An online video super-resolution method based on knowledge distillation

Abstract

This invention discloses an online video super-resolution method based on knowledge distillation, belonging to the field of image and video processing technology. During the training phase, a high-performance bidirectional teacher network guides a lightweight unidirectional student network. A continuous subsequence is selected from the complete training frame sequence and input into the teacher network. At the last frame of this subsequence, teacher output distillation and reconstructed feature distillation are performed, ensuring that the distillation supervision does not include information from future frames after that moment. The student network is configured with the optical flow-guided pyramid deformable convolution alignment module proposed in this invention, utilizing optical flow motion priors and multi-scale deformable sampling to align historical propagation features step-by-step. After training, only the unidirectional student network is retained for online super-resolution inference in chronological order, thus balancing video reconstruction quality and online deployment requirements with lower model complexity.

Description

Technical Field

[0001] This invention belongs to the field of image and video processing technology, specifically relating to video super-resolution reconstruction, and more particularly to an online video super-resolution method based on a teacher-student knowledge distillation mechanism, which is used to improve the reconstruction performance of lightweight video super-resolution networks while maintaining online inference capabilities. Background Technology

[0002] In the field of computer vision, Video Super-Resolution (VSR) technology aims to reconstruct low-resolution video sequences into high-resolution video sequences to improve texture details, edge structure, and overall visual quality. This technology can be applied to streaming video enhancement, surveillance video analysis, mobile terminal displays, and other scenarios requiring video quality restoration. The key to video super-resolution lies in effectively utilizing the temporal correlation between adjacent frames and achieving accurate cross-frame information fusion under conditions of motion, occlusion, and deformation.

[0003] Existing video super-resolution methods can be broadly categorized into sliding window methods and recurrent neural network (RNN) methods. Sliding window methods upsample images by extracting and fusing features from a fixed-length local frame window. This method achieves a good trade-off between reconstruction accuracy and computational latency, but its fixed window limit its ability to effectively capture long-term temporal information, thus restricting its performance in complex dynamic scenes. RNN methods, on the other hand, employ a recursive mechanism, typically utilizing temporal information from previous and future video sequences to improve image restoration quality. This method can achieve more accurate spatiotemporal feature fusion by accumulating information from historical and future frames, generally exhibiting excellent performance in metrics such as peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). However, this bidirectional information fusion structure often leads to high computational complexity and inference latency, making it difficult to deploy in real-time scenarios or on resource-constrained devices.

[0004] One-way online recursive networks can generate super-resolution results for the current frame based solely on the current frame and historical frame states, thus exhibiting characteristics such as causal inference, low latency, and ease of deployment. However, lightweight one-way networks are limited by model capacity and the range of available temporal information, and their reconstruction capabilities are typically lower than those of complex bidirectional networks. Knowledge distillation can transfer knowledge from high-performance teacher models to lightweight student models, providing an effective way to improve the performance of one-way student networks.

[0005] However, directly using a bidirectional teacher network to perform frame-by-frame distillation on a unidirectional online student network presents a temporal information asymmetry problem: for intermediate moments in the sequence, the bidirectional teacher's output may utilize future frames after that moment, while the unidirectional student cannot obtain this future information during actual online inference. If the student is directly required to fit the teacher's objective containing future information, there is an inconsistency between the distillation supervision and the information available to the student. Furthermore, if only the final super-resolution output is constrained, it is difficult to fully transfer the high-quality detailed representations formed by the teacher during the reconstruction phase.

[0006] Therefore, a knowledge distillation method for unidirectional online video super-resolution networks is needed. This method should leverage the reconstruction capabilities of high-performance bidirectional teacher networks while limiting future frame information that is not available in the distillation supervision. Furthermore, it should improve the reconstruction performance of lightweight student networks through joint distillation of output and reconstruction features. Summary of the Invention

[0007] To address the issues of high inference overhead in bidirectional high-performance models, limited reconstruction performance in unidirectional lightweight online models, and information asymmetry between bidirectional teachers and unidirectional students regarding future information, this invention provides an online video super-resolution method based on causal distillation of the last frame of a random continuous subsequence. During the training phase, the unidirectional student network receives a complete training frame sequence, while a random continuous subsequence is selected from this sequence as input to the bidirectional teacher network. Teacher output distillation and reconstructed feature distillation are performed only at the last frame of this continuous subsequence. Since the teacher network's input terminates at this last frame, the distillation supervision does not include information about future video frames after that moment, thus ensuring causal consistency between the teacher's supervisory knowledge and the information available to the unidirectional student network at the corresponding moment.

[0008] A core improvement of this invention lies in the random continuous subsequence final frame distillation mechanism. The student network's training input is a complete low-resolution frame sequence, from which a start and end position are randomly selected to form a continuous teacher subsequence. The teacher network processes only this continuous subsequence, extracting super-resolution output and reconstruction stage features at the final frame of the subsequence; the student network, meanwhile, propagates chronologically from the beginning of the complete sequence, extracting the corresponding output and reconstruction stage features at the same final frame position. At this point, the student network can utilize the complete historical propagation state, while teacher supervision does not include information beyond the final frame, thus avoiding the requirement for a one-way student to learn future information unavailable during the online inference stage. By randomly changing the position of the continuous subsequence during training, different time positions have the opportunity to become distillation endpoints.

[0009] Another core improvement of this invention lies in the combined use of teacher output distillation and reconstructed feature distillation. Teacher output distillation constrains the student network's final super-resolution result in the last frame of the random subsequence to approximate the teacher network's output; reconstructed feature distillation further constrains the intermediate representation formed by the student network in the high-resolution reconstruction stage to approximate the corresponding representation of the teacher network. When the channel dimensions of the reconstructed features of the teacher and student are inconsistent, the student's reconstructed features are dimensionally transformed through a channel mapping module. Compared with only constraining the final output, this method can simultaneously transfer knowledge from the result layer and knowledge from the reconstructed representation.

[0010] This invention proposes a Flow-Guided Pyramid Deformable Convolution Alignment (FGPCDA) module for unidirectional online video super-resolution student networks. This module first obtains motion priors for adjacent frames based on optical flow, then predicts deformable convolution offsets in a multi-scale feature space, and performs progressive adaptive correction on historical propagation features, thereby improving the cross-frame feature alignment accuracy of motion regions.

[0011] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0012] A knowledge distillation-based online video super-resolution method includes the following steps:

[0013] Step 1: Collect and construct a video super-resolution training dataset, and divide the training dataset into a training set and a test set;

[0014] The video super-resolution training dataset contains low-resolution video sequences. and high-resolution ground truth video sequences corresponding to low-resolution video sequences. M represents the number of frames in the video sequence;

[0015] Step 2: Construct the teacher network and complete the teacher model training; specifically, a two-stage training method is adopted, which first pre-trains on a large-scale video dataset and then fine-tunes it on the target scene dataset to obtain the teacher model for knowledge distillation.

[0016] The teacher network described is a bidirectional recursive video super-resolution network;

[0017] Step 3: Construct and configure the student network; specifically, receive low-resolution video sequences through the student network. As input, the corresponding high-resolution video sequence is output. The student network described above includes at least an optical flow estimation module, the FGPCDA module proposed in this invention, a forward propagation backbone, and a reconstruction head.

[0018] The student network described is a one-way online transmission network;

[0019] Step 4, Freeze the teacher network T parameters Only update the S-parameters of the student network. During training, the teacher network and the student network each output the final super-resolution results. and and intermediate features and When the number of intermediate feature channels in teacher network T and student network S is inconsistent, channel mapping is used. Perform dimensional alignment on the intermediate features of the student network;

[0020] The intermediate features include at least propagation features, alignment features, reconstruction input features, and reconstruction stage features;

[0021] Step 5: Using the last frame of the continuous subsequence as the distillation position, obtain the super-resolution output and reconstruction stage features of the teacher network and student network at this position; calculate the output distillation loss between the teacher output and the student output, and calculate the reconstruction feature distillation loss between the teacher reconstruction stage features and the student reconstruction stage features after channel mapping; the pixel-level reconstruction loss of the student network is calculated based on the student output of the complete training frame sequence and the high-resolution ground truth; update the student network parameters by combining the reconstruction loss, output distillation loss and reconstruction feature distillation loss.

[0022] Step 6: After training is complete, only the student network is retained for online video super-resolution inference, with low-resolution video frames input in chronological order. The system sequentially performs optical flow estimation, feature alignment, forward propagation state update, and reconstruction head output to generate high-resolution video sequences in real time. .

[0023] Furthermore, let the complete low-resolution frame sequence of the i-th training sample be... Randomly select a continuous subsequence Where 1 ≤ s ≤ e ≤ M, and e represents the last frame index of the continuous subsequence. The starting position and sequence length of the continuous subsequence are randomly determined during training.

[0024] Furthermore, the teacher network only receives during the knowledge distillation phase. As input, and using only the teacher output at the last frame e and the teacher reconstruction stage features as distillation supervision, the student network receives the complete frame sequence. .

[0025] Furthermore, the teacher network in step 2 is a bidirectional recursive network that uses forward and backward temporal information of the video sequence for feature fusion;

[0026] The pixel-level reconstruction loss L is used when training the teacher model. T :

[0027]

[0028] Where N represents the number of samples in a training batch; M represents the number of frames in the video sequence; and These represent the super-resolution output of the i-th sample in the m-th frame of the teacher model and the corresponding high-resolution ground truth frame, respectively. This represents a smoothing term (to avoid gradient instability).

[0029] Furthermore, the student network in step 3 is a unidirectional online propagation network that relies solely on the current frame and historical frame states for inference. Its feature channel size is smaller than that of the teacher network to meet the requirements of low-latency real-time inference.

[0030] Furthermore, in step 3, the optical flow estimation module is used to estimate the forward optical flow based on the current low-resolution video frame and adjacent historical low-resolution video frames;

[0031] In step 3, the FGPCDA module proposed in this invention is used to perform cross-frame alignment of historical propagation features based on the forward optical flow, the propagation features of the previous time step, and the shallow features of the current frame, and output the aligned historical features.

[0032] In step 3, the forward propagation backbone and reconstruction head are used to output the super-resolution result of the current frame based on the splicing features of the current frame features and the aligned historical features.

[0033] Furthermore, the channel mapping in step 4 is implemented using 1×1 convolution. Without changing the feature space resolution, the intermediate features of the student network are linearly transformed by channels to make them consistent with the channel dimensions of the corresponding intermediate features of the teacher network.

[0034] Furthermore, the loss function in step 5 includes the reconstruction loss L. rec Output distillation loss L out Reconstructing characteristic distillation loss L feat The knowledge distillation-related loss is calculated only at the last frame of the random continuous subsequence:

[0035]

[0036] in, and s and e represent the super-resolution output of the i-th sample in the student model and teacher model at the e-th frame, respectively; s and e represent the start and end frame positions of the random continuous subsequence, respectively, and the e-th frame is the distillation position.

[0037] Furthermore, in step 5, the characteristic distillation loss L is reconstructed. feat The intermediate features output by the teacher network and student network are extracted for reconstruction. After dimensional alignment of the student features by the channel mapping module, the reconstruction feature distillation loss between the student network and the teacher network is calculated to further constrain the intermediate representation formed by the student network in the high-resolution reconstruction stage to approximate the corresponding representation of the teacher network. The loss function is:

[0038]

[0039] in, and Let represent the reconstructed features extracted by the student network and the teacher network for the i-th sample at the e-th frame, respectively; This represents the video sequence from frame 1 to frame e input to the student network. This represents a continuous subsequence from frame s to frame e, randomly selected from the complete video sequence and input into the teacher network. The channel mapping module is used to map student features to the same channel dimension as the teacher.

[0040] Furthermore, the reconstruction features are intermediate representations of the teacher network and student network before or within the reconstruction phase, used to generate the final super-resolution result.

[0041] Furthermore, the alignment module in the student network adopts the FGPCDA module proposed in this invention. Specifically, it receives the propagation features of the previous time step, the features of the current frame, and the forward optical flow output by the optical flow estimation module, and performs step-by-step alignment of the propagation features of the previous time step at multiple scales, and outputs historical features aligned with the current frame.

[0042] Furthermore, the FGPCDA module first uses the optical flow of the corresponding scale to warping pre-align historical features at each scale, then predicts the deformable convolution offset based on the pre-aligned historical features and the scale features of the current frame, and uses the deformable convolution to perform adaptive sampling correction on the pre-alignment result; wherein, the offset information and alignment features predicted at lower scales are upsampled and then passed to higher scales to achieve progressively refined alignment.

[0043] Compared with the prior art, the present invention has the following advantages:

[0044] (1) To address the problem of future information asymmetry between bidirectional teachers and unidirectional students, this invention uses random continuous subsequence last frame distillation to truncate the teacher input at the distillation time, so that the teacher distillation target does not contain future frame information after that time, thereby improving the consistency of teacher supervision and online information conditions between unidirectional students.

[0045] (2) Simultaneously performing teacher output distillation and reconstruction feature distillation in the last frame of the random subsequence not only constrains the final super-resolution result, but also transfers the detailed representation formed by the teacher network in the reconstruction stage, which is beneficial to improving the reconstruction quality of the lightweight student network.

[0046] (3) The present invention proposes an FGPCDA module, which combines optical flow motion prior with multi-scale deformable sampling to perform step-by-step alignment and adaptive correction of historical propagation features, thereby improving the cross-frame feature misalignment problem in complex motion regions and enhancing the effective utilization of historical time sequence information.

[0047] (4) Knowledge distillation is used only during the training phase. During deployment, the teacher network, channel mapping, and other training branches used only for distillation are removed, leaving only a lightweight unidirectional student network. Therefore, the online inference overhead of the teacher network is not introduced, which enables the student network to achieve better super-resolution reconstruction capabilities while maintaining low latency and easy deployment characteristics. This meets the dual requirements of real-time performance and image quality in online video scenarios, achieving a balance between performance and efficiency in the overall technical approach. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 Overall framework diagram of the online video super-resolution method of this invention;

[0050] Figure 2 A schematic diagram of an optically flow-guided pyramid-shaped deformable convolution alignment module.

[0051] Figure 3 Comparison of local super-resolution results for REDS4 sequence 000;

[0052] Figure 4 Comparison of local super-resolution results for REDS4 sequence 011;

[0053] Figure 5 Comparison of local super-resolution results of REDS4 sequence 015;

[0054] Figure 6 Comparison of local super-resolution results for REDS4 sequence 020. Detailed Implementation

[0055] To gain a deeper understanding of this invention, we will provide a comprehensive and detailed description. However, this invention has various implementations and is not limited to the specific examples listed herein. These examples are presented to enhance a full understanding of the disclosure of this invention.

[0056] Example 1: Network Construction and Knowledge Distillation Training

[0057] Step 1: Collect low-resolution video sequences and their corresponding high-resolution ground truth video sequences to construct training and testing sets. In this embodiment, the REDS dataset is used as the target dataset for both training and validation.

[0058] Step 2: Design and train a bidirectional recursive teacher network (e.g., BasicVSR), with a larger parameter scale than the student network. The teacher network is obtained through a two-stage training process: first, pre-training on large-scale video data (e.g., Vimeo-90K) to learn general spatiotemporal priors; then fine-tuning on the target dataset (e.g., REDS) to obtain the target domain teacher weights. Teacher training uses pixel reconstruction loss.

[0059]

[0060] in, and represents the super-resolution output of the i-th sample in the t-th frame of the teacher model and the corresponding high-resolution ground truth frame, respectively; N represents the number of samples in a training batch; M represents the number of frames in the video sequence; This represents a smoothing term (to avoid gradient instability). In this embodiment, the training segment length is set to M = 15 frames; the training clipping size is 256×256; flipping and rotation enhancements are used to improve generalization ability; the Adam optimizer is used, with parameters set to... = 0.9, =0.99, weight decay=0; initial learning rate 2e-4; cosine annealing learning rate scheduling strategy is adopted, with a minimum learning rate of 1e-7. The optimizer, learning rate, and scheduling parameters can be adjusted equivalently according to the training data size and hardware resources. In the subsequent distillation stage, the teacher parameters are frozen and used only as a source of supervision signals.

[0061] Step 3 involves constructing a unidirectional online student network. Parameter compression is achieved by removing reverse branches and pruning channel depth. The proposed FGPCDA module is introduced to improve cross-frame alignment accuracy, thereby meeting the requirements of online real-time inference while ensuring reconstruction quality. Without departing from the technical concept of this invention, the parameter compression can also be achieved through, for example, depthwise separable convolution, structured pruning, quantization, or a combination thereof.

[0062] Step 4: Perform knowledge distillation to train the student network. For a complete training segment of length M, the student network receives the segment and propagates it unidirectionally in chronological order. Simultaneously, consecutive subsequences are randomly extracted from the same complete training segment. This is used as input to the teacher network. The teacher network only processes this subsequence and adds it to its last frame x. e The teacher super-resolution output and teacher reconstruction stage features are generated for distillation.

[0063] Because the teacher's input sequence is in x e The process terminates at this point; the teacher does not have x at this distillation position. e The subsequent input frames; the student network in x e The position output is from x1 to x e The current and historical information is generated. Therefore, the teacher supervision in this final frame does not introduce future frame information after the endpoint that is not available to students during online reasoning. Students can also learn from the high-quality reconstruction results formed by the teacher under conditions where temporal information is available by utilizing a longer historical state accumulated from the start position of the complete segment.

[0064] Step 5: Calculate the joint loss and update the student network. If the teacher and student reconstruct different numbers of feature channels, use a 1×1 convolution φ(·) to complete the channel mapping. The total loss is:

[0065]

[0066] Step 6: After training, only the student network is retained for inference. The actual online video input enters the student network frame by frame in chronological order. The features propagated from the previous time step are aligned to the current frame by the FGPCDA module proposed in this invention and fused with the features of the current frame. The reconstruction head then outputs the high-resolution result of the current frame. The teacher network and the channel mapping module used for distillation do not participate in deployment and inference.

[0067] Example 2: REDS4×4 Performance Test

[0068] The baseline student network represents a lightweight one-way student model that does not use the FGPCDA module proposed in this invention and does not employ knowledge distillation; "only FGPCDA" means that the FGPCDA module proposed in this invention is added to the baseline student network, but teacher knowledge distillation is not employed; the complete method of this invention employs both FGPCDA and knowledge distillation.

[0069] Table 1 Test results on REDS4×4

[0070]

[0071] Example 3 Comparison with the disclosed method

[0072] To further illustrate the reconstruction performance of the present invention on the REDS4×4 video super-resolution task, the method of the present invention is compared with methods disclosed in recent years.

[0073] Table 2 Comparison of REDS4×4 results with publicly available video / image super-resolution methods

[0074]

[0075] Example 4: Visual Effect Comparison

[0076] To visually verify the reconstruction effect of the method of this invention in typical detail regions, four test sequences 000, 011, 015, and 020 of REDS4 were selected for local visualization comparison. Each figure shows the low-resolution input (LR) and local region, and compares the high-resolution ground truth (GT), Bicubic interpolation, EDVR, baseline student network, student network with the proposed FGPCDA module, and the complete method with further knowledge distillation at the same location. All methods are compared within the same frame and the same local region.

[0077] The contents not described in detail in this specification are conventional techniques that can be understood and implemented by those skilled in the art. The above embodiments are only used to illustrate the technical concept of this invention and are not intended to limit the scope of protection of this invention; without departing from the technical concept of this invention, those skilled in the art can make equivalent adjustments or substitutions to the teacher network type, student network channel size, sequence length, loss weight, and specific implementation of the alignment module.

Claims

1. An online video super-resolution method based on knowledge distillation, characterized in that, Includes the following steps: Step 1: Collect and construct a video super-resolution training dataset, which includes low-resolution video sequences and corresponding high-resolution ground truth video sequences. Step 2: Construct and train the teacher network, which is a bidirectional recursive video super-resolution network capable of feature fusion using forward and backward temporal information; Step 3: Construct a student network, which is a unidirectional online propagation network that relies solely on the current frame and historical frame states for inference. It includes at least an optical flow estimation module, an optical flow-guided pyramid deformable convolution alignment module, a forward propagation backbone, and a reconstruction head. Step 4: Freeze the teacher network parameters and perform knowledge distillation training on the student network; For a complete training frame sequence of length M, the student network receives the complete training frame sequence and propagates it unidirectionally in chronological order; Select a continuous subsequence from the same complete training frame sequence as the input to the teacher network, let the starting frame position of the continuous subsequence be s and the ending frame position be e, where 1≤s≤e≤M; The teacher network only processes the continuous subsequence and obtains the teacher super-resolution output and teacher reconstruction features at the last frame e of the continuous subsequence, and the student network obtains the student super-resolution output and student reconstruction features at the corresponding e-th frame; Step 5: Calculate the output distillation loss between the teacher's super-resolution output and the student's super-resolution output at the e-th frame, and calculate the reconstruction feature distillation loss between the teacher's reconstructed features and the student's reconstructed features; the pixel-level reconstruction loss of the student network is calculated based on the student output of the complete training frame sequence and the corresponding high-resolution ground truth; the pixel-level reconstruction loss, output distillation loss and reconstruction feature distillation loss are weighted and combined to update the student network parameters; Step 6: After training is complete, remove the teacher network and the training branch used only for knowledge distillation, and keep only the student network for online video super-resolution inference. Input low-resolution video frames in chronological order, and perform optical flow estimation, historical feature alignment, forward propagation state update and high-resolution reconstruction in sequence to generate the corresponding high-resolution video sequence.

2. The online video super-resolution method based on knowledge distillation according to claim 1, characterized in that, The teacher network is obtained by first pre-training on a video dataset and then fine-tuning on a target scene dataset. The teacher network training adopts the pixel-level reconstruction loss between the teacher's super-resolution output and the corresponding high-resolution ground truth.

3. The online video super-resolution method based on knowledge distillation according to claim 1, characterized in that, The feature channel size of the student network is smaller than that of the teacher network; the optical flow estimation module is used to estimate the forward optical flow based on the current low-resolution video frame and adjacent historical low-resolution video frames; the forward propagation backbone is used to fuse the current frame features with aligned historical propagation features; the reconstruction head is used to generate the super-resolution result of the current frame.

4. The online video super-resolution method based on knowledge distillation according to claim 1, characterized in that, The starting frame position and / or sequence length of the continuous subsequence are randomly determined during training, so that different time positions in the complete training frame sequence can be used as knowledge distillation positions.

5. The online video super-resolution method based on knowledge distillation according to claim 1, characterized in that, The knowledge distillation position is set to the last frame e of the continuous subsequence input by the teacher network, so that the supervision information of the teacher network at the distillation position does not include video frame information after the e-th frame; the student network obtains the corresponding output and reconstructed features at the e-th frame by using the historical propagation state accumulated from the start position of the complete training frame sequence to the e-th frame.

6. The online video super-resolution method based on knowledge distillation according to claim 1, characterized in that, When the channel dimensions of the reconstructed features of the teacher network and the student network are inconsistent, a channel mapping module is used to align the dimensions of the reconstructed features of the student network. The channel mapping module uses 1×1 convolution to map the reconstructed features of the student network to the same channel dimensions as the reconstructed features of the teacher network without changing the feature space resolution.

7. The online video super-resolution method based on knowledge distillation according to claim 1, characterized in that, The reconstruction features are intermediate representations used by the teacher network and student network to generate the final super-resolution result before or within the high-resolution reconstruction stage; the output distillation loss and reconstruction feature distillation loss are both calculated only at the last frame e of the continuous subsequence.

8. The online video super-resolution method based on knowledge distillation according to claim 1, characterized in that, The optical flow-guided pyramid deformable convolutional alignment module receives the propagation features from the previous time step, the features of the current frame, and the forward optical flow output by the optical flow estimation module. It then performs step-by-step alignment of the propagation features from the previous time step at multiple scales and outputs historical features aligned with the current frame.

9. The online video super-resolution method based on knowledge distillation according to claim 8, characterized in that, The optical flow-guided pyramid deformable convolution alignment module pre-aligns historical propagation features at each scale using the corresponding scale's optical flow, predicts the deformable convolution offset based on the pre-aligned historical propagation features and the corresponding scale features of the current frame, and uses deformable convolution to adaptively sample and correct the pre-alignment results. The offset information and alignment features obtained at lower scales are upsampled and then passed to higher scales to achieve progressively refined alignment.

10. The online video super-resolution method based on knowledge distillation according to claim 1, characterized in that, The teacher network, channel mapping module, and other training branches used solely for knowledge distillation are only used during the training phase and do not participate in computation during the deployment and inference phase.

Images