Image decompression method and device

Description

1 / 45 METHOD AND APPARATUS FOR IMAGE DECOMPRESSION TECHNICAL AREA

[0001] This application relates to image processing technologies and, in particular, to a method and apparatus for image decompression. BACKGROUND

[0002] As convolutional neural networks (CNNs) outperform conventional algorithms in computer vision tasks such as image recognition and object detection, more researchers are beginning to explore deep learning-based image / video compression methods. Some researchers design an end-to-end deep learning image / video compression algorithm. For example, modules such as an encoding network, an entropy estimation network, an entropy encoding mechanism, an entropy decoding mechanism, and a decoding network are integrated as a whole, or modules such as a prediction module and a residual compression module are integrated as a whole. The encoding network and the decoding network may also be called the transformation module and inverse transformation module, and typically include a convolutional layer and a nonlinear transformation unit.

[0003] The Artificial Intelligence (AI) of the Joint Expert Group on Photography (JPEG) is the first international standard for AI-based image compression. In its typical structure, an encoder performs RGBYUV color space conversion into images to decompose the images into Y and UV components, and then performs encoding and transmission separately; a decoder separately restores the Y and UV components, and then performs YUV-RGB color space conversion to obtain RGB images. In the former process, there is some specific information exchange between the Y and UV components for better effect. Petition 870250108800, dated 11 / 27 / 2025, page 8 / 64 2 / 45 compression. The exchange of information between different components is particularly important for performance optimization. SUMMARY

[0004] This application provides a method and an image decompression device that can improve performance.

[0005] According to a first aspect, this application provides a method for image decompression, including: obtaining a first feature tensor, wherein the first feature tensor corresponds to a first component of a reconstructed image; obtaining a second feature tensor, wherein the second feature tensor corresponds to a second component of the reconstructed image; performing concatenation and convolution on the first feature tensor and the second feature tensor to obtain a third feature tensor; obtaining a fourth feature tensor based on the third feature tensor and the second feature tensor; and obtaining the reconstructed image based on the first feature tensor and the fourth feature tensor.

[0006] In embodiments of this application, a structure for merging a Y component and a UV component is optimized. The fourth feature tensor (corresponding to a UV component of the reconstructed image) is first obtained based on the third feature tensor (i.e., a tensor output after concatenating the first feature tensor (corresponding to a Y component of the reconstructed image) and the second feature tensor (corresponding to the UV component of the reconstructed image) and performing the convolution) and the second feature tensor, and then the reconstructed image is obtained based on the first feature tensor and the fourth feature tensor. There are three cases for the number of channels of the fourth feature tensor: less than, greater than, or equal to the sum of the number of channels of the first feature tensor. Petition 870250108800, dated 11 / 27 / 2025, p. 9 / 64 3 / 45 characteristic and the second characteristic tensor. In the case less than, the number of channels of the fourth characteristic tensor can be reduced to save computing power and implement a lightweight system. In the case greater than or equal to, even if the number of channels of the fourth characteristic tensor is not reduced, the fourth characteristic vector can be obtained by performing an operation on the third characteristic tensor and only on the second characteristic tensor, so that performance can be significantly improved.

[0007] Optionally, the first component can be a Y component of an image represented using a YUV color space, and the second component can be a UV component of the image represented using the YUV color space.

[0008] It should be noted that the first component and the second component may alternatively be other components based on a way of representing the image. This is not specifically limited in modalities of this application.

[0009] As described above, in a JPEG AI structure, an encoder side decomposes the images into Y components and UV components and then performs separate encoding and transmission; and a decoder side separately decodes and restores the Y and UV components. It can be learned that the Y component (i.e., the first component) and the UV component (i.e., the second component) are processed independently in two separate paths. In a path for processing the UV component on the decoder side, the Y component and the UV component are merged.

[0010] Concatenating (Concat) the first feature tensor and the second feature tensor can be done by performing channel concatenation on them. For example, if a scale of the first feature tensor is [128, H, W] and a scale of the second feature tensor is [64, H, W], Petition 870250108800, dated 11 / 27 / 2025, page 10 / 64 4 / 45 A scale of a concatenated characteristic tensor will be [192, H, W]. That is, a width and a height remain unchanged, and a channel number is a sum of its channel numbers. Then, convolution (Conv) is performed on the concatenated characteristic tensor. Convolution can increase / decrease / not change the channel number. Correspondingly, a channel number of the third output characteristic tensor is increased / decreased / not changed. For example, a scale of the third characteristic tensor will be [64, H, W], [128, H, W] or [192, H, W]. The 1st element in the previous scale represents a channel number of the characteristic tensor.

[0011] Optionally, the third feature tensor and the second feature tensor can be concatenated to obtain the fourth feature tensor.

[0012] Concatenation is channel concatenation. That is, channel concatenation is performed on the third feature tensor and the second feature tensor.

[0013] It should be noted that the number of channels of the fourth feature tensor is not specifically limited in the embodiments of this application. If the number of channels of the fourth feature tensor is less than the sum of the number of channels of the first feature tensor and the second feature tensor, the number of channels can be reduced to save computing power and implement a lighter structure. Even if the number of channels of the fourth feature tensor is greater than the sum of the number of channels of the first feature tensor and the second feature tensor, after convolution, element-by-element addition is transformed into concatenation, and element-by-element addition of an output from the convolution and a merged feature tensor of the Y component and the UV component is transformed into concatenation of the output and only the UV component, so that the Petition 870250108800, dated 11 / 27 / 2025, page 11 / 64 5 / 45 performance can be significantly improved.

[0014] Optionally, channel extraction can be performed on the second feature tensor to obtain a fifth feature tensor; and the third feature tensor and the fifth feature tensor can be concatenated to obtain the fourth feature tensor.

[0015] Channel reduction can be performed on the second feature tensor to obtain the fifth feature tensor, and then the third feature tensor and the fifth feature tensor can be concatenated.

[0016] It should be noted that the number of channels of the fourth feature tensor is not specifically limited in the embodiments of this application. If the number of channels of the fourth feature tensor is less than the sum of the number of channels of the first feature tensor and the second feature tensor, the number of channels can be reduced to save computing power and implement a lighter structure. Even if the number of channels of the fourth feature tensor is greater than the sum of the number of channels of the first feature tensor and the second feature tensor, after convolution, element-by-element addition is transformed into concatenation, and element-by-element addition of an output from the convolution and a merged feature tensor of the Y component and the UV component is transformed into concatenation of the output and only the UV component, so that performance can be significantly improved.

[0017] Optionally, the third feature tensor and the second feature tensor can be added to obtain the fourth feature tensor.

[0018] If the number of channels in the third feature tensor is equal to that of the second feature tensor, they can be added in an element-by-element manner. If the number of channels in the third feature tensor Petition 870250108800, dated 11 / 27 / 2025, page 12 / 64 If 6 / 45 is not equal to that of the second feature tensor, they can be added by adding the channels of the feature tensor with the fewest channels and some channels of the feature tensor with the most channels, in an element-by-element manner. For example, the third feature tensor has 64 channels and the second feature tensor has 32 channels. The 32 channels of the second feature tensor and the first 32 channels of the third feature tensor can be added in an element-by-element manner; the 32 channels of the second feature tensor and the last 32 channels of the third feature tensor can be added in an element-by-element manner; or similarly. One method of addition, in this case, is not specifically limited to the embodiments of this application.

[0019] It should be noted that the number of channels of the fourth feature tensor is not specifically limited in the embodiments of this application. If the number of channels of the fourth feature tensor is less than the sum of the number of channels of the first feature tensor and the second feature tensor, the number of channels can be reduced to save computing power and implement a lighter structure. Even if the number of channels of the fourth feature tensor is greater than the sum of the number of channels of the first feature tensor and the second feature tensor, after convolution, the element-by-element addition of a convolution output and a merged feature tensor of the Y component and the UV component is changed to element-by-element addition of the output and only the UV component, so that performance can be significantly improved.

[0020] Optionally, channel extraction can be performed on the second feature tensor to obtain a fifth feature tensor; and the third feature tensor and the fifth feature tensor can be added to obtain Petition 870250108800, dated 11 / 27 / 2025, page 13 / 64 7 / 45 the fourth characteristic tensioner.

[0021] If the number of channels in the third feature tensor is less than that in the second feature tensor, channel reduction can be performed on the second feature tensor to obtain the fifth feature tensor, so that the number of channels in the fifth feature tensor is equal to that of the second feature tensor; and then the third feature tensor and the fifth feature tensor can be added in an element-by-element manner.

[0022] It should be noted that the number of channels of the fourth feature tensor is not specifically limited in the embodiments of this application. In the previous case, if the number of channels of the fourth feature tensor is less than the sum of the number of channels of the first feature tensor and the second feature tensor, the number of channels can be reduced to save computing power and implement a lighter structure. Even if the number of channels of the fourth feature tensor is greater than the sum of the number of channels of the first feature tensor and the second feature tensor, after convolution, the element-by-element addition of a convolution output and a merged feature tensor of the Y component and the UV component is changed to element-by-element addition of the output and only the UV component, so that performance can be significantly improved.

[0023] The UV component of the reconstructed image can be obtained through some processing (see the following embodiments) on the fourth feature tensor obtained through fusion and then combined with the Y component of the reconstructed image to form the complete reconstructed image.

[0024] According to a second aspect, this application provides an image decompression apparatus, including: a module of Petition 870250108800, dated 11 / 27 / 2025, page 14 / 64 8 / 45 acquisition, configured to obtain a first feature tensor, where the first feature tensor corresponds to a first component of a reconstructed image, and obtain a second feature tensor, where the second feature tensor corresponds to a second component of the reconstructed image; a processing module, configured to perform concatenation and convolution on the first feature tensor and the second feature tensor to obtain a third feature tensor, and obtain a fourth feature tensor based on the third feature tensor and the second feature tensor; and a reconstruction module, configured to obtain the reconstructed image based on the first feature tensor and the fourth feature tensor.

[0025] In one possible implementation, the processing module is specifically configured to concatenate the third feature tensor and the second feature tensor to obtain the fourth feature tensor.

[0026] In one possible implementation, the processing module is specifically configured to add the third feature tensor and the second feature tensor to obtain the fourth feature tensor.

[0027] In one possible implementation, the processing module is specifically configured to perform channel extraction on the second feature tensor to obtain a fifth feature tensor; and to concatenate the third feature tensor and the fifth feature tensor to obtain the fourth feature tensor.

[0028] In one possible implementation, the processing module is specifically configured to perform channel extraction on the second feature tensor to obtain a fifth feature tensor; and add the third feature tensor and the fifth feature tensor to obtain Petition 870250108800, dated 11 / 27 / 2025, page 15 / 64 9 / 45 the fourth characteristic tensioner.

[0029] In one possible implementation, the first component is a Y component and the second component is a UV component.

[0030] According to a third aspect, this application provides an electronic device, including: one or more processors; and a memory, configured to store one or more programs. When the one or more programs are executed by the one or more processors, one or more processors are enabled to implement the method according to any of the implementations of the first aspect.

[0031] According to a fourth aspect, this application provides a computer-readable storage medium, including a computer program. When the computer program is run on a computer, the computer is enabled to perform the method according to any of the implementations of the first aspect.

[0032] According to a fifth aspect, this application provides a computer program. When the computer program is executed by a computer, the computer is configured to perform the method according to any of the implementations of the first aspect.

[0033] According to a sixth aspect, this application further provides a computer program product. The computer program product includes computer program code. When the computer program code is run on a computer, the computer is enabled to perform the operations and / or processing performed by the electronic device in any of the previous method embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is an example diagram of an end-to-end deep learning image coding structure; Figure 2 is an example diagram of an end-to-end deep learning video coding structure; Petition 870250108800, dated 11 / 27 / 2025, page 16 / 64 10 / 45 Figure 3 is an example diagram of an application scenario according to one modality of this request; Figure 4 is an example diagram of an application scenario according to one modality of this request; Figure 5a is a diagram of a typical structure in a JPEG AI format; Figure 5b is a diagram of a structure in the thick-lined box in Figure 5a; Figure 6 is a flowchart of a process 600 of an image decompression method according to an embodiment of this application; Figure 7a is a diagram of a fusion structure; Figure 7b is a diagram of a fusion structure; Figure 7c is a diagram of a fusion structure; Figure 7d is a diagram of a fusion structure; Figure 8 is a diagram of a JPEG AI structure using a merging structure; Figure 9 is a diagram of a JPEG AI structure using a merging structure; Figure 10 is a diagram of a JPEG AI structure using a merging structure; Figure 11 is a diagram of a JPEG AI structure using a fusion structure; and Figure 12 is a diagram of a structure of a 1200 image decompression apparatus according to an embodiment of this application. DESCRIPTION OF THE MODALITIES

[0035] To make the objectives, technical solutions and advantages of this application clearer, the following describes clearly and completely the technical solutions contained in this application, with reference to the attached drawings. It is evident that the embodiments described are only some, and not all, of the embodiments of this application. All other embodiments obtained by a technician in Petition 870250108800, dated 11 / 27 / 2025, page 17 / 64 11 / 45 subject matter, based on the modalities of this application, without creative effort, will be covered by the scope of protection of this application.

[0036] In the descriptive report, embodiments, claims and accompanying drawings of this application, terms such as first and second are used merely for differentiation and description, but should not be understood as an indication or implication of relative importance or an indication or implication of a sequence. Furthermore, the terms include, have and any variant thereof are intended to encompass non-exclusive inclusion, for example, including a series of steps or units. A method, system, product or device is not necessarily limited to the steps or units expressly listed, but may include other steps or units not expressly listed or inherent to such process, method, product or device.

[0037] It should be understood that, in this request, at least one (item) means one or more, and a plurality of means two or more. The term and / or is used to describe an association relationship between associated objects and represents the possibility of three relationships existing. For example, A and / or B can represent the following three cases: only A exists, only B exists, and both A and B exist, where A and B can be singular or plural. The character / generally indicates an OR relationship between associated objects. The expression at least one of the following items (pieces) or a similar expression means any combination of these items, including a single item (piece) or any combination of a plurality of items (pieces). For example, at least one of a, b, or c can indicate a, b, c, aeb, aec, bec, or a, bec, where a, bec can be singular or plural.

[0038] Figure 1 is an example diagram of an end-to-end deep learning image coding structure. Petition 870250108800, dated 11 / 27 / 2025, page 18 / 64 12 / 45 As shown in Figure 1, the image encoding structure includes an encoder side and a decoder side. The encoder side includes: an encoding network (Encoder), a quantization module, a hyperprior encoding network (Hyper Encoder), a hyperprior decoding network, entropy estimation, and an entropy encoding mechanism (common entropy encoding mechanisms include an arithmetic encoder (AE), an asymmetric numeral system (ANS), and the like). The decoder side includes: a hyperprior decoding network (Hyper Decoder), a decoding network (Decoder), entropy estimation, and an entropy decoding mechanism (common entropy decoding mechanisms include an arithmetic decoder (AD), an ANS, and the like).

[0039] On the encoder side, an original image is transformed from an image domain to a feature domain after being processed by the encoding network. A transformed image feature is encoded into a bitstream to be transmitted or stored after being processed by the quantization module and the entropy encoding mechanism. On the decoder side, a bitstream is decoded into an image feature after being processed by the entropy decoding mechanism. The image feature is transformed from the feature domain to the image domain after being processed by the decoding network, to obtain a reconstructed image. The entropy estimation network estimates and obtains an estimated probability value of each feature element based on the image feature. The probability value is used for processing by the entropy encoding mechanism and the entropy decoding mechanism.

[0040] In the modalities, both the encoding network (Encoder) Petition 870250108800, dated 11 / 27 / 2025, page 19 / 64 13 / 45 as for the decoding network (Decoder) they have a non-linear transformation unit.

[0041] Figure 2 is an example diagram of an end-to-end deep learning video coding structure. As shown in Figure 2, the video encoding structure includes a prediction model module and a residual compression module.

[0042] The prediction module predicts a current frame using a reconstructed image of a previous frame to obtain a predicted image. The residual compression module compresses a residue between an original image and a predicted image of the current frame, then decompresses the compressed residue to obtain a reconstructed residue and sums the reconstructed residue and the predicted image to obtain a reconstructed image of the current frame. Both an encoding subnetwork and a decoding subnetwork in the prediction module and the residual compression module have a nonlinear transformation unit.

[0043] In the modalities, both the predict model module and the residual compression module have the non-linear transformation unit.

[0044] Figure 3 is an example diagram of an application scenario according to one embodiment of this request. As shown in Figure 3, the application scenario can be a service related to capturing, storing, or transmitting images / video on a terminal, a cloud server, or video surveillance, for example, photography / video recording by a terminal, an album, a cloud album, or video surveillance.

[0045] Encoder side: A camera captures an image / video. An image / video encoding network with artificial intelligence (AI) performs feature extraction on the image / video to obtain a low-redundancy image feature and then performs entropy encoding based on the feature. Petition 870250108800, dated 11 / 27 / 2025, page 20 / 64 14 / 45 image to obtain a bitstream / image file.

[0046] Decoder side: when the image / video needs to be played back, an AI image / video decoding network performs entropy decoding on the bitstream / image file to obtain an image feature, and then performs reverse feature extraction on the image feature to obtain a reconstructed image / video.

[0047] A storage / transmission module stores (e.g., by taking photos from a terminal, video surveillance, or a cloud server) or transmits (e.g., a cloud service or live streaming technology) the bitstream / image file obtained through compression to different services.

[0048] Figure 4 is an example diagram of an application scenario according to one embodiment of this application. As shown in Figure 4, the application scenario can be a service related to capturing, storing, or transmitting images / video in the cloud or video surveillance, for example, a cloud album, video surveillance, or live streaming.

[0049] Encoder side: A local side obtains an image / video, encodes the image / video to obtain a compressed image / video, and then sends the compressed image / video to a cloud side. The cloud side decodes the compressed image / video to obtain the image / video and then performs AI encoding on the image / video to obtain a bitstream / image file and stores the bitstream / image file.

[0050] Decoder side: When the local side needs to obtain the image / video from the cloud side, the cloud side performs AI decoding on the bitstream / image file to obtain the image / video, encodes the image / video to obtain a compressed image / video, and sends the compressed image / video to the local side. The local side then decodes the image / video. Petition 870250108800, dated 11 / 27 / 2025, page 21 / 64 15 / 45 compressed to obtain the image / video. For a cloud-side structure and the use of each module, refer to the structure and use of each module in Figure 3. Details are not described in this document in the embodiments of this application.

[0051] Figure 5a is a diagram of a typical structure in a JPEG AI standard. As shown in Figure 5a, an encoder side performs RGB-YUV color space conversion to decompose images into Y and UV components, and then performs separate encoding and transmission; and a decoder side separately restores the Y and UV components, and then performs YUV-RGB color space conversion to obtain RGB images. On both the encoder and decoder sides, there is a specific exchange of information between the Y and UV components for better compression effect.

[0052] Figure 5b is a diagram of a structure in the thick line box of Figure 5a. As shown in Figure 5b, a scale of yY1 is [128, A, L], a scale of yuvé is [64, A, L], and they have equal width and height. The processing steps are as follows: (1) Perform subsampling once on yY to obtain yY1. (2) Insert yY1e yu^ into a concatenation module (Concat), and perform channel concatenation on them to obtain xi. A scale of xi is [192, H, W]. (3) Insert xi into a convolution module (Conv) and output X2. A scale of X2 is [192, H, W]. In the convolution module, 192 indicates a number of output channels of the convolution module, 3x3 indicates a size of a convolution kernel of the convolution module, and S1 indicates a step of the convolution module. (4) Add xi and X2 in an element-by-element manner to obtain xs. A scale of xs is [192, H, W].

[0053] The structure in the thick line box is a ResBlock structure, which has the following characteristic: a tensioner of Petition 870250108800, dated 11 / 27 / 2025, page 22 / 64 The input feature and output feature tensor of the convolution module have the same number of channels, so the feature tensors before and after convolution can be added directly via jump connections. However, in the JPEG AI framework, an input feature tensor is formed by concatenating the Y and UV components, often resulting in a large number of channels. If the number of channels is reduced to implement weight reduction, the computational complexity is reduced, but performance deteriorates drastically.

[0054] To solve the above technical problem, in the above network and encoding / decoding application scenario, the embodiments of this application provide an image decompression method to improve the performance of an image / video compression algorithm. It should be noted that an image in the embodiments of this application can be a separate image, for example, a photo or a picture, or it can be image frames in a video. In other words, the method in the embodiments of this application can be used to process a separate image or a sequence of image frames in a video. Unless otherwise specified, a video / image is collectively referred to as an image in this descriptive report.

[0055] Figure 6 is a flowchart of a 600 process of an image decompression method according to an embodiment of this application. The 600 process can be performed by the decoder in the previous embodiments. The 600 process is described as a series of steps or operations. It should be understood that the steps or operations of the 600 process can be performed in several sequences and / or simultaneously, and are not limited to the execution sequence shown in Figure 6. The 600 process includes the following steps.

[0056] Step 601: Obtain a first feature tensor. The first feature tensor corresponds to a first Petition 870250108800, dated 11 / 27 / 2025, page 23 / 64 17 / 45 component of a reconstructed image.

[0057] Step 602: Obtain a second feature tensor. The second feature tensor corresponds to a second component of the reconstructed image.

[0058] Optionally, the first component can be a Y component of an image represented using a YUV color space, and the second component can be a UV component of the image represented using the YUV color space.

[0059] It should be noted that the first component and the second component may alternatively be other components based on a way of representing the image. This is not specifically limited in modalities of this application.

[0060] As described above, in a JPEG AI structure, on the encoder side, an image is decomposed into a Y component and a UV component, which are then encoded and transmitted separately; and on the decoder side, the Y component and the UV component are decoded and restored separately. It can be learned that the Y component (i.e., the first component) and the UV component (i.e., the second component) are processed independently in two separate paths. With reference to Figure 5a, in a path for processing the UV component on the decoder side, the Y component and the UV component are merged, i.e., operations on the thick line box shown in Figure 5b are performed. It is assumed that the structure (configured to merge the Y component and the UV component) Let the UV component in the thick-line box be considered a black box. In this case, the black box has two inputs: the first feature tensor (corresponds to component Y and comes from a path for processing component Y; for its specific process of obtaining it, see the following embodiments) and the second feature tensor (corresponds to component UV and comes from the path for processing component UV; for its specific process of obtaining it, see the following embodiments). Petition 870250108800, dated 11 / 27 / 2025, page 24 / 64 18 / 45 modes). The black box has an output (a fourth tensioner with the characteristic below).

[0061] In some embodiments of this application, a framework for implementing a merge function is enhanced. See the following steps.

[0062] Step 603: Perform concatenation and convolution on the first feature tensor and the second feature tensor to obtain a third feature tensor.

[0063] Concatenating (Concat) the first feature tensor and the second feature tensor can be done by performing channel concatenation on them. For example, if a scale of the first feature tensor is [128, H, W] and a scale of the second feature tensor is [64, H, W], a scale of a concatenated feature tensor will be [192, H, W]. That is, a width and a height remain unchanged, and a channel quantity is a sum of their channel quantities. Then, convolution (Conv) is performed on the concatenated feature tensor. Convolution can increase / decrease / not change the channel quantity. Correspondingly, a channel quantity of the third output feature tensor is increased / decreased / not changed. For example, a scale of the third feature tensor will be [64, H, W], [128, H, W] or [192, H, W].

[0064] Step 604: Obtain a fourth feature tensor based on the third feature tensor and the second feature tensor.

[0065] Optionally, the third feature tensor and the second feature tensor can be concatenated to obtain the fourth feature tensor.

[0066] Concatenation is channel concatenation. That is, channel concatenation is performed on the third feature tensor and the second feature tensor.

[0067] Figure 7a is a diagram of a fusion structure. Petition 870250108800, dated 11 / 27 / 2025, page 25 / 64 19 / 45 As shown in Figure 7a, the fusion structure concatenates the first feature tensor and the second feature tensor and performs a convolution to obtain the third feature tensor, and concatenates the third feature tensor and the second feature tensor to obtain the fourth feature tensor. A scale of the first feature tensor yY1 is [cY, H, W]. A scale of the second feature tensor yuvé [ cuv, H, W]. They have equal width and height. The steps are as follows: (1) Insert yY1 and yuv into a Concat module and perform channel concatenation on them to obtain xi. A scale of xi is [cY + Cuv, H, W]. (2) Insert xi into a Conv module and output the third characteristic tensor x2. A scale of x2 is [cUv, H, W]. (3) Insert yuve x2 into a Concat module and perform channel concatenation on them to obtain the fourth characteristic tensor x 3. A scale of x 3 is [cUv +cuv, H, W].

[0068] It should be noted that the number of channels of the fourth feature tensor is not specifically limited in the embodiments of this application. For example, c'uv + cuv may be less than cY + cuv; c'uv + cuv may be greater than cY + cuv; c'uv + cuv may be less than cy; or similar. In the previous case, if the number of channels of the fourth feature tensor is less than the sum of the number of channels of the first feature tensor and the second feature tensor, the number of channels may be reduced to save computing power and implement a lighter structure. Even if the number of channels of the fourth feature tensor is greater than the sum of the number of channels of the first feature tensor and the second feature tensor, compared to the embodiment shown in Figure 5b, in this embodiment, after convolution, element-by-element addition is changed to concatenation, and element-by-element addition of Petition 870250108800, dated 11 / 27 / 2025, page 26 / 64 In a 20 / 45 configuration, the output of the convolution and a fused characteristic tensor of the Y component and the UV component is changed to concatenation of the output and only the UV component, so that performance can be significantly improved.

[0069] Optionally, channel extraction can be performed on the second feature tensor to obtain a fifth feature tensor; and the third feature tensor and the fifth feature tensor can be concatenated to obtain the fourth feature tensor.

[0070] Channel reduction can be performed on the second feature tensor to obtain the fifth feature tensor, and then the third feature tensor and the fifth feature tensor can be concatenated.

[0071] Figure 7b is a diagram of a fusion structure. As shown in Figure 7b, the fusion structure concatenates the first feature tensor and the second feature tensor and performs a convolution to obtain the third feature tensor, performs channel extraction on the second feature tensor to obtain the fifth feature tensor, and concatenates the third feature tensor and the fifth feature tensor to obtain the fourth feature tensor. A scale of the first feature tensor yyi is [cy, H, W]. A scale of the second feature tensor yuvé is [cuv, H, W]. They have equal width and height. The steps are as follows: (1) Insert yY1 and yuv into a Concat module and perform channel concatenation on them to obtain xi. A scale of xi is [c Y + Cuv, H, W]. (2) Insert xi into a Conv module and output the third characteristic tensor x 2. A scale of x 2 is [cUv, H, W]. (3) Extract c^vchannels from yuv according to a predefined rule to obtain the fifth characteristic tensor yUv. A scale of yUv is [uV,, H, W]. cW may or may not be equal to cUv. Petition 870250108800, dated 11 / 27 / 2025, p. 27 / 64 21 / 45 This is not specifically limited to this document. (4) Insert yúv and x2 into a Concat module and perform channel concatenation on them to obtain the fourth tensor of characteristic x 3. A scale of x 3 is [cúv+ c^v, H, W].

[0072] It should be noted that the number of channels in the fourth feature tensor is not specifically limited in the embodiments of this application. For example, c'vv+cüv may be less than cY+cvv; cuv+cuV may be greater than cY+cvv; cuv+c'Vy may be less than cYu+cvv; or similar. In the previous case, if the number of channels in the fourth feature tensor is less than the sum of the number of channels in the first feature tensor and the second feature tensor, the number of channels may be reduced to save computing power and implement a lighter structure.Even though the number of channels in the fourth feature tensor is greater than the sum of the number of channels in the first and second feature tensors, compared to the embodiment shown in Figure 5b, in this embodiment, after convolution, the element-by-element addition is changed to concatenation, and the element-by-element addition of an output from the convolution and a merged feature tensor of the Y component and the UV component is changed to concatenation of the output and only the UV component, so that the performance can be significantly improved.

[0073] Optionally, the third feature tensor and the second feature tensor can be added to obtain the fourth feature tensor.

[0074] If the number of channels in the third feature tensor is equal to that of the second feature tensor, they can be added in an element-by-element manner. If the number of channels in the third feature tensor is not equal to that of the second feature tensor, they can be added by adding the channels of the tensor of Petition 870250108800, dated 11 / 27 / 2025, page 28 / 64 22 / 45 characteristic tensor with the fewest channels and some channels of the characteristic tensor with the most channels, in an element-by-element manner. For example, the third characteristic tensor has 64 channels and the second characteristic tensor has 32 channels. The 32 channels of the second characteristic tensor and the first 32 channels of the third characteristic tensor can be added in an element-by-element manner; the 32 channels of the second characteristic tensor and the last 32 channels of the third characteristic tensor can be added in an element-by-element manner; or similar. A method of addition, in this case, is not specifically limited to the embodiments of this application.

[0075] Figure 7c is a diagram of a fusion structure. As shown in Figure 7c, the fusion structure concatenates the first feature tensor and the second feature tensor and performs a convolution to obtain the third feature tensor, and adds the third feature tensor and the second feature tensor to obtain the fourth feature tensor. A scale of the first feature tensor yY1 is [cY, H, W]. A scale of the second feature tensor yuvé [cuv, H, W]. They have equal width and height. The steps are as follows: (1) Insert yY1 and yuv into a Concat module and perform channel concatenation on them to obtain xi. A scale of xi is [cY + cuv, H, W]. (2) Insert xi into a Conv module and output the third characteristic tensor x2. A scale of x2 is [^v, H, W]. (3) Add yuve x2 to obtain the fourth characteristic tensor x 3. A scale of x 3 is [c'Uv, H, W]. is the largest cuve cuv.

[0076] It should be noted that the number of channels of the fourth characteristic tensor is not specifically limited in the embodiments of this application. For example, cW may be less than cY+ Petition 870250108800, dated 11 / 27 / 2025, p. 29 / 64 23 / 45 cuv; cuv can be greater than cY+ cuv; u VY can be less than cy; or similar. In the previous case, if the number of channels of the fourth feature tensor is less than the sum of the number of channels of the first feature tensor and the second feature tensor, the number of channels can be reduced to save computing power and implement a lighter structure. Even if the number of channels of the fourth feature tensor is greater than the sum of the number of channels of the first feature tensor and the second feature tensor, compared to the embodiment shown in Figure 5b, in this embodiment, after convolution, the element-by-element addition of a convolution output and a merged feature tensor of the Y component and the UV component is changed to element-by-element addition of the output and only the UV component, so that performance can be significantly improved.

[0077] Optionally, channel extraction can be performed on the second feature tensor to obtain a fifth feature tensor; and the third feature tensor and the fifth feature tensor can be added to obtain the fourth feature tensor.

[0078] If the number of channels in the third feature tensor is less than that in the second feature tensor, channel reduction can be performed on the second feature tensor to obtain the fifth feature tensor, so that the number of channels in the fifth feature tensor is equal to that of the second feature tensor; and then the third feature tensor and the fifth feature tensor can be added in an element-by-element manner.

[0079] Figure 7d is a diagram of a fusion structure. As shown in Figure 7d, the fusion structure concatenates the first feature tensor and the second feature tensor. Petition 870250108800, dated 11 / 27 / 2025, p. 30 / 64 24 / 45 feature and performs a convolution to obtain the third feature tensor, performs channel extraction on the second feature tensor to obtain the fifth feature tensor, and adds the third and fifth feature tensors to obtain the fourth feature tensor. A scale of the first feature tensor yY1 is [cY, H, W]. A scale of the second feature tensor yuvé is [cuv, H, W]. They have equal width and height. The steps are as follows: (1) Insert yY1 and yuv into a Concat module and perform channel concatenation on them to obtain x i. A scale of xi is [cY + Cuv, H, W]. (2) Insert xi into a Conv module and output the third characteristic tensor x2. A scale of x2 is [cUv, H, W]. (3) Extract cUv channels from yuv according to a predefined rule to obtain the fifth characteristic tensor yúv · A scale of yUv is [cUv, H, W]. (4) Add yUv and x2 to obtain the fourth characteristic tensor x 3. A scale of x 3 is [cUv, H, W].

[0080] It should be noted that the number of channels of the fourth feature tensor is not specifically limited in the embodiments of this application. For example, cUv may be less than cY + cUv; cUv may be greater than CY + cuv; c^v may be less than tyou cuv; or similar. In the previous case, if the number of channels of the fourth feature tensor is less than the sum of the number of channels of the first feature tensor and the second feature tensor, the number of channels may be reduced to save computing power and implement a lighter structure. Even if the number of channels of the fourth feature tensor is greater than the sum of the number of channels of the first feature tensor and the second feature tensor, compared to the embodiment shown in Figure 5b, in this embodiment, after convolution, the Petition 870250108800, dated 11 / 27 / 2025, page 31 / 64 The 25 / 45 element-by-element addition of a convolution output and a fused characteristic tensor of the Y component and the UV component is changed to element-by-element addition of the output and only the UV component, so that performance can be significantly improved.

[0081] Step 605: Obtain the reconstructed image based on the first feature tensor and the fourth feature tensor.

[0082] The UV component of the reconstructed image can be obtained through some processing (see the following embodiments) on the fourth feature tensor obtained through fusion and then combined with the Y component of the reconstructed image obtained by processing the first feature tensor, to form the complete reconstructed image.

[0083] In some embodiments of this application, the structure for merging the Y component and the UV component is optimized. This can reduce the number of channels, saving computing power and implementing a lightweight design. Even if the number of channels is not reduced, performance can still be significantly improved.

[0084] The following describes the technical solutions for this application through various specific modalities.

[0085] Figure 8 is a diagram of a JPEG AI structure using a merging structure. As shown in Figure 8, JPEG AI uses the merging structure shown in Figure 7a or Figure 7b, and its processing is as follows:

[0086] Encoder side: 1. Convert an original RGB image into a YUV444 image through color space conversion. 2. Decompose the YUV444 image obtained in step 1 by channel, to obtain a Y component and a UV component. 3. Insert the Y component obtained in step 2 into a Y component encoding network (Y Encoder) and perform the reduction of Petition 870250108800, dated 11 / 27 / 2025, page 32 / 64 26 / 45 sampling four times to obtain y_Y. 4. Process y_Y obtained in step 3 using a Y-component hyperprior encoding network (Y Hyper Encoder Net) and then perform nearest-rounding quantization to obtain z_Y_hat. 5. Process z_Y_hat obtained in step 4 using an entropy encoder (Lossless Encoder) to obtain a first bit stream, which is denoted as bit_z_Y. 6. Process z_Y_hat obtained in step 4 using a Y-component hyperpriority decoder network to obtain mean_Y. 7. Process z_Y_hat obtained in step 4 using a Y-component hyperprior variance decoder network to obtain variance_Y. 8. Subtract mean_Y obtained in step 6 from y_Y obtained in step 3 to obtain residual_Y and perform rounding quantization to the nearest to obtain residual_Y_hat. 9. Obtain a Gaussian probability distribution function with a mean of 0 and a standard deviation of variance_Y based on variance_Y generated in step 7, and process the residual_Y_hat obtained in step 8 using an entropy encoder (lossless encoder) to obtain a second bit stream, denoted as bit_y_Y. 10. Separately perform subsampling once on the Y component and the UV component obtained in step 2, and then insert the subsampled components into a UV component encoding network (UV Encoder) and perform subsampling three times on the UV Encoder network to obtain y_UV. (The UV Encoder network includes three times the subsampling.) 11. Process y_UV obtained in step 10 using a UV component hyperpriority encoding network (UV Hyper Encoder Net) and then perform nearest-rounding quantization to obtain z_UV_hat. Petition 870250108800, dated 11 / 27 / 2025, page 33 / 64 27 / 45 12. Process z_UV_hat obtained in step 11 through the entropy encoder (Lossless Encoder) to obtain a third bit stream, which is denoted as bit_z_UV. 13. Process z_UV_hat obtained in step 11 using a UV component hyperpriority decoding network to obtain mean_UV. 14. Process z_UV_hat obtained in step 11 using a UV component hyperprior variance decoder network to obtain variance_UV. 15. Subtract the mean_UV output from step 13 from the y_UV output from step 10 to obtain residual_UV and perform rounding quantization to the nearest to obtain residual_UV_hat. 16. Obtain a Gaussian probability distribution function with a mean of 0 and a standard deviation of variance_UV based on the variance_UV generated in step 14 and process residual_UV_hat obtained in step 15 using the entropy encoder (Lossless Encoder) to obtain a fourth bit stream, denoted as bit_y_UV. 17. Obtain a sum of the four previous bit streams as a final encoded bit stream.

[0087] Decoder side: 1. Process the first bit stream bit_z_Y using an entropy decoder (lossless decoder) to obtain z_Y_hat. 2. Process z_Y_hat obtained in step 1 using the Y-component hyperpriority decoder network to obtain mean_Y. 3. Process z_Y_hat obtained in step 1 using the Y-component hyperprior variance decoder network to obtain variance_Y. 4. Obtain a Gaussian probability distribution function with a mean of 0 and a standard deviation of variance_Y based on variance_Y generated in step 3 and process the second Petition 870250108800, dated 11 / 27 / 2025, page 34 / 64 28 / 45 bit stream bit_y_Y through an entropy decoder (lossless decoder) to obtain residual_Y_hat. 5. Add residual_Y_hat obtained in step 4 and mean_Y obtained in step 2, to obtain y1_Y. 6. Process y1_Y obtained in step 5 using a Y-component decoding network to obtain a Y-component from a reconstructed image. 7. Process the third bit stream bit_z_UV through the entropy decoder (lossless decoder) to obtain z_UV_hat. 8. Process z_UV_hat obtained in step 7 using the UV component hyperpriority decoding network to obtain mean_UV. 9. Process z_UV_hat obtained in step 7 using the UV Hyper Scale Decoder network to obtain variance_UV. 10. Obtain a Gaussian probability distribution function with a mean of 0 and a standard deviation of variance_UV based on the variance_UV generated in step 9, and process the fourth bitstream bit_y_UV using the entropy decoder (Lossless Decoder) to obtain residual_UV_hat. 11. Add residual_UV_hat obtained in step 10 and mean_UV obtained in step 8, to obtain y1_UV. 12. Perform channel concatenation on y1_Y obtained in step 5 and y1_UV obtained in step 11, to obtain y1. 13. Process y1 obtained in step 12 using a convolutional layer to obtain y2. 14. Perform channel concatenation on y2 obtained in step 13 and y1_UV obtained in step 11, to obtain y3. 15. Insert y3 obtained in step 14 into a UV component decoding network (UV Decoder), perform oversampling three times on the UV Decoder network (optionally, the three oversampling times can be...). Petition 870250108800, dated 11 / 27 / 2025, page 35 / 64 29 / 45 implemented by performing one oversampling scale via a deconvolution module, then two oversampling scales via pixel shuffling, and then performing oversampling once on an output of the UV Decoder (optionally, oversampling can be implemented via nearest neighbor interpolation) to obtain a UV component of the reconstructed image. (The UV Decoder network includes three oversampling times.) 16. Concatenate the Y component obtained in step 6 and the UV component obtained in step 15, and then perform the color conversion to obtain a reconstructed RGB image.

[0088] Figure 9 is a diagram of a JPEG AI structure using a merging structure. As shown in Figure 9, JPEG AI uses the merging structure shown in Figure 7c or Figure 7d, and its processing is as follows:

[0089] Encoder side: 1. Convert an original RGB image into a YUV444 image through color space conversion. 2. Decompose the YUV444 image obtained in step 1 by channel, to obtain a Y component and a UV component. 3. Insert the Y component obtained in step 2 into a Y component encoding network (Y Encoder) and perform four times downsampling to obtain y_Y. 4. Process y_Y obtained in step 3 using a Y-component hyperprior encoding network (Y Hyper Encoder Net) and then perform nearest-rounding quantization to obtain z_Y_hat. 5. Process z_Y_hat obtained in step 4 using an entropy encoder (Lossless Encoder) to obtain a first bit stream, which is denoted as bit_z_Y. 6. Process z_Y_hat obtained in step 4 using a Y-component hyperpriority decoding network. Petition 870250108800, dated 11 / 27 / 2025, page 36 / 64 30 / 45 to get mean_Y. 7. Process z_Y_hat obtained in step 4 using a Y-component hyperprior variance decoder network to obtain variance_Y. 8. Subtract mean_Y obtained in step 6 from y_Y obtained in step 3 to obtain residual_Y and perform rounding quantization to the nearest to obtain residual_Y_hat. 9. Obtain a Gaussian probability distribution function with a mean of 0 and a standard deviation of variance_Y based on variance_Y generated in step 7, and process the residual_Y_hat obtained in step 8 using an entropy encoder (lossless encoder) to obtain a second bit stream, denoted as bit_y_Y. 10. Separately perform subsampling once on the Y component and the UV component obtained in step 2, and then insert the subsampled components into a UV component encoding network (UV Encoder) and perform subsampling three times on the UV Encoder network to obtain y_UV. (The UV Encoder network includes three times the subsampling.) 11. Process y_UV obtained in step 10 using a UV component hyperpriority encoding network (UV Hyper Encoder Net) and then perform nearest-rounding quantization to obtain z_UV_hat. 12. Process z_UV_hat obtained in step 11 through the entropy encoder (Lossless Encoder) to obtain a third bit stream, which is denoted as bit_z_UV. 13. Process z_UV_hat obtained in step 11 using a UV component hyperpriority decoding network to obtain mean_UV. 14. Process z_UV_hat obtained in step 11 using a UV component hyperprior variance decoder network to obtain variance_UV. 15. Subtract the mean_UV output from step 13 from the y_UV output in Petition 870250108800, dated 11 / 27 / 2025, page 37 / 64 Step 10 (31 / 45) to obtain residual_UV and perform rounding quantization to the nearest value to obtain residual_UV_hat. 16. Obtain a Gaussian probability distribution function with a mean of 0 and a standard deviation of variance_UV based on the variance_UV generated in step 14 and process residual_UV_hat obtained in step 15 using the entropy encoder (Lossless Encoder) to obtain a fourth bit rate, which is denoted as bit_y_UV. 17. Obtain a sum of the four previous bit streams as a final encoded bit stream.

[0090] Decoder side: 1. Process the first bit stream bit_z_Y using an entropy decoder (lossless decoder) to obtain z_Y_hat. 2. Process z_Y_hat obtained in step 1 using the Y-component hyperpriority decoder network to obtain mean_Y. 3. Process z_Y_hat obtained in step 1 using the Y-component hyperprior variance decoder network to obtain variance_Y. 4. Obtain a Gaussian probability distribution function with a mean of 0 and a standard deviation of variance_Y based on variance_Y generated in step 2, and process the second bit stream bit_y_Y using an entropy decoder (lossless decoder) to obtain residual_Y_hat. 5. Add residual_Y_hat obtained in step 4 and mean_Y obtained in step 2, to obtain y1_Y. 6. Process y1_Y obtained in step 5 using a Y-component decoding network to obtain a Y-component from a reconstructed image. 7. Process the third bit stream bit_z_UV through the entropy decoder (lossless decoder) to obtain zUVhat. Petition 870250108800, dated 11 / 27 / 2025, page 38 / 64 32 / 45 8. Process z_UV_hat obtained in step 7 using the UV component hyperpriority decoding network to obtain mean_UV. 9. Process z_UV_hat obtained in step 7 using the UV Hyper Scale Decoder network to obtain variance_UV. 10. Obtain a Gaussian probability distribution function with a mean of 0 and a standard deviation of variance_UV based on the variance_UV generated in step 9, and process the fourth bitstream bit_y_UV using the entropy decoder (Lossless Decoder) to obtain residual_UV_hat. 11. Add residual_UV_hat obtained in step 10 and mean_UV obtained in step 8, to obtain y1_UV. 12. Perform channel concatenation on y1_Y obtained in step 11 and y1_UV obtained in step 12, to obtain y1. 13. Process y1 obtained in step 12 using a convolutional layer to obtain y2. 14. Add y2 obtained in step 13 and y1_UV obtained in step 11 to obtain y3. 15. Insert y3 obtained in step 14 into a UV component decoding network (UV Decoder), perform three amplification steps in the UV Decoder network, and then perform one amplification step at an output of the UV Decoder to obtain a UV component of the reconstructed image. (The UV Decoder network includes three oversampling steps.) 16. Concatenate the Y component obtained in step 6 and the UV component obtained in step 15, and then perform the color conversion to obtain a reconstructed RGB image.

[0091] A difference between this mode and the mode shown in Figure 8 lies in steps 12 to 14, in bold, on the decoder side, above. The method of adding two feature tensors is used in this mode.

[0092] Figure 10 is a diagram of a JPEG AI structure. Petition 870250108800, dated 11 / 27 / 2025, page 39 / 64 33 / 45 using a merging structure. As shown in Figure 10, JPEG AI uses the merging structure shown in Figure 7a or Figure 7b, and its processing is as follows:

[0093] Encoder side: 1. Convert an original RGB image into a YUV444 image through color space conversion. 2. Decompose the YUV444 image obtained in step 1 by channel, to obtain a Y component and a UV component. 3. Insert the Y component obtained in step 2 into a Y component encoding network (Y Encoder) and perform four times the sample reduction to obtain y_Y. 4. Process y_Y obtained in step 3 using a Y-component hyperprior encoding network (Y Hyper Encoder Net) and then perform nearest-rounding quantization to obtain z_Y_hat. 5. Process z_Y_hat obtained in step 4 using an entropy encoder (Lossless Encoder) to obtain a first bit stream, which is denoted as bit_z_Y. 6. Process z_Y_hat obtained in step 4 using a Y-component hyperpriority decoder network to obtain mean_Y. 7. Process z_Y_hat obtained in step 4 using a Y-component hyperprior variance decoder network to obtain variance_Y. 8. Subtract mean_Y obtained in step 6 from y_Y obtained in step 3 to obtain residual_Y and perform rounding quantization to the nearest to obtain residual_Y_hat. 9. Obtain a Gaussian probability distribution function with a mean of 0 and a standard deviation of variance_Y based on variance_Y generated in step 7, and process the residual_Y_hat obtained in step 8 using an entropy encoder (lossless encoder) to obtain a second bit stream, denoted as bit_y_Y. Petition 870250108800, dated 11 / 27 / 2025, page 40 / 64 34 / 45 10. Insert the Y component and the UV component obtained in step 2 into a UV component encoding network (UV Encoder) and perform four subsampling on the UV Encoder network to obtain y_UV. (The UV Encoder network includes four times the subsampling.) 11. Process y_UV obtained in step 10 using a UV component hyperpriority encoding network (UV Hyper Encoder Net) and then perform nearest-rounding quantization to obtain z_UV_hat. 12. Process z_UV_hat obtained in step 11 through the entropy encoder (Lossless Encoder) to obtain a third bit stream, which is denoted as bit_z_UV. 13. Process z_UV_hat obtained in step 11 using a UV component hyperpriority decoding network to obtain mean_UV. 14. Process z_UV_hat obtained in step 11 using a UV component hyperprior variance decoder network to obtain variance_UV. 15. Subtract the mean_UV output from step 13 from the y_UV output from step 10 to obtain residual_UV and perform rounding quantization to the nearest to obtain residual_UV_hat. 16. Obtain a Gaussian probability distribution function with a mean of 0 and a standard deviation of variance_UV based on the variance_UV generated in step 14 and process residual_UV_hat obtained in step 15 using the entropy encoder (Lossless Encoder) to obtain a fourth bit rate, which is denoted as bit_y_UV. 17. Obtain a sum of the four previous bit streams as a final encoded bit stream.

[0094] Decoder side: 1. Process the first bit stream bit_z_Y using an entropy decoder (lossless decoder) to obtain zYhat. Petition 870250108800, dated 11 / 27 / 2025, page 41 / 64 35 / 45 2. Process z_Y_hat obtained in step 1 using the Y-component hyperpriority decoder network to obtain mean_Y. 3. Process z_Y_hat obtained in step 1 using the Y-component hyperprior variance decoder network to obtain variance_Y. 4. Obtain a Gaussian probability distribution function with a mean of 0 and a standard deviation of variance_Y based on variance_Y generated in step 2, and process the second bit stream bit_y_Y using an entropy decoder (lossless decoder) to obtain residual_Y_hat. 5. Add residual_Y_hat obtained in step 4 and mean_Y obtained in step 2, to obtain y1_Y. 6. Process y1_Y obtained in step 5 using a Y-component decoding network to obtain a Y-component from a reconstructed image. 7. Process the third bit stream bit_z_UV through the entropy decoder (lossless decoder) to obtain z_UV_hat. 8. Process z_UV_hat obtained in step 7 using the UV component hyperpriority decoding network to obtain mean_UV. 9. Process z_UV_hat obtained in step 7 using the UV Hyper Scale Decoder network to obtain variance_UV. 10. Obtain a Gaussian probability distribution function with a mean of 0 and a standard deviation of variance_UV based on the variance_UV generated in step 9, and process the fourth bitstream bit_y_UV using the entropy decoder (Lossless Decoder) to obtain residual_UV_hat. 11. Add residual_UV_hat obtained in step 10 and mean_UV obtained in step 8, to obtain y1_UV. 12. Perform channel concatenation on y1_Y obtained in step Petition 870250108800, dated 11 / 27 / 2025, page 42 / 64 36 / 45 and y1_UV obtained in step 11, to obtain y1. 13. Process y1 obtained in step 12 using a convolutional layer to obtain y2. 14. Perform channel concatenation on y2 obtained in step 13 and y1_UV obtained in step 11, to obtain y3. 15. Insert y3 obtained in step 14 into a UV component decoding network (UV Decoder) and perform four oversamplings on the UV Decoder network (optionally, the four oversamplings can be implemented by performing two oversampling scales using two deconvolution modules and then two oversampling scales using pixel shuffling) to obtain a UV component of the reconstructed image. (The UV Decoder network includes four oversamplings.) 16. Concatenate the Y component obtained in step 6 and the UV component obtained in step 15, and then perform the color conversion to obtain a reconstructed RGB image.

[0095] A difference between this mode and the mode shown in Figure 8 lies in step 10 in bold on the encoder side and step 15 in bold on the decoder side, above. Specifically, on the encoder side, no subsampling is performed on the Y component and the UV component, and the number of times subsampling is performed on the UV encoder changes from three to four; and on the decoder side, the number of times oversampling is performed on the UV decoder also changes from three to four, and no oversampling is performed on the UV decoder output.

[0096] Figure 11 is a diagram of a JPEG AI structure using a merging structure. As shown in Figure 11, JPEG AI uses the merging structure shown in Figure 7a or Figure 7b, and its processing is as follows:

[0097] Encoder side: Petition 870250108800, dated 11 / 27 / 2025, pp. 43 / 64 37 / 45 1. Convert an original RGB image into a YUV444 image through color space conversion. 2. Decompose the YUV444 image obtained in step 1 by channel, to obtain a Y component and a UV component. 3. Insert the Y component obtained in step 2 into a Y component encoding network (Y Encoder) and perform four times downsampling to obtain y_Y. 4. Process y_Y obtained in step 3 using a Y-component hyperprior encoding network (Y Hyper Encoder Net) and then perform nearest-rounding quantization to obtain z_Y_hat. 5. Process z_Y_hat obtained in step 4 using an entropy encoder (Lossless Encoder) to obtain a first bit stream, which is denoted as bit_z_Y. 6. Process z_Y_hat obtained in step 4 using a Y-component hyperpriority decoder network to obtain mean_Y. 7. Process z_Y_hat obtained in step 4 using a Y-component hyperprior variance decoder network to obtain variance_Y. 8. Subtract mean_Y obtained in step 6 from y_Y obtained in step 3 to obtain residual_Y and perform rounding quantization to the nearest to obtain residual_Y_hat. 9. Obtain a Gaussian probability distribution function with a mean of 0 and a standard deviation of variance_Y based on variance_Y generated in step 7, and process the residual_Y_hat obtained in step 8 using an entropy encoder (lossless encoder) to obtain a second bit stream, denoted as bit_y_Y. 10. Separately perform subsampling once on the Y component and the UV component obtained in step 2, and then insert the subsampled components into a UV component encoding network (UV Encoder) and perform subsampling. Petition 870250108800, dated 11 / 27 / 2025, pp. 44 / 64 38 / 45 four times on the UV Encoder network to obtain y_UV. (The UV Encoder network includes four times the subsampling.) 11. Process y_UV obtained in step 10 using a UV component hyperpriority encoding network (UV Hyper Encoder Net) and then perform nearest-rounding quantization to obtain z_UV_hat. 12. Process z_UV_hat obtained in step 11 through the entropy encoder (Lossless Encoder) to obtain a third bit stream, which is denoted as bit_z_UV. 13. Process z_UV_hat obtained in step 11 using a UV component hyperpriority decoding network to obtain mean_UV. 14. Process z_UV_hat obtained in step 11 using a UV component hyperprior variance decoder network to obtain variance_UV. 15. Subtract the mean_UV output from step 13 from the y_UV output from step 10 to obtain residual_UV and perform rounding quantization to the nearest to obtain residual_UV_hat. 16. Obtain a Gaussian probability distribution function with a mean of 0 and a standard deviation of variance_UV based on the variance_UV generated in step 14 and process residual_UV_hat obtained in step 15 using the entropy encoder (Lossless Encoder) to obtain a fourth bit rate, which is denoted as bit_y_UV. 17. Obtain a sum of the four previous bit streams as a final encoded bit stream.

[0098] Decoder side: 1. Process the first bit stream bit_z_Y using an entropy decoder (lossless decoder) to obtain z_Y_hat. 2. Process z_Y_hat obtained in step 1 using the Y-component hyperpriority decoder network to obtain mean_Y. Petition 870250108800, dated 11 / 27 / 2025, page 45 / 64 39 / 45 3. Process z_Y_hat obtained in step 1 using the Y-component hyperprior variance decoder network to obtain variance_Y. 4. Obtain a Gaussian probability distribution function with a mean of 0 and a standard deviation of variance_Y based on variance_Y generated in step 2, and process the second bit stream bit_y_Y using an entropy decoder (lossless decoder) to obtain residual_Y_hat. 5. Add residual_Y_hat obtained in step 4 and mean_Y obtained in step 2, to obtain y1_Y. 6. Process y1_Y obtained in step 5 using a Y-component decoding network to obtain a Y-component from a reconstructed image. 7. Process the third bit stream bit_z_UV through the entropy decoder (lossless decoder) to obtain z_UV_hat. 8. Process z_UV_hat obtained in step 7 using the UV component hyperpriority decoding network to obtain mean_UV. 9. Process z_UV_hat obtained in step 7 using the UV Hyper Scale Decoder network to obtain variance_UV. 10. Obtain a Gaussian probability distribution function with a mean of 0 and a standard deviation of variance_UV based on the variance_UV generated in step 9, and process the fourth bitstream bit_y_UV using the entropy decoder (Lossless Decoder) to obtain residual_UV_hat. 11. Add residual_UV_hat obtained in step 10 and mean_UV obtained in step 8, to obtain y1_UV. 12. Perform subsampling once on y1_Y obtained in step 5 to obtain y2_Y and perform channel concatenation on y2_Y and y1_UV obtained in step 11 to obtain y1. 13. Process y1 obtained in step 12 using a layer. Petition 870250108800, dated 11 / 27 / 2025, pp. 46 / 64 40 / 45 convolutional to obtain y2. 14. Perform channel concatenation on y2 obtained in step 13 and y1_UV obtained in step 11, to obtain y3. 15. Insert y3 obtained in step 14 into a UV component decoding network (UV Decoder), perform four oversampling steps on the UV Decoder network, and then perform one oversampling step on an output of the UV Decoder to obtain a UV component of the reconstructed image. (The UV Decoder network includes four oversampling steps.) 16. Concatenate the Y component obtained in step 6 and the UV component obtained in step 15, and then perform the color conversion to obtain a reconstructed RGB image.

[0099] A difference between this mode and the mode shown in Figure 8 lies in step 10 in bold on the encoder side and step 15 in bold on the decoder side, above. Specifically, on the encoder side, the number of subsampling times in the UV encoder changes from three to four; and on the decoder side, the number of oversampling times in the UV decoder also changes from three to four.

[00100] Figure 12 is a diagram of the structure of an image decompression device 1200, according to an embodiment of this application. As shown in Figure 12, the image decompression device 1200 in this embodiment can be used as a decoder. The image decompression device 1200 may include an acquisition module 1201, a processing module 1202, and a reconstruction module 1203.

[00101] The acquisition module 1201 is configured to obtain a first feature tensor, where the first feature tensor corresponds to a first component of a reconstructed image; and to obtain a second feature tensor, where the second feature tensor corresponds to a second component of the reconstructed image. The processing module 1202 is configured to perform concatenation and Petition 870250108800, dated 11 / 27 / 2025, page 47 / 64 41 / 45 convolution on the first feature tensor and the second feature tensor to obtain a third feature tensor; and to obtain a fourth feature tensor based on the third feature tensor and the second feature tensor. The reconstruction module 1203 is configured to obtain the reconstructed image based on the first feature tensor and the fourth feature tensor.

[00102] In one possible implementation, the 1202 processing module is specifically configured to concatenate the third feature tensor and the second feature tensor to obtain the fourth feature tensor.

[00103] In one possible implementation, the 1202 processing module is specifically configured to add the third feature tensor and the second feature tensor to obtain the fourth feature tensor.

[00104] In one possible implementation, the 1202 processing module is specifically configured to perform channel extraction on the second feature tensor to obtain a fifth feature tensor; and to concatenate the third feature tensor and the fifth feature tensor to obtain the fourth feature tensor.

[00105] In one possible implementation, the 1202 processing module is specifically configured to perform channel extraction on the second feature tensor to obtain a fifth feature tensor; and to add the third feature tensor and the fifth feature tensor to obtain the fourth feature tensor.

[00106] In one possible implementation, the first component is a Y component and the second component is a UV component.

[00107] The device in this embodiment can be configured to implement the technical solutions of the embodiment shown in Figure 6. The implementation principles and technical effects are similar. Details are not described here. Petition 870250108800, dated 11 / 27 / 2025, pp. 48 / 64 42 / 45 document again.

[00108] In an implementation process, the steps in the previous method modalities can be performed by a hardware integrated logic circuit in a processor or by the use of instructions in software. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component. The general-purpose processor can be a microprocessor, or the processor can be any conventional processor or similar.The steps of the method described in the embodiments of this application may be presented directly as being performed and completed by a hardware encoding processor, or performed and completed by a combination of hardware in the encoding processor and a software module. The software module may be located in a storage medium mature in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or a register. The storage medium is located in a memory. The processor reads the information in the memory and completes the steps of the previous method in combination with the processor hardware.

[00109] Memory in the preceding embodiments may be volatile or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), or read-only memory. Petition 870250108800, dated 11 / 27 / 2025, page 49 / 64 43 / 45 erasable programmable memory (EPROM), an electrically erasable programmable read-only memory (EPROM, EEPROM), or a flash memory. Volatile memory can be random access memory (RAM), used as an external cache. By way of example, but not exhaustively, many forms of RAM can be used, for example, static random access memory (SRAM), dynamic random access memory (dynamic RAM, DRAM), synchronous dynamic random access memory (synchronous DRAM, SDRAM), double data rate synchronous dynamic random access memory (double data rate SDRAM, DDR SDRAM), enhanced synchronous dynamic random access memory (enhanced SDRAM, ESDRAM), synchlink dynamic random access memory (synchlink DRAM, SLDRAM), and direct Rambus random access memory (direct Rambus RAM, DR RAM). It should be noted that the system and method memory described in this descriptive report includes, among others, these and any other appropriate memory types.

[00110] A person skilled in the art may be aware that, in combination with the examples described in the embodiments disclosed in this descriptive report, algorithm units and steps may be implemented by electronic hardware or by a combination of software and electronic hardware. Whether the functions are performed by hardware or software depends on the specific applications and design constraints of the technical solutions. A person skilled in the art may use different methods to implement the described functions for each specific application, but the implementation should not be considered to go beyond the scope of this application.

[00111] It may be clearly understood by a person skilled in the art that, for the purposes of convenient and brief description, for a detailed working process of the aforementioned system, apparatus and unit, please refer to a corresponding process in Petition 870250108800, dated 11 / 27 / 2025, pages 50 / 64 44 / 45 methods as described above. Details are not described again in this document.

[00112] In the various embodiments provided in this application, it should be understood that the disclosed system, apparatus, and method may be implemented in other ways. For example, the described embodiment of the apparatus is merely an example. For example, the division into units is merely a logical function division and may be another division in a real implementation. For example, a plurality of units or components may be combined or integrated into another system, or some features may be omitted or not implemented. Furthermore, the mutual couplings or direct couplings or communication connections shown or discussed may be implemented using some interfaces. Indirect couplings or communication connections between the apparatuses or units may be implemented in electronic, mechanical, or other forms.

[00113] The units described as separate parts may or may not be physically separate, and the parts displayed as units may or may not be physical units, may be located in one position, or may be distributed across a plurality of network units. Some or all of the units may be selected based on the actual requirements to achieve the objectives of the modal solutions.

[00114] Furthermore, functional units in the modalities of this application may be integrated into a processing unit, each of the units may physically exist alone, or two or more units may be integrated into one unit.

[00115] If implemented in the form of a functional software unit and sold or used as a standalone product, the functions may be stored on a computer-readable storage medium. Based on this understanding, the technical solutions of this application essentially constitute either the part that contributes to the conventional technology, or a part of the solutions. Petition 870250108800, dated 11 / 27 / 2025, pp. 51 / 64 45 / 45 techniques can be implemented in the form of a software product. The software product is stored on a storage medium and includes various instructions to instruct a computer device (a personal computer, a server, a network device, or similar) to perform all or some of the steps of the method in the embodiments of this application. The aforementioned storage medium includes any medium that can store program code, such as a USB flash drive, a removable hard disk, read-only memory (ROM), random access memory (RAM), a magnetic disk, or an optical disk.

[00116] The descriptions above are merely specific implementations of this application, but are not intended to limit the scope of protection of this application. Any variation or substitution easily understood by a person skilled in the art, within the technical scope disclosed in this application, will be covered by the scope of protection of this application. Therefore, the scope of protection of this application will be subject to the scope of protection of the claims. Petition 870250108800, dated 11 / 27 / 2025, pp. 52 / 64

Claims

1 / 5 CLAIMS 1. Image decompression method, characterized in that it comprises: obtaining a first feature tensor, wherein the first feature tensor corresponds to a first component of a reconstructed image; obtaining a second feature tensor, wherein the second feature tensor corresponds to a second component of the reconstructed image; performing concatenation and convolution on the first feature tensor and the second feature tensor to obtain a third feature tensor; obtaining a fourth feature tensor based on the third feature tensor and the second feature tensor; and obtaining the reconstructed image based on the first feature tensor and the fourth feature tensor.

2. A method according to claim 1, characterized in that obtaining the fourth feature tensor based on the third feature tensor and the second feature tensor comprises: concatenating the third feature tensor and the second feature tensor to obtain the fourth feature tensor.

3. A method according to claim 1, characterized in that obtaining the fourth feature tensor based on the third feature tensor and the second feature tensor comprises: adding the third feature tensor and the second feature tensor to obtain the fourth feature tensor.

4. Method, according to any one of claims 1 to 3, characterized in that performing the concatenation and convolution on the first feature tensor and the second feature tensor to obtain the third feature tensor comprises: concatenating the first feature tensor and the second feature tensor to obtain a concatenated feature tensor; and performing convolution on the concatenated tensor to obtain the third feature tensor.

5. A method according to claim 4, characterized in that the number of channels for the concatenated feature tensor is the sum of the number of channels for the first feature tensor and the number of channels for the second feature tensor; the number of channels for the third feature tensor is equal to the number of channels for the second feature tensor.

6. A method according to any one of claims 1 to 5, characterized in that obtaining the fourth feature tensor based on the third feature tensor and the second feature tensor comprises: performing channel extraction on the second feature tensor to obtain a fifth feature tensor; and concatenating the third feature tensor and the fifth feature tensor to obtain the fourth feature tensor.

7. A method according to any one of claims 1 to 5, characterized in that obtaining the fourth feature tensor based on the third feature tensor and the second feature tensor comprises: performing channel extraction on the second feature tensor to obtain a fifth feature tensor; and adding the third feature tensor and the fifth feature tensor to obtain the fourth feature tensor. Petition 870250095035, dated 10 / 17 / 2025, p. 18 / 26 3 / 5 8. Method, according to any one of claims 1 to 7, characterized in that the first component is a Y component and the second component is a UV component.

9. Image decompression apparatus, characterized in that it comprises: an acquisition module, configured to obtain a first feature tensor, wherein the first feature tensor corresponds to a first component of a reconstructed image; and to obtain a second feature tensor, wherein the second feature tensor corresponds to a second component of the reconstructed image; a processing module, configured to perform concatenation and convolution on the first feature tensor and the second feature tensor to obtain a third feature tensor; and to obtain a fourth feature tensor based on the third feature tensor and the second feature tensor; and a reconstruction module, configured to obtain the reconstructed image based on the first feature tensor and the fourth feature tensor.

10. Apparatus, according to claim 9, characterized in that the processing module is specifically configured to concatenate the third feature tensor and the second feature tensor to obtain the fourth feature tensor.

11. Apparatus, according to claim 9, characterized in that the processing module is specifically configured to add the third feature tensor and the second feature tensor to obtain the fourth feature tensor.

12. Apparatus, according to any one of claims 9 to 11, characterized in that the processing module is specifically configured to: Petition 870250095035, dated 10 / 17 / 2025, page 19 / 26 4 / 5 concatenate the first feature tensor and the second feature tensor to obtain a concatenated feature tensor; and perform convolution on the concatenated tensor to obtain the third feature tensor.

13. Apparatus, according to claim 12, characterized in that the number of channels for the concatenated feature tensor is the sum of the number of channels for the first feature tensor and the number of channels for the second feature tensor; the number of channels for the third feature tensor is equal to the number of channels for the second feature tensor.

14. Apparatus, according to any one of claims 9 to 13, characterized in that the processing module is specifically configured to perform channel extraction on the second feature tensor to obtain a fifth feature tensor; and to concatenate the third feature tensor and the fifth feature tensor to obtain the fourth feature tensor.

15. Apparatus, according to any one of claims 9 to 14, characterized in that the processing module is specifically configured to perform channel extraction on the second feature tensor to obtain a fifth feature tensor; and to add the third feature tensor and the fifth feature tensor to obtain the fourth feature tensor.

16. Apparatus, according to any one of claims 9 to 15, characterized in that the first component is a Y component and the second component is a UV component.

17. Electronic device, characterized in that it comprises: one or more processors; and a memory, configured to store one or more programs; Petition 870250095035, dated 10 / 17 / 2025, page 20 / 26 5 / 5 wherein when one or more programs are executed by one or more processors, the one or more processors are enabled to implement the method as defined in any of claims 1 to 8.

18. Computer-readable storage medium characterized in that it comprises a computer program, wherein when the computer program is executed on a computer, the computer is enabled to perform the method as defined in any one of claims 1 to 8.

19. Computer program product, characterized in that the computer program product comprises computer program code and, when the computer program code is run on a computer, the computer is enabled to perform the method as defined in any of claims 1 to 8. Petition 870250095035, dated 10 / 17 / 2025, pp. 21 / 26

Images