An image down-sampling method and system

Abstract

The application is directed to the field of image scaling and image processing, and discloses an image down-sampling method and system. Firstly, scale-related features are extracted from a high-resolution image by a scale-adaptive feature extraction module, and then the scale-related features are re-sampled according to sampling coordinates obtained by coordinate mapping to obtain multiple groups of candidate features with the same size as a low-resolution image. At the same time, a weight generator generates adaptive weights for each group of candidate features according to the sampling coordinates. Finally, the down-sampled low-resolution image is obtained by weighting and summing the candidate features using the generated adaptive weights. The image down-sampling model proposed by the method is trained under the guidance of the existing interpolation-based image super-resolution reconstruction method, and can effectively solve the problem that the low-resolution image obtained by the interpolation-based image down-sampling method is generally not fully applicable to the existing image super-resolution reconstruction method.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image scaling and image processing, in particular to an image down-sampling method and system. BACKGROUND

[0002] Image scaling is a comprehensive research field involving image signal processing, image down-sampling, image interpolation and image super-resolution reconstruction, aiming to adjust the size of digital images, mainly divided into two parts: image down-sampling and image super-resolution reconstruction.

[0003] Image down-sampling is mainly used to reduce the size of the original image, so that the reduced image conforms to the size of the display area on the display device. At the same time, image down-sampling is also used to generate a thumbnail corresponding to the original image. For a high-resolution image with an input size of HxW, after s-fold down-sampling, a low-resolution image with an output size of ( / )x( / ) is obtained. Image super-resolution reconstruction is a technology opposite to image down-sampling, aiming to recover the corresponding high-resolution image from the low-resolution image, which can be used to view image details or display images on a display device with higher resolution. Image super-resolution reconstruction mainly includes interpolation-based methods and learning-based methods. Compared with learning-based methods, interpolation-based methods have the advantages of fast processing speed, small amount of calculation and low space occupation, and are therefore adopted by various computer devices.

[0004] In related technologies, low-resolution images are usually obtained directly from high-resolution images by bicubic interpolation and other interpolation-based image down-sampling methods. Such methods can reduce the resolution of the original image while maintaining its visual quality. However, interpolation-based image down-sampling methods do not consider the methods used in image super-resolution reconstruction, so the low-resolution images obtained by interpolation-based image down-sampling methods are usually not fully applicable to existing image super-resolution reconstruction methods.

[0005] At present, there is no effective solution to the problem that low-resolution images obtained by interpolation-based image down-sampling methods in related technologies are usually not fully applicable to existing image super-resolution reconstruction methods. SUMMARY

[0006] The present application provides an image down-sampling method and system to solve the problem that low-resolution images obtained by interpolation-based image down-sampling methods in related technologies are usually not fully applicable to existing image super-resolution reconstruction methods. The image down-sampling method and system provided by the present application mainly generate low-resolution images suitable for interpolation-based image super-resolution reconstruction methods.

[0007] In a first aspect, the present application provides an image down-sampling method, the specific steps are as follows:

[0008] Step 1: The overall image down-sampling model construction. The present application adopts the form of convolutional neural network to construct the overall model, which approximates the nonlinear function and has the ability to map pixels to a higher dimensional feature space than traditional RGB images. The present application mainly uses convolutional layers, channel attention layers, residual layers, multi-layer perceptron and other modules as basic components to construct the image down-sampling model. The image down-sampling model construction steps are as follows:

[0009] Step 1.1: Calculate the coordinate map in high resolution space and low resolution space respectively. First, define the value range of the scale factor that controls the size of the low resolution image after down-sampling in the image down-sampling process as Λ, then the scale factor can be sampled from Λ, that is, s∈Λ. According to the high resolution image size [H,W] and the scale factor s, the target low resolution image size can be expressed as

[0010]

[0011] where represents the floor function. The coordinate map in high resolution space and low resolution space can be normalized as

[0012]

[0013] where the values in the coordinate map in high resolution space and low resolution space are normalized to the interval [-1,1], avoiding excessively large coordinate values, C HR and C LR are the coordinate maps in high resolution space and low resolution space, respectively, and the sizes of the two coordinate maps are [H,W] and [h,w], respectively. HR and LR represent high resolution and low resolution, respectively; i and j represent the value index of the coordinate map in the vertical direction (H / h direction) and the horizontal direction (W / w direction), respectively. For example, for C HR , the length of the vertical direction (H direction) is H, so the value of i is 0, 1, 2, …, H-1, and the length of the horizontal direction (W direction) is W, so the value of j is 0, 1, 2, …, W-1, so the value of (i,j) is the combination of the values of i and j; represents the set of integers; represents the set of integer coordinates in the range of HxW.

[0014] Step 1.2: Adjust the coordinate map in the high-resolution space and the coordinate map in the low-resolution space to the same size. The coordinate map in the high-resolution space and the coordinate map in the low-resolution space reflect the mapping relationship of the pixel coordinates in the image downsampling process, but the sizes of the coordinate map in the high-resolution space and the coordinate map in the low-resolution space are different. In order to utilize the pixel coordinate mapping relationship, it is necessary to adjust the coordinate map in the high-resolution space and the coordinate map in the low-resolution space to the same size. The present application considers that the size of the feature map needs to be consistent with the size of the input high-resolution image in the feature extraction process, and therefore selects to upsample the coordinate map in the low-resolution space to enlarge it to the size of the coordinate map in the high-resolution space. The upsampled coordinate map can be expressed as

[0015]

[0016] wherein f upsampling (·) represents the upsample operation.

[0017] Step 1.3: Calculate the feature scaling factor for adjusting the feature distribution. Considering that the relative distance between the coordinate map in the high-resolution space and the coordinate map in the low-resolution space reflects the mapping relationship of the pixels in the high-resolution image and the low-resolution image, the present application takes the relative distance between the coordinate map in the high-resolution space and the coordinate map in the low-resolution space as the feature scaling factor for adjusting the feature distribution, in order to scale the feature in the numerical value. The feature scaling factor can be expressed as

[0018]

[0019] wherein f distance (·) represents the relative distance calculation function.

[0020] Step 1.4: Calculate the feature offset parameter for adjusting the feature distribution. In order to adapt the feature extraction process to different scale factors, the present application calculates the feature offset parameter from the perspective of the size relationship between the high-resolution image and the low-resolution image, which is used to continuously adjust the feature distribution in the feature extraction process. First, the scale vectors in the vertical and horizontal directions can be defined as

[0021]

[0022] Considering the tensor shape [middle_channel, H, W] of the deep feature in the feature extraction process, in order to make the feature offset parameter have the same size as the deep feature, the present application expands the scale vector into a scale tensor S, whose shape can be expressed as [2, H, W]. Next, the feature offset parameter is calculated according to the scale tensor S. The present application selects a multi-layer perceptron to calculate the feature offset parameter, and the dynamically generated feature offset parameter can be expressed as

[0023]

[0024] wherein denotes the multi-layer perception selected by the present application, which aims to convert the input feature with a size of 2 into an output feature with a size of middle_channel.

[0025] Step 1.5: Constructing the scale-adaptive feature enhancement module. The present application combines the core component invention feature scaling coefficient and feature offset parameter mentioned in steps 1.3 and 1.4 with the convolution layer to construct the scale-adaptive feature enhancement module. For the input feature F in , the output feature F out obtained after implementing the scale-adaptive feature enhancement can be expressed as

[0026]

[0027] wherein f conv (·) denotes the convolution operation. Based on residual learning, the present application adjusts the distribution of deep features in the feature extraction process to adapt to different scale factors.

[0028] Step 1.6: Constructing the scale-adaptive feature extraction module. The present application selects the backbone network of the classic image super-resolution reconstruction model RCAN after removing the upsampling module as the basic structure of the feature extraction module. The present application inserts the core component invention scale-adaptive feature enhancement module mentioned in step 1.5 into the basic structure of the feature extraction module. Specifically, a scale-adaptive feature enhancement module is inserted after each residual group in the basic structure to complete the construction of the scale-adaptive feature extraction module.

[0029] Step 1.7: Calculate the coordinate mapping of the pixel position in the low-resolution image to the high-resolution space. For the coordinate tensor L x and L y of the pixel in the low-resolution image in the vertical and horizontal directions, the coordinates in the high-resolution space can be expressed as

[0030]

[0031] Step 1.8: Determine the offset for sampling according to the scale factor. After completing the coordinate mapping, the features in the s×s neighborhood of the center [P x , P y ] in the high-resolution space need to be resampled. In order to generate the sampling coordinates, the offset sets in the vertical and horizontal directions are defined as

[0032]

[0033] Step 1.9: Determine the sampling coordinate set according to the mapping coordinates and the offset. According to the mapping coordinates [Px ,P y ] and offset set, the sampling coordinate set can be defined as

[0034]

[0035] Step 1.10: Resample the scale-dependent features according to the sampling coordinate set and get multiple groups of candidate features. First, use the core component invention scale-adaptive feature extraction module mentioned in step 1.6 to extract scale-dependent features F s from the input high-resolution image. Then, the invention resamples the scale-dependent features to get the resampled candidate feature set, which can be represented as

[0036] Ψ={F|F=f resample (F s ,H x ,H y ),H x ∈Φ x ,H y ∈Φ y}

[0037] Where f resample (·) represents the candidate feature resampling function.

[0038] Step 1.11: Generate adaptive weights for each candidate feature in the candidate feature set. The invention selects a multi-layer perceptron as the weight generator to adaptively generate corresponding weights for each candidate feature. The generated adaptive weight set can be represented as

[0039] Ω={W|W=f weight (δ x ,δ y ,s),δ x ∈Δ x ,δ y ∈Δ y}

[0040] Where f weight (·) represents a weight generator based on a multi-layer perceptron.

[0041] Step 1.12: Generate a down-sampled low-resolution image using adaptive weights. At this point, according to the candidate features and their corresponding adaptive sampling weights, the down-sampled low-resolution image can be calculated as

[0042]

[0043] Where f fusion (·) represents a channel feature fusion module used to generate a 3-channel RGB image.

[0044] Step 2: Constructing the loss function of the image downsampling model. The image downsampling method provided in this invention considers known interpolation-based image super-resolution reconstruction methods, aiming to generate low-resolution images suitable for the corresponding image super-resolution reconstruction methods. The training objective of the image downsampling model constructed in this invention is to minimize the distortion of the reconstructed image obtained by reconstructing the downsampled low-resolution image using the corresponding image super-resolution reconstruction method relative to the original image. Therefore, the loss function can be defined as follows:

[0045]

[0046] Among them, I HR For the input high-resolution image, I LR For the downsampled low-resolution image, f SR (·) indicates the known interpolation-based image super-resolution reconstruction method that needs to be adapted, f ref (·) represents the reference image generation function, which is used to supervise the content consistency during the downsampling process. λ represents the tradeoff coefficient used to balance the quality of the reconstructed image and the quality of the downsampled image. Represents high-resolution image I HR Belonging set, This indicates an averaging operation.

[0047] Step 3: Image downsampling model training and optimization. This invention applies gradient descent to the loss function in Step 2 for overall model training and optimization. During training, the scale factor is randomly sampled from the value range Λ, ensuring the model remains sensitive to the scale factor, thereby constructing the image downsampling model.

[0048] Secondly, the present invention provides an image downsampling system, comprising: an image downsampling device, a transmission device, and an image super-resolution reconstruction device. The image downsampling device is used to receive a user downsampling scale factor instruction, and to perform the image downsampling method described in any one of the first aspects on a high-resolution image with a specified scale factor to obtain a low-resolution image. The transmission device is used to send the low-resolution image downsampled by the image downsampling device to the image super-resolution reconstruction device. The image super-resolution reconstruction device is used to receive the low-resolution image sent by the transmission device, and to perform super-resolution reconstruction on the low-resolution image using a specified interpolation-based image super-resolution reconstruction method to obtain a reconstructed high-resolution image. Bicubic interpolation, Lanczos interpolation, and Bilinear interpolation can be selected.

[0049] Compared with the related art, the image downsampling method and system provided by the application considers the known image super-resolution reconstruction method adopted by the image super-resolution reconstruction device, receives a downsampling scale factor instruction, and downsamples the original high-resolution image to obtain a low-resolution image, thereby solving the problem that the low-resolution image obtained by the image downsampling method based on interpolation in the related art cannot be fully applied to the existing image super-resolution reconstruction method. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 Model details of the technology described in the application.

[0051] Figure 2 Overall flowchart of the technology described in the application. DETAILED DESCRIPTION

[0052] The overall flowchart of the application is shown in Figure 2 The specific implementation steps are as follows:

[0053] Step 1: overall image downsampling model construction. The image downsampling model construction steps are as follows,

[0054] Step 1.1: calculate the coordinate maps in the high-resolution space and the low-resolution space, respectively. First, define the value range of the scale factor that controls the size of the low-resolution image after downsampling in the image downsampling process, Λ, take The scale factor s in the image downsampling process controls the size of the low-resolution image after downsampling. According to the high-resolution image size [H, W] and the scale factor s, the target low-resolution image size can be expressed as

[0055]

[0056] where represents rounding down. The coordinate maps in the high-resolution space and the low-resolution space can be normalized as

[0057]

[0058] where the values in the coordinate maps in the high-resolution space and the low-resolution space are normalized to the interval [-1, 1], avoiding excessively large coordinate values, C HR and C LR are the coordinate maps in the high-resolution space and the low-resolution space, respectively, and the sizes of the two coordinate maps are [H, W] and [h, w], respectively, and HR and LR represent high resolution and low resolution, respectively; i and j represent the value index of the coordinate map in the vertical direction (H / h direction) and the horizontal direction (W / w direction), respectively. For example, for C HRFor example, the length of the vertical direction (H direction) is H, so i takes values of 0, 1, 2, …, H-1, the length of the horizontal direction (W direction) is W, so j takes values of 0, 1, 2, …, W-1, and thus (i, j) takes values of the combination of i and j; denotes an integer set; denotes an integer coordinate set in the HxW range.

[0059] Step 1.2: Adjust the coordinate maps in the high-resolution space and the low-resolution space to the same size. The coordinate maps in the high-resolution space and the low-resolution space reflect the mapping relationship of pixel coordinates in the image downsampling process, but the sizes of the coordinate maps in the high-resolution space and the low-resolution space are different. In order to utilize this pixel coordinate mapping relationship, it is necessary to adjust the coordinate maps in the high-resolution space and the low-resolution space to the same size. The present application considers that the size of the feature map needs to be consistent with the size of the input high-resolution image in the feature extraction process, and therefore selects to use the nearest neighbor sampling method to upsample the coordinate map in the low-resolution space to enlarge it to the size of the coordinate map in the high-resolution space. The sampled coordinate map can be expressed as

[0060]

[0061] where f nearest (·) denotes the nearest neighbor sampling method selected by the present application, which aims to enlarge the coordinate map in the low-resolution space to the target size without destroying the value of the coordinate map in the low-resolution space.

[0062] Step 1.3: Calculate the feature scaling factor for adjusting the feature distribution. Considering that the relative distance between the coordinate maps in the high-resolution space and the low-resolution space reflects the mapping relationship of pixels in the high-resolution image and the low-resolution image, the present application takes the relative distance between the coordinate maps in the high-resolution space and the low-resolution space as the feature scaling factor for adjusting the feature distribution, which is used to scale the feature in numerical value. The feature scaling factor can be expressed as

[0063]

[0064] where f subtraction (·) denotes the pixel-by-pixel subtraction operation between tensors, which is used to calculate the relative distance between the coordinate maps, and has the advantages of simplicity and efficiency.

[0065] Step 1.4: Calculate the feature offset parameter for adjusting the feature distribution. In order to adapt the feature extraction process to different scale factors, the present application calculates the feature offset parameter from the perspective of the size relationship between the high-resolution image and the low-resolution image, which is used to continuously adjust the feature distribution in the feature extraction process. First, the scale vectors in the vertical and horizontal directions can be defined as

[0066]

[0067] Considering the tensor shape of the deep features in the feature extraction process [middle_ch, H, W], in order to make the feature shift parameters have the same size as the deep features, the present application expands the scale vector into a scale tensor S, which can be expressed as [2, H, W] in shape. Next, the feature shift parameters are calculated according to the scale tensor S. The present application selects a multilayer perceptron (MLP) to calculate the feature shift parameters, and the dynamically generated feature shift parameters can be expressed as

[0068]

[0069] wherein MLP refers to a layer of multilayer perceptron selected by the present application, which aims to convert the input feature with a size of 2 into an output feature with a size of middle_channel. In this embodiment, middle_channel is taken as 64.

[0070] Step 1.5: Constructing the scale adaptive feature enhancement module. The present application combines the core component invention feature scaling coefficient and feature shift parameter mentioned in steps 1.3 and 1.4 with a convolutional layer to construct the scale adaptive feature enhancement module. For the input feature F in , the output feature F out obtained after implementing the scale adaptive feature enhancement can be expressed as

[0071]

[0072] wherein f conv (·) represents a convolutional layer with a kernel size of 3x3. Based on residual learning, the present application adjusts the distribution of the deep features in the feature extraction process to adapt to different scale factors.

[0073] Step 1.6: Constructing the scale adaptive feature extraction module. The present application selects the backbone network of the classical image super-resolution reconstruction model RCAN after removing the upsampling module as the basic structure of the feature extraction module, which is composed of n residual groups. The RCAN basic structure and residual groups used in this embodiment are not within the scope of this patent and are not described. In this embodiment, n is taken as 5. The present application inserts the core component invention scale adaptive feature enhancement module mentioned in step 1.5 into the basic structure of the feature extraction module, specifically, an scale adaptive feature enhancement module is inserted after each residual group, completing the construction of the scale adaptive feature extraction module.

[0074] Step 1.7: Calculate the coordinate mapping of pixel position in low resolution image to high resolution space. For the coordinates of a pixel in low resolution image in vertical and horizontal direction, L x and L y , the coordinates in high resolution space can be represented as

[0075]

[0076] Step 1.8: Determine the offsets for sampling according to the scale factor. After the coordinate mapping is completed, the features in the s x s neighborhood centered at the coordinates in high resolution space [P x , P y ] need to be resampled. In order to generate the sampling coordinates, the offset sets in vertical and horizontal direction are defined as

[0077]

[0078] Step 1.9: Determine the sampling coordinate set according to the mapping coordinates and the offset sets. According to the mapping coordinates [P x , P y ] and the offset sets, the sampling coordinate set can be defined as

[0079]

[0080] Step 1.10: Resample the scale-dependent features according to the sampling coordinate set and obtain multiple groups of candidate features. First, use the core component invention scale adaptive feature extraction module mentioned in step 1.6 to extract scale-dependent features F s from the input high resolution image. Then, the grid sampling method is selected to resample the scale-dependent features, and the resampled candidate feature set can be represented as

[0081] Ψ={F|F=f grid (F s ,H x ,H y ),H x ∈Φ x ,H y ∈Φ y}

[0082] where f grid (·) represents the grid sampling function.

[0083] Step 1.11: Generate adaptive weights for each candidate feature in the candidate feature set. The invention selects a multi-layer perceptron as the weight generator to adaptively generate corresponding weights for each candidate feature. The generated adaptive weight set can be represented as

[0084] Ω={W|W=f weight (δ x ,δ y ,s),δ x ∈Δ x ,δ y ∈Δ y}

[0085] where f weight (·) represents a weight generator based on multi-layer perception. In the present embodiment, the input dimension of the weight generator is 3, the output dimension is 1, and the hidden layer structure is [16, 16, 16, 16, 16].

[0086] Step 1.12: Generate the down-sampled low-resolution image using adaptive weights. So far, according to the candidate features and their corresponding adaptive sampling weights, the down-sampled low-resolution image can be calculated as

[0087]

[0088] where f fusion (·) represents a channel feature fusion module composed of two 3x3 convolutions, which is used to generate a 3-channel RGB image.

[0089] Step 2: Construct the loss function of the image down-sampling model. The image down-sampling method provided by the present application considers the known interpolation-based image super-resolution reconstruction method, aiming to generate a low-resolution image suitable for the corresponding image super-resolution reconstruction method. The training target of the image down-sampling model constructed by the present application is to minimize the distortion of the reconstructed image obtained by the corresponding image super-resolution reconstruction method after the down-sampled low-resolution image, relative to the original image, so that the loss function can be defined as

[0090]

[0091] where I HR is the input high-resolution image, I LR is the down-sampled low-resolution image, f SR (·) represents a known interpolation-based image super-resolution reconstruction method that needs to be adapted, f ref (·) is a reference image generation function that is used to supervise the content consistency in the down-sampling process, and λ represents a weighting coefficient used to balance the quality of the reconstructed image and the quality of the down-sampled image. In the present embodiment, λ is taken as 0.1, f sR (·) is set as a bicubic interpolation function, and f ref (·) is set as a bicubic down-sampling function.

[0092] Step 3: Image down-sampling model training optimization. The gradient descent method is applied to the loss function in step 2 for the training optimization of the overall model. During the training, the scale factor is from the value interval Random sampling, so that the model is always sensitive to the scale factor, and then build an image down-sampling model.

[0093] The specific embodiments described herein are merely illustrative of the spirit of the present application and various modifications or supplements or similar ways of substitution can be made to the described specific embodiments by those skilled in the art without departing from the spirit of the present application or exceeding the scope defined by the appended claims.

Claims

1. An image downsampling method, characterized in that: Step 1: constructing an image downsampling model, specifically comprising: Step 1.1: calculating the coordinate maps in the high-resolution space and the low-resolution space respectively, and normalizing them; Step 1.2: adjusting the coordinate maps in the high-resolution space and the low-resolution space to the same size; Step 1.3: calculating the feature scaling coefficients for adjusting the feature distribution; Step 1.4: calculating the feature offset parameters for adjusting the feature distribution; Step 1.5: constructing a scale-adaptive feature enhancement module; Step 1.6: constructing a scale-adaptive feature extraction module; Step 1.7: calculating the coordinate mapping of the low-resolution image pixel position to the high-resolution space; Step 1.8: determining the offset for sampling according to the scale factor; Step 1.9: determining the sampling coordinate set according to the mapping coordinates and the offset; Step 1.10: resampling the scale-dependent features according to the sampling coordinate set and obtaining multiple groups of candidate features; Step 1.11: generating adaptive weights for each candidate feature in the candidate feature set, specifically: using a multi-layer perceptron as the weight generator to adaptively generate corresponding weights for each candidate feature; the generated adaptive weight set is represented as Ω = { W | W = f weight (δ x , δ y , s), δ x ∈ Δ x , δ y ∈ Δ y} where f weight (·) represents a weight generator based on a multi-layer perceptron, δ x , δ y is a sampling offset, Δ x , Δ y is a set of sampling offsets, and s is a scale factor; Step 1.12: generating the final low-resolution image after downsampling using the adaptive weights, specifically: according to the candidate features and their corresponding adaptive sampling weights, the low-resolution image after downsampling can be calculated as where f fusion (·) represents a channel feature fusion module for generating a 3-channel RGB image, Ψ is a set of sampled candidate features, and F is a candidate feature; Step 2: constructing the loss function of the image downsampling model; Step 3: training and optimizing the image downsampling model.

2. The image downsampling method of claim 1, wherein: The specific calculation process of the coordinate maps in the high-resolution space and the low-resolution space in step 1.1 is as follows: First, define the value range of the scale factor Λ that controls the size of the low-resolution image after downsampling in the image downsampling process, then sample the scale factor from Λ, i.e. s∈Λ, according to the high-resolution image size [H,W] and the scale factor s, the target low-resolution image size is represented as wherein denotes rounding down, the coordinate maps under the high-resolution space and the low-resolution space can be normalized as where the values in the coordinate map in the high-resolution space and the low-resolution space are normalized to the interval [-1, 1], avoiding excessively large coordinate values, C HR and C LR are the coordinate maps in the high-resolution space and the low-resolution space, respectively, i and j represent the value indexes of the coordinate map in the vertical direction and the horizontal direction, respectively, H is the length of the high-resolution image in the vertical direction, and W is the length of the high-resolution image in the horizontal direction, represents the integer coordinate set in the HxW range; In step 1.2, the coordinate map in the low-resolution space is upsampled to the size of the coordinate map in the high-resolution space.

3. The image downsampling method of claim 1, wherein: The specific process of calculating the feature scaling coefficients for adjusting the feature distribution in step 1.3 is as follows: The relative distance between the coordinate maps in the high-resolution space and the low-resolution space is used as the feature scaling coefficient for adjusting the feature distribution, which is used to scale the features numerically, and the feature scaling coefficient is represented as where f distance (·) represents a relative distance computation function, C HR represents a coordinate map of a high-resolution space, represents a coordinate map after sampling.

4. The image downsampling method of claim 1, wherein, The specific process of calculating the feature offset parameters for adjusting the feature distribution in step 1.4 is as follows: First, the scale vectors in the vertical and horizontal directions are defined as where H is the length of the high-resolution image in the vertical direction, and W is the length of the high-resolution image in the horizontal direction. Considering the tensor shape of the deep feature in the feature extraction process [middle_channel, H, W], in order to make the feature offset parameter have the same size as the deep feature, the scale vector is expanded to a scale tensor S, which has a shape of [2, H, W]. Next, the feature offset parameter is calculated according to the scale tensor S. The multi-layer perception is selected to calculate the feature offset parameter. The dynamically generated feature offset parameter can be represented as wherein is a selected multi-layer perceptron for converting input features of size 2 to output features of size middle_channel.

5. The image downsampling method of claim 1, wherein, The specific process of constructing the scale adaptive feature enhancement module in step 1.5 is as follows: The scale adaptive feature enhancement module is constructed by combining the feature scaling coefficient R and the feature offset parameter B in step 1.3 and step 1.4 with the convolution layer, and the output feature F in obtained after implementing the scale adaptive feature enhancement is represented as out ​ where f conv (·) denotes a convolution operation, based on residual learning, by adjusting the distribution of deep features in the feature extraction process to adapt to different scale factors.

6. The image downsampling method of claim 1, wherein, The specific process of calculating the coordinate mapping of the low-resolution image pixel position to the high-resolution space in step 1.7 is as follows: For a pixel in a low resolution image, the coordinate tensor L x and L y in the vertical and horizontal directions, the coordinate in the high resolution space is represented as where s denotes a scale factor, [P x ,P y ] denotes a coordinate in a high-resolution space.

7. The image downsampling method of claim 1, wherein, The specific process of resampling the scale-related features according to the sampling coordinate set and obtaining multiple groups of candidate features in step 1.10 is as follows: First, the scale-dependent features F are extracted from the input high-resolution image using the scale-adaptive feature extraction module in step 1.6 s Then, the scale-dependent features are re-sampled to obtain a set of re-sampled candidate features, denoted as Ψ = { F | F = f resample (F s ,H x ,H y ), H x ∈ Φ x ,H y ∈ Φ y} where f resample (·) denotes a candidate feature resampling function, Φ x , Φ y is a set of sampling coordinates.

8. The image downsampling method of claim 1, wherein: The specific process of constructing the loss function of the image down-sampling model in step 2 is as follows: The reconstructed image obtained by making the down-sampled low-resolution image pass through the corresponding image super-resolution reconstruction method has the minimum distortion relative to the original image, so the loss function is defined as where I HR is the input high-resolution image, I LR is the down-sampled low-resolution image, f SR (·) denotes a known interpolation-based image super-resolution reconstruction method that needs to be adapted, f ref (·) is a reference image generation function that is used to supervise the content consistency in the down-sampling process, and λ denotes a trade-off coefficient used to balance the reconstructed image quality and the down-sampled image quality.

9. An image down-sampling system comprising: An image down-sampling device, a transmission device, and an image super-resolution reconstruction device; The image down-sampling device is configured to receive a user down-sampling scale factor instruction, perform the image down-sampling method of any one of claims 1-8 on a high-resolution image with a specified scale factor, and obtain a low-resolution image. The transmission device is configured to send the low-resolution image down-sampled by the image down-sampling device to the image super-resolution reconstruction device. The image super-resolution reconstruction device is configured to receive the low-resolution image sent by the transmission device, perform super-resolution reconstruction on the low-resolution image with a specified interpolation-based image super-resolution reconstruction method, and obtain a reconstructed high-resolution image. The specific process of constructing the scale adaptive feature enhancement module in step 1.5 is as follows:

Images