Learning Summary-Worthy Visual Representation for Abstractive Summarization in Video

1 Introduction

With the increasing popularity of video in user-generated content in recent years [Kim et al., 2021; Cherian et al., 2022], a large number of open-domain videos have been posted on the Web (e.g., YouTube). Usually, many of the videos are not accompanied by a briefly introduction to reflect the corresponding salient information, which prevents users from quickly finding their interested videos unless taking time to peruse each video. Obviously, in this scenario, it would be

∗ Work is done during internship at Noah's Ark Lab.

† Corresponding author.

Video Frames:

Transcript：so, now we 're going to talk about the backhand. basical ly, the same a grip. sometimes i feel like i want to adjust it [...] you 're hanging your arm, trying to hang your arm almost perpendicular to the ground. almost like a tennis stroke. [...] and just follow through, you can come through a little high, but basically, you want to get your arms back into the ready position. we 'll talk about that next.

Reference Summary： learn how to do the backhand stroke in ping pong in this free video lesson on ping pong .

Figure 1: An example of multimodal abstractive summarization. The unimportant textual content is omitted and replaced by [...]. It can be seen that some critical information in the reference is either highlighted in the transcript (e.g., backhand stroke) or only available from the video (e.g., ping pong).

valuable to develop an automatic abstractive summarization model [Libovick`y et al., 2018] that detects the highlights of each video and then generates a short textual description.

As illustrated in Figure 1, the task of multimodal abstractive summarization (MAS) aims to generate a concise textual summary according to multimodal resources, i.e., the video content and its transcript [Sanabria et al., 2018]. This task is challenging since both visual and textual modalities are complementary to each other, and thus, how to efficiently combine the multimodal information is the key to this task. To leverage information from both modalities, [Palaskar et al., 2019] employed separate encoders on visual and textual data, which was followed by a joint decoder with an attention mechanism to capture the intrinsic connection between the two modalities. Later, MFFG [Liu et al., 2020] and SFFG [Liu et al., 2022] brought the multimodal interaction into the encoder to obtain the fine-grained correlation between multimodal inputs to exploit the complementary information of each modality and achieved promising results.

Recently, inspired by the success of the large-scale generative pre-trained language model (GPLM) [Lewis et al., 2019; Raffel et al., 2020] on generating high-quality textual content (e.g., summary), researchers start to apply it to the MAS

task. To do this, VG-GPLM [Yu et al., 2021a] first extracts the video representation with the pre-trained visual feature extractor. Then, to allow the cross-modal interaction, the obtained visual feature of video is injected into each encoder layer of the GPLM with an attention-based text-vision fusion module. This attention module is designed to select the relevant visual information (key & value) based on the textual features (query). VG-GPLM has demonstrated the potentials of injecting visual representation into language models to improve the MAS performance.

However, existing methods mainly use a general-purpose visual feature extractor, which may potentially overlook some summary-worthy visual information, including the essentials that are merely available in the video (e.g., the 'ping pong' in Figure 1 only appears in video frames), and thus impede the quality of the generated summary. It is also possible for the existing methods to generate the novel concepts like 'ping pong' if sufficient labelled data is provided to welltrain the general-purpose extractor to identify task-specific objects. Unfortunately, it may lead to the contradiction between the data hungriness issue of the extractor and the scarcity of annotated summarization data.

To address the above issue, we aim to design a visual encoder that is aware of summary-worthy information. We present a novel method, named 𝑆𝑊𝑉𝑅, to learn SummaryWorthy Visual Representation for the MAS task. We use a Bidirectional Visual-Language Attention mechanism (BVLA) to encourage the visual encoder to exploit the summary-wothy information from the textual data, i.e., transcripts. Afterward, we further introduce a Self Distillation Mechanism (SDM) that takes the generated pseudo summary as the teacher to guide the learning of the visual encoder. This self-distillation encourages the visual encoder to have the ability of capturing and aligning to summary-worthy knowledge that appears in the generated summary after the decoder. Given the surprising and strong text generation ability of GPLM, such distillation helps the visual encoder pay more attention on corresponding key concepts in video frames. Since a sub-optimal placement may impede the model's performance, we enumerate almost all possible ways to combine and evaluate BVLA and SDM modules in the visual encoder. Experimental results on three public datasets show that our method outperforms all competing baselines, especially on the dataset with a relatively small size. Further studies demonstrate the effectiveness of each component, and suggest that the learned summary-worthy feature can help the model identify valuable information, thus benefiting the MAS task.

2 RelatedWork

Text-based Abstractive Summarization: Given a long article, abstractive text summarization aims to generate a brief summary that describes the article's most salient content. Early studies were based on statistical or linguistic rules, including extractive and compression [Knight and Marcu, 2002; Clarke and Lapata, 2010], templates [Zhou and Hovy, 2004], and statistics [Banko et al., 2000]. Later, the availability of large-scale summarization corpora has promoted the development of various neural networks, among which the representative Seq2Seq model [Sutskever et al., 2014] and the attention mechanism have greatly advanced the quality of summaries [Paulus et al., 2018; Wang et al., 2019; Zhang et al., 2020; Yu et al., 2021b]. Recently, in light of the powerful generative abilities, the large-scale pre-trained language models have led the mainstream in this field [Lewis et al., 2019; Raffel et al., 2020; Zhang et al., 2020; Qi et al., 2020].

Video-based Abstractive Summarization: Multimodal summarization has been developed for decades [Erol et al., 2003; Tjondronegoro et al., 2011; Evangelopoulos et al., 2013; Shah et al., 2016; Zhang et al., 2022]. The key method behind it, i.e., multimodal learning, has recently attracted a number of researchers' interest, while in fact little attention has been paid to the video content based summarization. Previous methods mainly focus on a simple situation where the video data contains synchronized signals, e.g., synchronized voice and captions. To tackle the video-based summarization in a more general and asynchronous scenario, [Li et al., 2017] collected a multimodal dataset containing 500 videos of English news articles with human-generated reference summaries. Later, to better promote the development of the MAS for videos, [Sanabria et al., 2018] introduced a large-scale human-annotated video dataset named How2, which contains videos of 2000 hours, and each video is annotated with a short summary. Thanks to the How2 dataset, of which the advent has accelerated the development of MAS methods using neural networks, e.g., the hierarchical attention in [Palaskar et al., 2019], the forget gate mechanism in [Liu et al., 2020], and the multi-stage fusion network in [Liu et al., 2022]. To further leverage the GPLM's generation ability, VG-GPLM [Yu et al., 2021a] studies multiple fusion methods that inject the visual information into GPLMs to improve the MAS for videos. However, VG-GPLM obtains the visual feature via a generalpurpose visual encoder, which likely ignores task-specific visual clues that are valuable to summarization. In contrast, our method bridges this gap by learning and then injecting the summary-worthy visual feature into the GPLMs.

3 Methodology

3.1 ProblemDefinition

Given a video 𝑉 and its associated textual transcript 𝑇, our multimodal summarization system is required to summarize the video by a generated concise summary 𝑆 with maximum probability 𝑝(𝑆|𝑉, 𝑇; 𝜃), where 𝜃 stands for the model parameters. It is worth knowing that both the transcript 𝑇 and generated summary 𝑆 are in the form of a sequence of words.

For the text-based abstractive summarization task, a prevailing way is to adopt the powerful GPLM, whose ability of generating high-quality texts has been widely demonstrated [Qi et al., 2020; Zhang et al., 2020]. When in the case of multimodal input like a video, a natural idea is to inject the visual features into the encoder of GPLM and then utilize its superior generation to obtain a better textual summary, which is how VG-GPLM [Yu et al., 2021a] works. In the following, we present the proposed 𝑆𝑊𝑉𝑅 model in detail, and the overall framework is shown in Figure 2.

3.2 Video Pre-processing

Given a video, we follow previous work [Yu et al., 2021a] to first extract the visual features with 2048 dimension for every 16 non-overlapping frames through a 3D ResNeXt-101 model [Hara et al., 2018], which is pre-trained on the Kinetics dataset [Kay et al., 2017]. Then a linear layer is adopted to project the visual features into 𝑑 dimension vector space and obtain Z𝑣 ∈ R 𝑛𝑣×𝑑 , where 𝑛𝑣 is the number of visual tokens.

3.3 Method Overview

As discussed above, we employ the GPLM as our backbone. Specifically, we construct our method with the sequence-tosequence BART model [Lewis et al., 2020], which consists of a textual encoder and a left-to-right decoder designed to generate the textual summary. Given the transcript as input, the textual encoder first tokenizes it and then embeds it into the textual features Z𝑡 ∈ R 𝑛𝑡 ×𝑑 , which will be fed into the transformer [Vaswani et al., 2017] encoder layers to exploit the contextual representation as:

where 𝐿𝑁 denotes the layer normalization [Ba et al., 2016], 𝑙 ∈ [1, 𝐿] denotes the 𝑙-th textual encoder layer, Z0 𝑡 is initialized with Z𝑡 , and 𝑀𝑢𝑙𝑡𝑖𝐴𝑡𝑡𝑛 and 𝐹𝐹𝑁 denotes two sublayers, i.e., the multi-headed self-attention mechanism and a feed-forward network, respectively1.

To inject the visual information into each layer of the textual encoder, we add a third sub-layer into each textual encoder layer. Specifically, assume the visual feature that is to be injected into the 𝑙-th textual encoder layer is Z𝑙 𝑣, then this sub-layer operates as:

where W 𝑙 𝑡 ∈ R 𝑑×𝑑 is themodelweight, and𝑉𝐿𝐴(Z𝑙 𝑣 , Z𝑙 𝑡 ) is a unidirectional visual-language attention function that extracts the relevant information from visual feature Z𝑙 𝑣 according to the textual feature Z𝑙 𝑡 , which will be described later. It can be seen from (3) that Z𝑙 𝑣 takes a great impact on our model. Therefore, our paper focuses on how to obtain a high-quality collection of visual feature, i.e., C𝑣 = {Z1 𝑣, · · · ,Z 𝐿 𝑣 }, where each Z𝑙 𝑣 can be injected into the 𝑙-th textual encoder layer.

Afterward, the output ˜Z 𝐿 𝑡 from the last layer of textual encoder will be provided to the decoder to generate a sequence (w1, . . . ,w𝑛𝑤 ) of vectors one element at a time. Since the decoder is auto-regressive, at each step, the generated vector w𝑖 is mapped to a new vector of vocabulary size, followed by a softmax function to output the summary word distribution ˆy𝑖 . The negative log-likelihood between the generated summary and the ground-truth summary is used to calculate the loss as:

3.4 Learning Summary-Worthy Visual Representation

A simple way to generate the collection C𝑣 is that we fill C𝑣 with the obtained Z𝑣 after video pre-processing, namely, C𝑣 = {Z𝑙 𝑣 = Z𝑣, ∀𝑙 = 1, · · · , 𝐿}. In thisway, the cross-modal interaction can be simply summarized as Figure 3(a).

Auxiliary Visual Encoder

However, the direct utilization of Z𝑣 may suffer from the below shortcomings: (𝑖) The temporal information of a video is ignored, which limits the expressiveness of visual feature; (𝑖𝑖) Since recent study [Raghu et al., 2021] finds that different encoder layers exploit different-level semantic, injecting the same visual feature into different textual encoder layers may potentially restrict the model to learn more meaningful crossmodal interactions.

1Additionally, there are residual operations inside each layer, which we omit here for brevity. For the interested readers, we kindly refer to [Vaswani et al., 2017] for more description.

Figure 3: The simple illustration of different types of visual information injected into the textual encoder layers. SWR stands for summary-worthy representation. The textual and visual information are represented in green and blue color, respectively. The yellow color stands for the feature map of generated summary.

To address the above issues, also inspired by the recent success of the two-tower architecture [Radford et al., 2021; Jia et al., 2021], we introduce an auxiliary visual encoder that takes the video feature as input and learns visual information at different levels of semantics. Specifically, the trainable positional encodings are first added to the Z𝑣 to preserve the temporal information. Afterward, this Z𝑣 can be used directly as input to vanilla transformer-based visual encoder of 𝐿 layers. Formally, given the input Z𝑣, the output of each visual layer is computed as:

where Z0 𝑣 is initialized with Z𝑣. We collect the output of each visual encoder layer and obtain a new visual collection C𝑣 = {Z1 𝑣, · · · ,Z 𝐿 𝑣 }, as shown in Figure 3(b). In this way, we inject the visual information into the textual feature from the same 'height' (e.g., both from the 𝑙-th layer), which could benefit the cross-modal correlation at different levels of semantics.

Summary-Worthy Information from Transcript

It can be observed that this introduced auxiliary visual encoder is still a general-purpose visual feature extractor. Since not all the information from the visual modality is valuable for summarization [Liu et al., 2020], the noise and redundancy in the generally learned visual feature makes it less effective for the model to summarize the highlights. We are thus motivated to propose to learn the summary-worthy visual representation.

Since the summary-worthy information from each modality may be complementary to the other, carrying this observation, we propose to exploit the summary-worthy information from the transcript, which may be complementary to and benefit learning the visual feature. Therefore, we use the multihead cross-attention to inject the summary-worthy information from the transcript into the visual representation. Technically, we extend the unidirectional 𝑉𝐿𝐴(·, ·) function (which is employed in (3)) into bi-directional visual-language attention function 𝐵𝑉𝐿𝐴(·, ·). Given textual feature Z𝑙 𝑡 ∈ R 𝑛𝑡 ×𝑑 and visual representation Z𝑙 𝑣 ∈ R 𝑛𝑣×𝑑 , the 𝐵𝑉𝐿𝐴(·, ·) will function as:

where Z𝑙 𝑡→𝑣 ∈ R 𝑛𝑣×𝑑 and Z𝑙 𝑣→𝑡 ∈ R 𝑛𝑡 ×𝑑 denote the information of transcript that is relevant to video or vice versa, respectively. Let us take Z𝑙 𝑡→𝑣 as an example to elaborate. Particularly, the Z𝑙 𝑡→𝑣 is calculated via a cross-modal multihead attention (𝐶𝑟𝑜𝑠𝑠𝑀𝑢𝑙𝑡𝑖𝐴𝑡𝑡𝑛) as:

where W 𝑙 𝑞, W 𝑙 𝑘 and W𝑣 ∈ R 𝑑×𝑑 are model weights. After obtaining the Z𝑙 𝑡→𝑣 , we inject it into the visual feature as

where W ′𝑙 𝑣 ∈ R 𝑑×𝑑 is the model weight. Similarly, the output of each visual encoder layer can be collected to form C𝑣 = { ˜Z1 𝑣, · · · , ˜Z 𝐿 𝑣 }, as like Figure 3(c).

Summary-Worthy Information from Summary

Since some vital summary concepts are only available in the video content rather than transcripts (e.g., the 'ping pong' in Figure 1), we also expect the visual encoder could have the ability to identify the novel summary-worthy visual feature that might not appear in transcripts. We are thus motivated to bridge the visual feature learning and the generated summary in a more direct manner, thus further enhancing our visual encoder.

One potential way of benefiting our visual encoder is to distil the task-specific knowledge from other large-scale summarization models. However, most of the available ones are textonly methods, of which the knowledge is not suitable for our visual encoder. To this end, instead of the traditional knowledge distillation, we take advantage of self distillation [Zhang et al., 2019], which has been proven to be an effective way to distil knowledge within the network itself. The basic idea is that we take the output of the decoder as the pseudo summary (i.e., hints) to teach the learning process of the visual encoder.

Technically, we apply the average pooling on the output vector sequence (w1, . . . ,w𝑛𝑤 ) from the last layer of the decoder to obtain w ∈ R 𝑑 , which can be regarded as the global representative of the generated pseudo summary. Then we distil this summary-worthy knowledge into our visual encoder layer via a self distillation mechanism (SDM) as

where 𝜆 is the hyper-parameters, 𝑎𝑣𝑔(·) is the average pooling layer. As seen from (11), this knowledge distillation works by decreasing the distance between the feature maps of the pseudo summary from the decoder and of the visual feature from the visual encoder. Since the feature maps are from two modalities with different dimensions, an extra linear projector is added to project the visual one to the same dimension as the other. Finally, the training loss L of our method is a sum of the objectives as:

During the training, as the visual feature in our SDM can gradually fit the feature map of the decoder output in a global representative way, the inexplicit knowledge and novel concepts that have not appeared in transcripts are injected into learning the visual feature, and thus we achieve a new summary-worthy collection C𝑣 = { ˆZ1 𝑣, · · · , ˆZ 𝐿 𝑣 }, see Figure 3(d).

3.5 Implementation Details

The BART-base model is adopted as the backbone of our model, in which 𝐿 = 6 for both the encoder and decoder. For the introduced auxiliary visual encoder, we use a 6-layer encoder with 8 attention heads and a 768 feed-forward dimension. Following previous work [Yu et al., 2021a], we set the max length of the generated summary to be 64 tokens; the decoding process can be stopped early if an End-of-Sequence (EOS) token is emitted. The Adam [Kingma and Ba, 2014] with 𝛽1 = 0.9, 𝛽2 = 0.999, and a weight decay of 1𝑒 -5 is employed as the optimizer.

4 Experiments

4.1 Experimental Setups

Datasets and Evaluation Metrics: We evaluate the proposed 𝑆𝑊𝑉𝑅 on three public datasets, including How2, How2-300 [Sanabria et al., 2018], and MM-AVS [Fu et al., 2021] dataset. The statistic of datasets is shown in Table 1. We follow [Yu et al., 2021a] to adopt the ROUGE-{1,2,L} [Lin and Hovy, 2003], BLEU-{1,2,3,4} [Papineni et al., 2002], METEOR [Denkowski and Lavie, 2011], CIDEr [Vedantam et al., 2015], and Content F1 [Palaskar et al., 2019] as the evaluation criteria.

Bacelines: We compare 𝑆𝑊𝑉𝑅 with the following two groups of baselines. 1) Methods using transcript only: In this group of baselines, we pick the existing text-only summarization methods, including S2S [Luong et al., 2015], PG [See et al., 2017], TF [Vaswani et al., 2017], T5 [Raffel et al., 2020], and BART [Lewis et al., 2020]. 2) Methods using both transcript and video: We consider strong MAS baselines including HA(RNN/Transformer) [Palaskar et al., 2019], MFFG(RNN/Transformer) [Liu et al., 2020], VG-GPLM(T5/BART) [Yu et al., 2021a], and SFFG [Liu et al., 2022].

4.2 Experimental Results

The results of How2 datasets are shown in Table 2. As observed, our proposed model significantly outperforms all previous methods, and we take an average improvement of 5.72% over all the criterion compared with the strongest baseline VG-BART. It confirms the effectiveness of our proposed summary-worthy mechanism. We can also see that the performance of abstractive summarization with the help of both transcripts and video content significantly outperform transcriptonly summarization models, demonstrating that visual features contain valuable information that is complementary to the transcript.

As for the How2-300 dataset, it can be observed from Table 3 that our model still performs the best. An interesting found is that, compared with the VG-BART model, the average improvement 4.9% (over R1, R2, and RL) on How2-300 is larger than that of 2.2% in the How2 dataset, namely, our model can work even better on small datasets like How2-300. We conjecture that it is because the general-purpose visual feature extractors require more training data to distinguish the task-specific clues that are valuable for the MAS task, so they don't work very well on small datasets. In contrast, thanks to the superiority of our proposed BVLA and SDM, the key and novel information is easier to capture and align, therefore achieving much better performance on the smaller dataset.

To further elaborate on the advantage of our model in the case of small datasets, we conduct experiments on MM-AVS, which has the minimum amount of data among benchmark datasets. We select T5, BART, VG-T5, and VG-BART for comparison. As shown in Table 4, if we directly fine-tune the baselines on this dataset, severe over-fitting phenomena occurs in both the VG-T5 and VG-BART models. It can be further observed that the above two models can only coverage when first training on the larger How2 dataset, and then continuously fine-tuning on the smaller MM-AVS dataset. This can be attributed to the fact that it is hard for the baseline models to disentangle enough summary-worthy information from the general visual features if limited training data is provided. Instead, with the proposed summary-worthy mechanism, our model works surprisingly well on MM-AVS, no matter whether the model is first trained on the larger How2 dataset or not.

4.3 Ablation Study

Impacts of Different Components

We evaluate the effectiveness of different components by gradually removing three modules, i.e., the SDM, BVLA, and auxiliary visual encoder. 1) SWVR w/o D: the self distillation is first removed; in such case, our model can only leverage the summary-related information from transcripts and becomes Fig.3(c). 2) SWVR w/o D+B: The BVLA is further replaced with the uni-directional VLA, then our model can be summarized as Fig.3(b). Note that in this case, our model is not aware of any summary-worthy information. 3) SWVR w/o D+B+A: Lastly, the introduced auxiliary visual encoder is dropped, and our model is degraded to the simplest form Fig.3(a).

We compare the three variants with the complete 𝑆𝑊𝑉𝑅 on the How2 dataset, and results are shown in Table 2. The general trend is that eliminating any of the modules (e.g., from SWVR w/o D to w/o D+B+A) will negatively impact the model's performance, confirming each component's benefits. Moreover, it is noticeable that the introduction of SMD can bring an improvement of 1.25% on average, suggesting the importance of equipping our visual encoder with the ability to capture summary-worthy values from the pseudo summary.

Impacts of the Hyper-parameter of 𝜆

In SDM, we introduce the hyper-parameter 𝜆, which controls the trade-off between the cross-entropy loss and the self distillation loss. To analysis the sensitivity of 𝜆, we manually select the values of 𝜆 from {0.01, 0.05, 0.1, 0.2, 0.5}. ROUGE-1 and BLEU-1 w.r.t 𝜆 on How2 datasets are illustrated in Figure 4. It can be observed that the performance of 𝑆𝑊𝑉𝑅 first increases and gets the peak when 𝜆 is not greater than 0.1. Afterward, the improvement is neutralized and even lost if 𝜆 becomes larger. This may be attributed to the fact that a simple linear layer in (11) may not mitigate the difference between different modalities, and thus, we speculate that a more well- Video：

Transcript：okay , now we do the same thing we just did with the ground strokes . we do the extreme angle volley [...] out of each shot as you can because then your practice is valuable . [No mention about tennis]

Ours w/o D：practice angle drill to practice a backhand with an extreme angle drill . learn how to do a technical drill in this free video lesson .

Ours：work on tennis volleys with an extreme angle drill . learn how to do technical tennis drills that will improve your control in this free video lesson .

Figure 5: An example of case study. We show the generated summaries of model using summary-related and summary-worthy visual feature, respectively.

designed transformation may further discover the poentials of self distillation, which however, is beyond the scope of this paper.

4.4 Case Study

We conduct a case study to empirically exhibit the effectiveness of using summary-worthy visual information. To do so, we conduct experiments using two cases: one only exploits summary-worthy information from the transcript (see Fig.3(c)), while the other exploits summary-worthy information from both the transcript and the generated pseudo summary (Fig.3(d)). Data samples can be seen in Figure 5, the former case of only leveraging the summary-worthy values from the transcript fails to predict the 'tennis', which concept is only available in video frames. With the assistance of distilled knowledge from the pseudo summary, our model captures the novel concept, or we call it visual object (i.e., tennis), from the video frames and successfully generates a high-quality summary result.

4.5 Where to Adopt the BVLA and SDM

In this section, we aim to further investigate the optimal location where to insert the BVLA and SDM modules. As depicted in Table 5, in general, leveraging BVLA to inject the cross-modal features into the unimodal representation can significantly boost the model's performance. A similar phenomenon of the utilization of SDM can also be observed in Table 6. Furthermore, we observe that inserting the BVLA or SDM modules to the top layers can achieve the best performance. We speculate that the lower layers of the encoder may tend to capture the local and low-level semantics, which are usually modality-specific. This makes it challenging for the model to capture better cross-modal semantic interaction. In contrast, the top layers of the encoder may primarily exploit the global and high-level semantics, where the high-level abstract information is more accessible for the model to perform cross-modal interaction. This observation is also consistent with [Xu et al., 2021]. Given this conclusion, our best model reported in Table 2 adopts the BVLA and SDM in {4,5,6}-th and {5,6}-th encoder layers, respectively.

5 Conclusion

In this paper, we propose to learn summary-worthy visual features to boost the MAS task. We introduce a bi-directional visual-language attention (BVLA) mechanism and a self distillation mechanism (SDM) to encourage the visual encoder to exploit the summary-worthy information from the textual data or the knowledge that distills from the pseudo summary. In addition, we enumerate almost all possible ways to combine and evaluate BVLA and SDM modules in the visual encoder. Experiments results show that our proposed method significantly outperforms all strong baselines on three public datasets. Further analysis demonstrates each component's effectiveness and suggests that the higher layers of the visual encoder are the optimal places to employ the BVLA and SMD.

Acknowledgments

We are grateful to ShuTing Cai for perfecting this article. This work is supported by the National Natural Science Foundation of China (No. 62276280, U1811264), Key R&D Program of Guangdong Province (No. 2018B010107005), Natural Science Foundation of Guangdong Province (No. 2021A1515012299), Science and Technology Program of Guangzhou (No. 202102021205).