序列。
在一些实施例中，确定单元801，还用于对第N个梯度图像进行上采样得到的第一个上采样图像，与第(N-1)个梯度图像进行合并，得到合并图像；N为大于或等于2的整数；在N为2的情况下，将所述合并图像确定为所述纹理图像；在N大于2的情况下，对得到的第j个合并图像进行上采样，得到第(j+1)个上采样图像；将所述第(j+1)个上采样图像与第(N-1-j)个梯度图像进行合并，得到第(j+1)个合并图像；j为大于或等于1的整数；将最后一次得到的合并图像确定为所述纹理图像。
在一些实施例中，融合单元802，还用于确定所述语义分割图像中至少一个区域对应的权重值；将所述语义分割图像中的至少一个区域的像素值，修改为所述至少一个区域对应的权重值，得到权重图像；将所述权重图像和所述纹理图像进行融合，得到所述融合图像。
在一些实施例中，融合单元802，还用于将所述纹理图像中的像素值与所述权重图像中的像素值对应相减，得到目标图像；将所述目标图像中大于第一阈值的像素值，修改为所述第一阈值，且将所述目标图像中小于第二阈值的像素值，修改为所述第二阈值，得到所述融合图像；所述第一阈值大于所述第二阈值。
在一些实施例中，确定单元801，还用于获取所述融合图像中第一像素的像素值；根据所述第一像素的像素值大小，为与所述第一像素分别对应的所述待处理图像中的第二像素设置第一降噪参数；其中，所述第一降噪参数包括标准差和窗口大小，若第m个第一像素的像素值大于第n个第一像素的像素值，则第m个第二像素的标准差小于第n个第二像素的标准差，且所述第m个第二像素的窗口大小小于所述第n个第二像素的窗口大小，m和n不同，m和n均为大于或等于1的整数。
在一些实施例中，所述至少两个区域包括第一区域、第二区域以及第三区域；确定单元801，还用于获取所述融合图像中第一像素的像素值；根据所述第一像素的像素值大小，将所述待处理图像分为所述第一区域、所述第二区域以及所述第三区域；其中，与所述第一区域中第二像素分别对应的所述融合图像中第一像素的像素值，为所述融合图像中像素值的最小值；与所述第三区域中第二像素分别对应的所述融合图像中第一像素的像素值，为所述融合图像中像素值的最大值；所述第二区域为所述待处理图像中除所述第一区域和所述第三区域之外的区域；所述第一区域对应的子降噪参数、所述第二区域对应的子降噪参数以及所述第三区域对应的子降噪参数的降噪力度依次减小。
在一些实施例中，所述至少两个区域包括第四区域和第五区域；确定单元801，还用于获取所述融合图像中第一像素的像素值；根据所述第一像素的像素值大小，将所述待处理图像分为所述第四区域和所述第五区域；其中，与所述第四区域中第二像素分别对应的所述融合图像中第一像素的像素值，大于第三阈值；与所述第五区域中第二像素分别对应的所述融合图像中第一像素的像素值，小于或等于所述第三阈值；所述第四区域对应的子降噪参数的降噪力度，小于所述第五区域对应的子降噪参数的降噪力度。
以上装置实施例的描述，与上述方法实施例的描述是类似的，具有同方法实施例相似的有益效果。对于本公开装置实施例中未披露的技术细节，请参照本公开方法实施例的描述而理解。
需要说明的是，本公开实施例中，如果以软件功能模块的形式实现上述的图像处理方法，并作为独立的产品销售或使用时，也可以存储在一个计算机可读存储介质中。基于这样的理解，本公开实施例的技术方案本质上或者说对相关技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台图像处理装置执行本公开各个实施例所述方法的全部或部分。
图9为本公开实施例提供的一种图像处理设备的硬件实体示意图，如图9所示，该图像处理设备900的硬件实体包括：处理器901和存储器902，其中，存储器902存储有可在处理器901上运行的计算机程 序，处理器901执行程序时实现上述任一实施例的图像处理方法。
存储器902存储有可在处理器上运行的计算机程序，存储器902配置为存储由处理器901可执行的指令和应用，还可以缓存待处理器901以及图像处理设备900中各模块待处理或已经处理的数据(例如，图像数据、音频数据、语音通信数据和视频通信数据)，可以通过闪存(FLASH)或随机访问存储器(Random Access Memory，RAM)实现。
处理器901执行程序时实现上述任一项的图像处理方法。处理器901通常控制图像处理设备900的总体操作。
本公开实施例提供一种计算机可读存储介质，计算机可读存储介质存储有一个或者多个程序，该一个或者多个程序可被一个或者多个处理器执行，以实现如上任一实施例的图像处理方法。计算机可读存储介质可以为易失性存储介质或非易失性存储介质。
本公开实施例还可以提供一种芯片，芯片包括处理器，处理器可以从存储器中调用并运行计算机程序，以实现本公开实施例中的图像处理方法。
芯片还可以包括存储器。其中，处理器可以从存储器中调用并运行计算机程序，以实现本公开实施例中的图像处理方法。
其中，存储器可以是独立于处理器的一个单独的器件，也可以集成在处理器中。
在一些实施例中，该芯片还可以包括输入接口。其中，处理器可以控制该输入接口与其他设备或芯片进行通信，具体地，可以获取其他设备或芯片发送的信息或数据。
在一些实施例中，该芯片还可以包括输出接口。其中，处理器可以控制该输出接口与其他设备或芯片进行通信，具体地，可以向其他设备或芯片输出信息或数据。
在一些实施例中，该芯片可应用于本公开实施例中的图像处理设备，并且该芯片可以实现本公开实施例的各个方法中由图像处理设备实现的相应流程，为了简洁，在此不再赘述。
应理解，本公开实施例提到的芯片还可以称为系统级芯片，系统芯片，芯片系统或片上系统芯片等。
本公开实施例还提供了一种计算机程序产品，所述计算机程序产品包括计算机可读存储介质，所述计算机可读存储介质存储计算机程序，所述计算机程序包括能够由至少一个处理器执行的指令，当所述指令由所述至少一个处理器执行时实现本公开实施例中的图像处理方法。
本公开实施例还提供了一种计算机程序所述计算机程序使得计算机执行本公开实施例中的图像处理方法。
这里需要指出的是：以上图像处理设备、计算机可读存储介质、芯片、计算机程序产品、计算机程序实施例的描述，与上述方法实施例的描述是类似的，具有同方法实施例相似的有益效果。对于本公开图像处理设备、计算机可读存储介质、芯片、计算机程序产品、计算机程序实施例中未披露的技术细节，请参照本公开方法实施例的描述而理解。
上述图像处理装置、芯片或处理器可以包括以下任一个或多个的集成：特定用途集成电路(Application Specific Integrated Circuit，ASIC)、数字信号处理器(Digital Signal Processor，DSP)、数字信号处理装置(Digital Signal Processing Device，DSPD)、可编程逻辑装置(Programmable Logic Device，PLD)、现场可编程门阵列(Field Programmable Gate Array，FPGA)、中央处理器(Central Processing Unit，CPU)、图形处理器(Graphics Processing Unit，GPU)、嵌入式神经网络处理器(neural-network processing units，NPU)、控制器、微控制器、微处理器、可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。可以理解地，实现上述处理器功能的电子器件还可以为其它，本公开实施例不作具体限定。
上述计算机可读存储介质/存储器可以是只读存储器(Read Only Memory，ROM)、可编程只读存储器(Programmable Read-Only Memory，PROM)、可擦除可编程只读存储器(Erasable Programmable Read-Only Memory，EPROM)、电可擦除可编程只读存储器(Electrically Erasable Programmable Read-Only Memory，EEPROM)、磁性随机存取存储器(Ferromagnetic Random Access Memory，FRAM)、快闪存储器(Flash Memory)、磁表面存储器、光盘、或只读光盘(Compact Disc Read-Only Memory，CD-ROM)等存储器；也可以是包括上述存储器之一或任意组合的各种终端，如移动电话、计算机、平板设备、个人数字助理等。
应理解，说明书通篇中提到的“一个实施例”或“一实施例”或“本公开实施例”或“前述实施例”或“一些实施方式”或“一些实施例”意味着与实施例有关的特定特征、结构或特性包括在本公开的至少一个实施例中。因此，在整个说明书各处出现的“在一个实施例中”或“在一实施例中”或“本公开实施例”或“前述实施例”或“一些实施方式”或“一些实施例”未必一定指相同的实施例。此外，这些特定的特征、结构或特性可以任意适合的方式结合在一个或多个实施例中。应理解，在本公开的各种实施例中，上述各过程的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定，而不应对本公开实施例的实施过程构成任何限定。上述本公开实施例序号仅仅为了描述，不 代表实施例的优劣。
在未做特殊说明的情况下，图像处理装置执行本公开实施例中的任一步骤，可以是图像处理装置的处理器执行该步骤。除非特殊说明，本公开实施例并不限定图像处理装置执行下述步骤的先后顺序。另外，不同实施例中对数据进行处理所采用的方式可以是相同的方法或不同的方法。还需说明的是，本公开实施例中的任一步骤是图像处理装置可以独立执行的，即图像处理装置执行上述实施例中的任一步骤时，可以不依赖于其它步骤的执行。
此外，术语第一、第二仅用于描述目的，而不能理解为指示或暗示相对重要性或者隐含指明所指示的技术特征的数量。由此，限定有第一、第二的特征可以明示或者隐含地包括一个或者更多个所述特征。
在本公开所提供的几个实施例中，应该理解到，所揭露的设备和方法，可以通过其它的方式实现。以上所描述的设备实施例仅仅是示意性的，例如，所述单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，如：多个单元或组件可以结合，或可以集成到另一个系统，或一些特征可以忽略，或不执行。另外，所显示或讨论的各组成部分相互之间的耦合、或直接耦合、或通信连接可以是通过一些接口，设备或单元的间接耦合或通信连接，可以是电性的、机械的或其它形式的。
上述作为分离部件说明的单元可以是、或也可以不是物理上分开的，作为单元显示的部件可以是、或也可以不是物理单元；既可以位于一个地方，也可以分布到多个网络单元上；可以根据实际的需要选择其中的部分或全部单元来实现本实施例方案的目的。
另外，在本公开各实施例中的各功能单元可以全部集成在一个处理单元中，也可以是各单元分别单独作为一个单元，也可以两个或两个以上单元集成在一个单元中；上述集成的单元既可以采用硬件的形式实现，也可以采用硬件加软件功能单元的形式实现。
本公开所提供的几个方法实施例中所揭露的方法，在不冲突的情况下可以任意组合，得到新的方法实施例。
本公开所提供的几个产品实施例中所揭露的特征，在不冲突的情况下可以任意组合，得到新的产品实施例。
本公开所提供的几个方法或设备实施例中所揭露的特征，在不冲突的情况下可以任意组合，得到新的方法实施例或设备实施例。
本领域普通技术人员可以理解：实现上述方法实施例的全部或部分步骤可以通过程序指令相关的硬件来完成，前述的程序可以存储于计算机可读存储介质中，该程序在执行时，执行包括上述方法实施例的步骤；而前述的存储介质包括：移动存储设备、只读存储器(Read Only Memory，ROM)、磁碟或者光盘等各种可以存储程序代码的介质。
或者，本公开上述集成的单元如果以软件功能模块的形式实现并作为独立的产品销售或使用时，也可以存储在一个计算机可读存储介质中。基于这样的理解，本公开实施例的技术方案本质上或者说对相关技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机、服务器、或者网络设备等)执行本公开各个实施例所述方法的全部或部分。而前述的存储介质包括：移动存储设备、ROM、磁碟或者光盘等各种可以存储程序代码的介质。
在本公开实施例中，不同实施例中相同步骤和相同内容的说明，可以互相参照。在本公开实施例中，术语“并”不对步骤的先后顺序造成影响，例如，图像处理装置执行A，并执行B，可以是图像处理装置先执行A，再执行B，或者是图像处理装置先执行B，再执行A，或者是图像处理装置执行A的同时执行B。
在本公开实施例和所附权利要求书中所使用的单数形式的一种、所述和该也旨在包括多数形式，除非上下文清楚地表示其他含义。
应当理解，本文中使用的术语“和/或”仅仅是一种描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况。
需要说明的是，本公开所涉及的各个实施例中，可以执行全部的步骤或者可以执行部分的步骤，只要能够形成一个完整的技术方案即可。
工业实用性：
本公开实施例所描述的图像处理方法适用于计算机视觉领域，并且特别地适用于对待处理的视频和/或图像数据，比如，图像处理装置拍摄的视频和/或图像、图像处理装置存储空间中存储的视频和/或图像、图像处理装置接收其他设备发送的视频和/或图像等，进行有效地降噪处理，使处理后的图像，在去除噪声的同时保留了图像细节，大大提高了图像质量。
例如，当用户在暗光场景室内室外拍照时，发现噪声太大或细节涂抹严重或拍照耗时太久时，采用本方法可以有效的去除噪声且保留细节。当用户在夜晚或暗光场景下拍摄视频后保存下来进行后处理时， 应用本方法，可以有效快速的处理成高质量的视频。当用户在确定需要对本地保存的图像进行去噪的时候，可以选择该图像，图像处理装置将该图像确定为待处理图像，用户可以选择显示的去噪按钮，进而应用本方法得到高质量的图像。
本公开对工业实用性不作限制，在应用过程中，对于需要提高待处理图像质量的场景，比如需要对待处理图像降噪，且减少待处理图像中的细节丢失的场景，可使用本公开描述的图像处理方法。
以上所述，仅为本公开的实施方式，但本公开的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本公开揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本公开的保护范围之内。因此，本公开的保护范围应以所述权利要求的保护范围为准。

Claims (16)

  1. 一种图像处理方法，所述方法包括：
    确定待处理图像的语义分割图像，以及确定所述待处理图像的纹理图像；
    对所述语义分割图像和所述纹理图像进行融合，得到融合图像；
    基于所述融合图像，确定所述待处理图像对应的第一降噪参数；
    基于所述第一降噪参数，对所述待处理图像进行降噪处理。
  2. 根据权利要求1所述的图像处理方法，其中，所述基于所述融合图像，确定所述待处理图像对应的第一降噪参数，包括：
    基于所述融合图像，确定所述待处理图像中的像素分别对应的第一降噪参数；其中，不同像素值的像素对应不同的第一降噪参数；
    所述基于所述第一降噪参数，对所述待处理图像进行降噪处理，包括：
    基于所述像素对应的第一降噪参数，对所述像素进行降噪处理。
  3. 根据权利要求1所述的图像处理方法，其中，所述第一降噪参数包括至少两个子降噪参数，所述至少两个子降噪参数对应的降噪力度不同；
    所述基于所述融合图像，确定所述待处理图像对应的第一降噪参数，包括：
    基于所述融合图像，将所述待处理图像分为至少两个区域；
    根据所述至少两个子降噪参数，分别确定所述区域对应的子降噪参数；
    所述基于所述第一降噪参数，对所述待处理图像进行降噪处理，包括：
    基于所述区域对应的子降噪参数，分别对所述区域进行降噪处理。
  4. 根据权利要求1至3任一项所述的图像处理方法，其中，所述确定所述待处理图像的纹理图像，包括：
    确定与所述待处理图像对应的梯度图像序列；所述梯度图像序列包括至少两个不同尺度的归一化梯度图像；
    基于所述梯度图像序列，确定所述纹理图像。
  5. 根据权利要求4所述的图像处理方法，其中，所述确定与所述待处理图像对应的梯度图像序列，包括：
    确定第一图像序列；所述第一图像序列包括N个第一图像，所述N个第一图像中第i个第一图像是对第(i-1)个第一图像进行下采样得到，第一个第一图像为所述待处理图像；N为大于或等于2的整数，i为大于或等于2的整数；
    分别对所述第一图像序列中的所述第一图像进行图像梯度处理，得到所述梯度图像序列。
  6. 根据权利要求5所述的图像处理方法，其中，所述分别对所述第一图像序列中的所述第一图像进行图像梯度处理，得到所述梯度图像序列，包括：
    分别对所述第一图像序列中的所述第一图像均采用第二降噪参数进行降噪处理，得到第二图像序列；
    分别对所述第二图像序列中的第二图像进行图像梯度处理，得到所述梯度图像序列。
  7. 根据权利要求4所述的图像处理方法，其中，所述基于所述梯度图像序列，确定所述纹理图像，包括：
    对第N个梯度图像进行上采样得到的第一个上采样图像，与第(N-1)个梯度图像进行合并，得到合并图像；N为大于或等于2的整数；
    在N为2的情况下，将所述合并图像确定为所述纹理图像；
    在N大于2的情况下，对得到的第j个合并图像进行上采样，得到第(j+1)个上采样图像；将所述第(j+1)个上采样图像与第(N-1-j)个梯度图像进行合并，得到第(j+1)个合并图像；j为大于或等于1的整数；将最后一次得到的合并图像确定为所述纹理图像。
  8. 根据权利要求1至3任一项所述的图像处理方法，其中，所述对所述语义分割图像和所述纹理图像进行融合，得到融合图像，包括：
    确定所述语义分割图像中至少一个区域对应的权重值；
    将所述语义分割图像中的至少一个区域的像素值，修改为所述至少一个区域对应的权重值，得到权重图像；
    将所述权重图像和所述纹理图像进行融合，得到所述融合图像。
  9. 根据权利要求8所述的图像处理方法，其中，所述将所述权重图像和所述纹理图像进行融合，得到所述融合图像，包括：
    将所述纹理图像中的像素值与所述权重图像中的像素值对应相减，得到目标图像；
    将所述目标图像中大于第一阈值的像素值，修改为所述第一阈值，且将所述目标图像中小于第二阈 值的像素值，修改为所述第二阈值，得到所述融合图像；所述第一阈值大于所述第二阈值。
  10. 根据权利要求2所述的图像处理方法，其中，所述基于所述融合图像，确定所述待处理图像中的像素分别对应的第一降噪参数，包括：
    获取所述融合图像中第一像素的像素值；
    根据所述第一像素的像素值大小，为与所述第一像素分别对应的所述待处理图像中的第二像素设置第一降噪参数；
    其中，所述第一降噪参数包括标准差和窗口大小，若第m个第一像素的像素值大于第n个第一像素的像素值，则第m个第二像素的标准差小于第n个第二像素的标准差，且所述第m个第二像素的窗口大小小于所述第n个第二像素的窗口大小，m和n不同，m和n均为大于或等于1的整数。
  11. 根据权利要求3所述的图像处理方法，其中，所述至少两个区域包括第一区域、第二区域以及第三区域；所述基于所述融合图像，将所述待处理图像分为至少两个区域，包括：
    获取所述融合图像中第一像素的像素值；
    根据所述第一像素的像素值大小，将所述待处理图像分为所述第一区域、所述第二区域以及所述第三区域；
    其中，与所述第一区域中第二像素分别对应的所述融合图像中第一像素的像素值，为所述融合图像中像素值的最小值；
    与所述第三区域中第二像素分别对应的所述融合图像中第一像素的像素值，为所述融合图像中像素值的最大值；
    所述第二区域为所述待处理图像中除所述第一区域和所述第三区域之外的区域；
    所述第一区域对应的子降噪参数、所述第二区域对应的子降噪参数以及所述第三区域对应的子降噪参数的降噪力度依次减小。
  12. 根据权利要求3所述的图像处理方法，其中，所述至少两个区域包括第四区域和第五区域；所述基于所述融合图像，将所述待处理图像分为至少两个区域，包括：
    获取所述融合图像中第一像素的像素值；
    根据所述第一像素的像素值大小，将所述待处理图像分为所述第四区域和所述第五区域；
    其中，与所述第四区域中第二像素分别对应的所述融合图像中第一像素的像素值，大于第三阈值；
    与所述第五区域中第二像素分别对应的所述融合图像中第一像素的像素值，小于或等于所述第三阈值；
    所述第四区域对应的子降噪参数的降噪力度，小于所述第五区域对应的子降噪参数的降噪力度。
  13. 一种图像处理装置，包括：
    确定单元，用于确定待处理图像的语义分割图像，以及确定所述待处理图像的纹理图像；
    融合单元，用于对所述语义分割图像和所述纹理图像进行融合，得到融合图像；
    所述确定单元，还用于基于所述融合图像，确定所述待处理图像对应的第一降噪参数；
    降噪单元，用于基于所述第一降噪参数，对所述待处理图像进行降噪处理。
  14. 一种图像处理设备，包括：存储器和处理器，
    所述存储器存储有可在所述处理器上运行的计算机程序，
    所述处理器执行所述计算机程序时实现权利要求1至12中任一项所述的图像处理方法。
  15. 一种计算机可读存储介质，所述计算机可读存储介质存储有一个或者多个程序，所述一个或者多个程序可被一个或者多个处理器执行，以实现权利要求1至12中任一项所述的图像处理方法。
  16. 一种计算机程序产品，包括计算机可读代码，或者承载有计算机可读代码的计算机可读存储介质，当所述计算机可读代码在电子设备的处理器中运行时，所述电子设备中的处理器执行用于实现权利要求1至12中的任一权利要求所述的图像处理方法。
PCT/CN2022/095025 2021-10-21 2022-05-25 图像处理方法、装置、设备、存储介质及计算机程序产品 WO2023065665A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111226036.8A CN113689373B (zh) 2021-10-21 2021-10-21 图像处理方法、装置、设备及计算机可读存储介质
CN202111226036.8 2021-10-21

Publications (1)

Publication Number Publication Date
WO2023065665A1 true WO2023065665A1 (zh) 2023-04-27

Family

ID=78587658

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/095025 WO2023065665A1 (zh) 2021-10-21 2022-05-25 图像处理方法、装置、设备、存储介质及计算机程序产品

Country Status (2)

Country Link
CN (1) CN113689373B (zh)
WO (1) WO2023065665A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116894796A (zh) * 2023-09-11 2023-10-17 归芯科技(深圳)有限公司 图像降噪和畸变矫正联合处理方法及装置

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113689373B (zh) * 2021-10-21 2022-02-11 深圳市慧鲤科技有限公司 图像处理方法、装置、设备及计算机可读存储介质
CN116029951A (zh) * 2022-05-27 2023-04-28 荣耀终端有限公司 图像处理方法与电子设备

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107507149A (zh) * 2017-08-31 2017-12-22 深圳市智图医疗技术有限责任公司 一种核磁共振成像图像的降噪方法及装置
US20170372455A1 (en) * 2016-06-22 2017-12-28 Adobe Systems Incorporated Enhanced Vectorization of Raster Images
CN108335306A (zh) * 2018-02-28 2018-07-27 北京市商汤科技开发有限公司 图像处理方法及装置、电子设备和存储介质
CN110781899A (zh) * 2019-10-23 2020-02-11 维沃移动通信有限公司 图像处理方法及电子设备
CN111583142A (zh) * 2020-04-30 2020-08-25 深圳市商汤智能传感科技有限公司 图像降噪方法及装置、电子设备和存储介质
CN113689373A (zh) * 2021-10-21 2021-11-23 深圳市慧鲤科技有限公司 图像处理方法、装置、设备及计算机可读存储介质

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7933454B2 (en) * 2007-06-25 2011-04-26 Xerox Corporation Class-based image enhancement system
CN105913402B (zh) * 2016-05-20 2019-04-16 上海海洋大学 一种基于ds证据理论的多幅遥感图像融合去噪方法
CN108122206A (zh) * 2016-11-29 2018-06-05 深圳市中兴微电子技术有限公司 一种低照度图像降噪方法及装置
CN112150371B (zh) * 2019-06-28 2024-02-06 浙江宇视科技有限公司 图像降噪方法、装置、设备及存储介质
CN110516678B (zh) * 2019-08-27 2022-05-06 北京百度网讯科技有限公司 图像处理方法和装置
CN111127347A (zh) * 2019-12-09 2020-05-08 Oppo广东移动通信有限公司 降噪方法、终端及存储介质
CN112053358A (zh) * 2020-09-28 2020-12-08 腾讯科技（深圳）有限公司 图像中像素的实例类别确定方法、装置、设备及存储介质
CN113011425A (zh) * 2021-03-05 2021-06-22 上海商汤智能科技有限公司 图像分割方法、装置、电子设备及计算机可读存储介质
CN113486956B (zh) * 2021-07-07 2022-03-25 山东力聚机器人科技股份有限公司 目标分割系统及其训练方法、目标分割方法及设备

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170372455A1 (en) * 2016-06-22 2017-12-28 Adobe Systems Incorporated Enhanced Vectorization of Raster Images
CN107507149A (zh) * 2017-08-31 2017-12-22 深圳市智图医疗技术有限责任公司 一种核磁共振成像图像的降噪方法及装置
CN108335306A (zh) * 2018-02-28 2018-07-27 北京市商汤科技开发有限公司 图像处理方法及装置、电子设备和存储介质
CN110781899A (zh) * 2019-10-23 2020-02-11 维沃移动通信有限公司 图像处理方法及电子设备
CN111583142A (zh) * 2020-04-30 2020-08-25 深圳市商汤智能传感科技有限公司 图像降噪方法及装置、电子设备和存储介质
CN113689373A (zh) * 2021-10-21 2021-11-23 深圳市慧鲤科技有限公司 图像处理方法、装置、设备及计算机可读存储介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116894796A (zh) * 2023-09-11 2023-10-17 归芯科技(深圳)有限公司 图像降噪和畸变矫正联合处理方法及装置
CN116894796B (zh) * 2023-09-11 2023-12-26 归芯科技(深圳)有限公司 图像降噪和畸变矫正联合处理方法及装置

Also Published As

Publication number Publication date
CN113689373A (zh) 2021-11-23
CN113689373B (zh) 2022-02-11

Similar Documents

Publication Publication Date Title
WO2023065665A1 (zh) 图像处理方法、装置、设备、存储介质及计算机程序产品
JP6431245B1 (ja) エッジ認識双方向画像処理
WO2020192483A1 (zh) 图像显示方法和设备
Chen et al. Robust image and video dehazing with visual artifact suppression via gradient residual minimization
Wan et al. CoRRN: Cooperative reflection removal network
CN108895981B (zh) 一种三维测量方法、装置、服务器和存储介质
CN111259758B (zh) 一种针对密集区域的两阶段遥感图像目标检测方法
Park et al. Single image dehazing with image entropy and information fidelity
US11276177B1 (en) Segmentation for image effects
US20210004962A1 (en) Generating effects on images using disparity guided salient object detection
WO2021164234A1 (zh) 图像处理方法以及图像处理装置
WO2021063341A1 (zh) 图像增强方法以及装置
CN112446380A (zh) 图像处理方法和装置
CN112446834A (zh) 图像增强方法和装置
CN108765333B (zh) 一种基于深度卷积神经网络的深度图完善方法
WO2021003936A1 (zh) 图像分割方法、电子设备和计算机可读存储介质
CN115049717B (zh) 一种深度估计方法及装置
CN110717864B (zh) 一种图像增强方法、装置、终端设备及计算机可读介质
CN114581318A (zh) 一种低照明度图像增强方法及系统
CN116883770A (zh) 深度估计模型的训练方法、装置、电子设备及存储介质
CN110544221B (zh) 训练方法及装置、去雨方法、终端设备、存储介质
Jayanthi et al. Underwater haze removal using contrast boosted grayscale image
Yoshida et al. Saturated Reflection Detection for Reflection Removal Based on Convolutional Neural Network
CN116171463A (zh) 一种堆叠物体的识别方法、装置、设备及计算机存储介质
CN116883257A (zh) 图像去雾方法、装置、计算机设备和存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22882290

Country of ref document: EP

Kind code of ref document: A1