Neural Network-Based Atlas Enhancement in MPEG Immersive Video

Abstract

Recently, the demand for immersive videos has surged with the expansion of virtual reality, augmented reality, and metaverse technologies. As an international standard, moving picture experts group (MPEG) has developed MPEG immersive video (MIV) to efficiently transmit large-volume immersive videos. The MIV encoder generates atlas videos to convert extensive multi-view videos into low-bitrate formats. When these atlas videos are compressed using conventional video codecs, compression artifacts often appear in the reconstructed atlas videos. To address this issue, this study proposes a feature-extraction-based convolutional neural network (FECNN) to reduce the compression artifacts during MIV atlas video transmission. The proposed FECNN uses quantization parameter (QP) maps and depth information as inputs and consists of shallow feature extraction (SFE) blocks and deep feature extraction (DFE) blocks to utilize layered feature characteristics. Compared to the existing MIV, the proposed method improves the Bjontegaard delta bit-rate (BDBR) by −4.12% and −6.96% in the basic and additional views, respectively.

1. Introduction

Recently, the expansion of virtual reality, augmented reality, and the metaverse with multimodal interactions has highlighted the need for efficient immersive video streaming services. As a kind of immersive video, multi-view texture and depth videos are obtained from various viewpoints using an omnidirectional light-field camera comprising an array of micro lenses. These videos require massive amounts of raw data, including multiple texture videos with diverse perspectives and depth videos displaying distances between objects and the cameras. Consequently, powerful video codecs for immersive videos are required to transmit them within the constrained network bandwidth. As shown in Figure 1, the number of two-dimensional (2D) video codecs for immersive videos is determined by the number of input videos with multi-view texture and depth videos, resulting in high complexity. Moving picture experts group (MPEG) has attempted to resolve this complexity issue through MPEG immersive video (MIV) standardization [1]. MPEG released the test model for MIV (TMIV) in 2020 [2], serving as a reference software for compressing the atlas videos derived from multi-view texture and depth videos. Currently, TMIV uses either high efficiency video coding (HEVC) [3] or versatile video coding (VVC) [4].

Figure 2 shows the overall architecture of TMIV, in which several multi-view texture and depth videos constituting the input videos are injected into TMIV encoder. For each texture and depth video, the TMIV encoder generates texture and depth atlas videos for basic and additional views. As atlas videos, each texture and depth basic view is the representative video generated by the multiple texture and depth videos. In contrast, the additional view is composed of rectangular patches obtained from the residual after removing the overlapped regions of the basic view. These basic and additional views exhibit distinct characteristics compared to natural videos, as illustrated in Figure 3. Figure 3a shows a texture atlas used as the basic view where large coherent regions are packed to minimize patch boundaries. Figure 3b shows the texture additional view that gathers residual patches not assigned to the basic view. Figure 3c shows a depth atlas used as the basic view which exhibits smooth layers with strong gradients at object boundaries. Figure 3d shows the depth additional view that contains sparse patches for the remaining depth regions.

Therefore, this study proposes a neural network-based post-processing filter to significantly reduce the compression artifacts of atlas videos. We present a feature extraction-based convolutional neural network (FECNN) that consists of shallow feature extraction (SFE) and deep feature extraction (DFE) blocks to improve the quality of texture atlas videos. The proposed method takes advantage of the FECNN’s performance by utilizing a depth atlas and quantization parameter (QP) information as input data. Additionally, various ablation studies were conducted to evaluate the efficiency of the proposed design and its effect on the performance of neural networks. The performance of the proposed method was evaluated under the test sequences provided by MIV and measured using Bjontegaard delta bit-rate (BDBR) as performance metrics. The primary contributions of this study are summarized as follows:

• The proposed FECNN newly deployed depth atlas and QP information were used as inputs to enhance the coding performance as well as visual quality. To the best of our knowledge, this is the first study that uses such neural networks to enhance the atlas quality of MIV in the literature.
• To develop FECNN, we designed SFE and DFE blocks to improve the visual quality of the texture atlas compared with that of existing TMIV. Specifically, it can provide noticeable visual improvements between objects and patch boundary lines in the atlas videos.
• The proposed FECNN improved BDBR by −4.12% and −6.96% on average in the basic and additional views, respectively.

The remainder of this paper is organized as follows: Section 2 reviews previous coding methods for immersive video and various post-processing studies for artifact reduction. Section 3 describes the proposed FECNN in terms of network architecture and training. Finally, experimental results and conclusions are given in Section 4 and Section 5, respectively.

2. Related Works

2.1. Previous Coding Methods for Immersive Video

To improve the coding performance of immersive video, several studies investigated new views and enhanced the accuracy of input depth videos. For instance, Jeong et al. proposed a method that assigned different QP values due to the varying characteristics between the basic and additional views of atlas videos [5]. Milovanovic et al. reduced the amount of transmitted depth data by leveraging the characteristics of the depth information included in the texture view [6]. Lee et al. designed a multi-view streaming system that employed group-based immersive video by deriving weights from these groups and transmitting bitstreams based on the given viewports [7]. Garus et al. proposed the depth recovery concept and implemented a decoder-side depth recovery method that exploited motion information in a texture bitstream [8]. Jeong et al. combined texture and depth information to generate a single frame, which was then compressed into subpicture bitstreams to form a unified bitstream [9]. Mieloch et al. proposed the following two techniques to improve the depth map accuracy: (i) input depth map assistance (IDMA) technique to modify the depth map using the global multi-view optimization method and (ii) the extended IDMA technique to enhance the quality of the estimated depth map by reprojecting the input depth map onto different views [10]. Lim et al. extended the dynamic depth range using the patchwise min-max linear scaling method to accurately transmit depth information [11]. Dziembowski et al. applied two methods to improve the coding performance of immersive video. The first method removed the constant components of all luma and chroma constituents of each patch in the atlas, and the second method adjusted the dynamic range of the depth atlas to match the quality of the input depth map [12]. Lee et al. proposed a high bit depth geometry representation for MIV that preserves sub-bit details by assigning the post-quantization residual bits to the chroma channels and adds tailored preprocessing for YUV 420 format [13]. Oh et al. proposed an efficient atlas coding strategy for object based MIV that detects unused areas in the atlas and reduces the atlas resolution during encoding to cut computation and improve coding efficiency [14]. In contrast to the aforementioned studies, the proposed method deploys a neural network to enhance the visual quality of atlas video as well as the coding performance of immersive video.

2.2. Artifact Reduction of Video Coding

Several studies on visual quality enhancement methods have utilized neural networks to remove compression artifacts caused by lossy coding schemes. To improve the quality of HEVC I-frames, Dai et al. developed a deep neural network-based in-loop filter, known as the variable-filter-size residue-learning convolutional neural network (VRCNN) [15]. Yu et al. proposed a high-frequency guided CNN to enhance the quality of the chroma components using high-frequency information from the input luma component [16]. Hoang et al. investigated a deep recursive residual network with block information (B-DRRN) to provide better quality for compressed frames using coding block information [17]. B-DRRN includes additional network branches to utilize block information and reduce network complexity by applying recursive residual structures and weight-sharing techniques. Wang et al. designed a unified single-model solution that is a CNN-based in-loop filter applicable to video coding [18]. They also proposed an attention based dual-scale CNN to remove the artifacts from reconstructed videos using QP and coding unit (CU) partitioning information. Zhao et al. proposed variable-filter-size residue-learning CNN with batch normalization (VRCNN-BN) [19]. VRCNN-BN was designed as an end-to-end model and applied separately to the luma and chroma components of the video. Das et al. developed a CNN-based post-processing framework that utilized QP information to enhance the visual quality of reconstructed videos [20]. Zhang et al. proposed a weakly connected dense attention neural network (WCDANN) to remove visual artifacts based on residual learning scheme [21]. Santamaria et al. proposed a content-adaptive CNN post-filter with per-picture activation signaled via Supplemental Enhancement Information (SEI), built upon the Neural Network-based Post Filter (NNPF) [22] architecture. The filter minimizing Sum of Squared Differences (SSD) is selected among four candidates, enabling decoder-side operation without bitstream changes [23]. Although these methods have demonstrated high performance in reducing compression artifacts from natural videos, there are limitations in applying them to texture and depth atlas videos due to distinct characteristics compared with natural videos. Thus, research on neural network-based algorithms suitable for atlas video is required.

3. Proposed Method

As MIV specified a common test condition (CTC), four different QP values were used in each test sequence under CTC [24].

Table 1. QP values for atlas encoding in MIV CTC.

Class	Sequences	Texture QP				Depth QP
		RP1	RP2	RP3	RP4	RP1	RP2	RP3	RP4
A01	Classroom Video	26	30	38	51	7	10	16	27
B01	Museum	29	40	47	51	9	18	23	27
B02	Chess	18	27	35	45	1	7	14	22
B03	Guitarist	22	24	29	39	3	5	9	17
C01	Hijack	19	24	34	49	1	5	13	25
C02	Cyberpunk	21	24	29	39	3	5	9	17
J01	Kitchen	18	26	33	41	1	7	12	19
J02	Cadillac	22	31	41	51	3	11	19	27
J03	Mirror	26	33	42	51	7	12	19	27
J04	Fan	32	38	45	51	11	16	22	27
W01	Group	26	31	37	46	7	11	15	23
W02	Dancing	20	24	28	40	2	5	8	18
D01	Painter	24	32	43	51	5	11	20	27
D02	Breakfast	25	30	35	43	6	10	14	20
D03	Barn	25	30	35	42	6	10	14	19
E01	Frog	29	34	40	46	9	13	18	23
E02	Carpark	23	28	37	47	4	8	15	23
E03	Street	21	25	32	41	3	6	11	19
L01	Fencing	23	28	39	51	4	8	17	27
L02	CBABaskeball	24	27	31	43	5	7	11	20
L03	MartialArts	24	27	31	43	5	7	11	20

shows that a test sequence has four rate points (RP) corresponding to the texture and depth atlas videos. Because depth atlas videos are compressed with lower QP values, reconstructed depth atlas videos have minor compression artifacts. Conversely, texture atlas videos employing a maximum QP of 51 generate significant compression artifacts and lead to the degradation of final views.

Figure 4 shows the integrated MIV framework with the proposed FECNN. The roles of the blocks in Figure 4 are as follows in a compact form: Multi view videos provide the texture and depth video inputs, and the TMIV encoder converts the multi view inputs into atlases and metadata while removing redundancy, with the texture atlas carrying patches for the basic and additional views. The VVC encoder and decoder compress and reconstruct the atlases with QP maps, and FECNN enhances the decoded texture atlases using depth and QP to reduce artifacts without blurring edges. From a modeling standpoint, decoded atlases can be regarded as clean signals perturbed by quantization noise whose magnitude depends on the QP and by boundary-localized artifacts that arise near patch or object boundaries. To address these distortions, we guide the restoration using depth, which provides a geometry-aware cue that preserves object boundaries, and we use the QP maps as a spatial confidence signal that allocates stronger correction to neighborhoods encoded with high QP while keeping edge structures intact. In this study we deploy VVC as the video codec, and we apply FECNN as a decoder-side post-processing filter to the reconstructed texture atlases for both basic and additional views. FECNN enhances the texture atlas using SFE and DFE blocks; SFE extracts boundary and patch-aware low-level features, and DFE aggregates broader spatial context to suppress blocking and ringing without blurring edges.

3.1. Network Architecture

Figure 5 shows the architecture of the proposed neural network, comprising the input layer, SFE, DFE, and output layer. The input layer feeds the following four inputs: the texture atlas, depth atlas, texture QP map, and depth QP map. Texture and depth atlas videos are reconstructed video sequences from the video codecs of Figure 4, with the latter being half the size of the former in terms of width and height. To generate QP maps of texture and depth, the texture and depth QP values (${QP}_{texture}$, ${QP}_{depth}$) used in the atlas videos encoding process have normalized values that range from zero to one by dividing the maximum QP value (${QP}_{max}$). Then, QP maps of texture and depth are subsequently filled with the normalized values with the same size of input texture and depth atlas videos.

According to the VVC specification, ${QP}_{max}$ value was set to 63. Input layer, output feature maps were extracted using a 3 × 3 filter to account for the spatial correlations of the texture and depth atlas videos. Because the input QP maps were insufficient to represent the spatial information, the feature maps were extracted using a 1 × 1 filter. The convolutional operation of FECNN between output feature maps ${(F}_i)$ and previous feature maps ${(F}_{i-1})$ is expressed in Equation (1):

$$F_i={\delta }_i(W_i⨂F_{i-1}+B_i)$$

(1)

where δ(•), $W_i$, $B_i$, and ⨂ represent the parametric rectified linear unit (PReLU) as one of the activation functions, filter weights, biases, and convolutional operations, respectively. In this study, the notation of a convolutional operation is denoted as Conv (filter size, channel depth of generated feature maps).

Atlases are piecewise smooth with discontinuities at patch and object boundaries, and these regions are amplified at high QP. The SFE uses two branches to learn complementary texture-driven and depth-driven features so that boundary structures are preserved. The DFE aggregates global context to mitigate blocking and ringing without blurring edges. We use separable 1 × 3 and 3 × 1 filters to approximate 3 × 3 responses with fewer parameters, and we use PixelShuffle to bring half-resolution depth features to the texture resolution by channel-to-space rearrangement, which avoids interpolation blur near boundaries.

The SFE layer was designed to extract each feature from the concatenated texture, depth, and QP feature map. In SFE layers, the output feature maps derived from depth information were upsampled by a factor of two using the pixel shuffle method [25] and concatenated with those derived from texture information. Figure 6a shows a SFE block structure comprising four convolution layers and one skip connection [26]. After the intermedia features were extracted using 1 × 1 and 3 × 3 filters, combinations of 1 × 3 and 3 × 1 filters were used to achieve a performance comparable to that of the 3 × 3 filter with lower complexity. In addition, the feature maps extracted using the 1 × 1 filter were used as the skip connection to transmit initial input information.

Figure 6b shows the DFE block structure, comprising six convolution layers and one skip connection. The input feature maps of DFE were passed through a 3 × 3 convolution layer, which was split into two branches for deep feature extraction. Here, deep feature refers to features that closely resemble texture information. Subsequently, the output feature maps from each branch were concatenated, and the channel depth was reduced to alleviate the memory burden of the DFE. Then, the feature maps with the reduced channel depth were calculated by two convolution layers with 1 × 3 and 3 × 1 filters. Similarly to SFE, the feature maps extracted from the initial 3 × 3 filter were used as the skip connection.

3.2. Network Training

The proposed FECNN was trained on an immersive video dataset provided by MIV. As shown in

Table 2. MIV test sequences on CTC.

Content Categories	Class	Sequences	Resolution	Number of Source Views
Computer generated	A	Classroom Video	4096 × 2048	15
	B	Museum	2048 × 2048	24
		Chess	2048 × 2048	10
		Guitarist	2048 × 2048	46
	C	Hijack	4096 × 2048	10
		Cyberpunk	2048 × 2048	10
	J	Kitchen	1920 × 1080	25
		Cadillac	1920 × 1080	15
		Mirror	1920 × 1080	15
		Fan	1920 × 1080	15
	W	Group	1920 × 1080	21
		Dancing	1920 × 1080	24
Natural	D	Painter	2048 × 2048	16
		Breakfast	1920 × 1080	15
		Barn	1920 × 1080	15
	E	Frog	1920 × 1080	13
		Carpark	1920 × 1088	9
		Street	1920 × 1088	9
	L	Fencing	1920 × 1088	10
		CBABasketball	2048 × 1088	34
		Martial Arts	1920 × 1080	15

, the dataset consists of 21 video sequences [27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44] and eight sequences labeled “optionally” were used as the training dataset. The video sequences in the training dataset were compressed under CTC configurations and subsequently converted to the NPY data type for network training. The converted training data were extracted without overlapping regions from the texture and depth atlas videos with a size of 256 × 256 and 128 × 128, respectively, and collected a total of 451,904 patches. For additional views, we masked out gray background regions and sampled patches only from valid atlas pixels.

The proposed network was trained to minimize the mean square error between the original and reconstructed videos with the loss function as in Equation (2):

$$L\left(\theta \right)={\displaystyle \frac{1}{N}}\sum _{i=0}^{N-1}{‖O_i-Y_i‖}_2^2$$

(2)

where θ, N, O, and Y represent the network parameters such as filter weights, batch size, original, and reconstructed videos, respectively.

Table 3. Hyper-parameters of the proposed method.

Hyper-Parameters	Options
Loss function	Mean Squared Error (MSE)
Optimizer	Adam
Number of epochs	50
Batch size	32
Learning rate	${10}^{-4}$ $\mathrm{to}{10}^{-6}$
Activation function	PReLU

shows the hyper-parameters for the training of FECNN. The proposed network measured L2 loss with a batch size of 32 and used Adam optimizer [45] with a learning rate of ${10}^{-4}$ over 50 epochs, where the learning rate was reduced by a factor of 10 every 20 epochs. To calculate the mean gradients during the training, the hyper-parameters β1 and β2 of Adam optimizer were set to 0.9 and 0.999, respectively.

4. Experimental Results

4.1. Experimental Environments

All training and test procedures were performed using an Intel Xeon Gold 5218 CPU (16 cores @ 2.30 GHz) with 258 GB of RAM and three NVIDIA Tesla V100 GPUs. To evaluate the performance of the proposed method, we used the “mandatory” video sequences among the MIV CTC dataset. These sequences represent essential videos that must be used to evaluate the coding performance of MIV. They comprised eight sequences containing the classes B, D, E, J, L, and W. The video encoding and decoding processes were performed using the software listed in

Table 4. Software and library versions used in our experiments.

Software	Version
TMIV	17.0
VVenC	1.7.0
VVdeC	1.6.0
PyTorch	1.14.0
CUDA	11.6

.

4.2. Performance Measurements

The Bjontegaard delta bit-rate (BDBR) is used to evaluate coding performance [46]. In general, a BDBR increase of 1% corresponds to a decrease of 0.05 dB in the BD-peak signal-to-noise ratio (BD-PSNR). Thus, a higher BDBR indicates coding loss. This metric is valuable for comparing the coding performance of different video codecs as it considers both bit-rate and video quality. By quantifying the difference in bit-rate needed to achieve the same video quality, BDBR objectively evaluates coding performance. Equation (3) calculates the weighted average BDBR of the Y, U, and V color components in the test sequences recommended by MIV CTC.

$${BDBR}_{YUV}={\displaystyle \frac{6\times {BDBR}_Y+{BDBR}_U+{BDBR}_V}{8}}$$

(3)

Here, ${BDBR}_Y$, ${BDBR}_U$, and ${BDBR}_V$ denote the BDBR of the Y, U, and V color components of the texture atlas, respectively.

Table 5. Comparison of coding performance between the proposed method and CNN-based post-filters.

Class	Sequences	Proposed Method		Ref. [20]		Ref. [22]
		Basic	Additional	Basic	Additional	Basic	Additional
		BDBR-YUV	BDBR-YUV	BDBR-YUV	BDBR-YUV	BDBR-YUV	BDBR-YUV
B02	Chess	−2.51%	−4.85%	−1.56%	−3.96%	1.74%	0.89%
B03	Guitarist	−1.98%	−3.37%	−0.25%	−2.54%	0.68%	0.20%
J02	Cadillac	−3.13%	−7.25%	−1.32%	−4.29%	0.21%	0.11%
J04	Fan	−8.88%	−13.19%	−1.25%	−5.40%	−0.83%	−1.02%
W01	Group	−2.15%	−6.74%	−0.12%	−4.31%	−0.55%	−0.46%
D01	Painter	−4.96%	−9.37%	−1.33%	−6.59%	−0.40%	−0.18%
E01	Frog	−4.10%	−4.88%	−2.77%	−4.06%	−0.82%	−0.81%
L02	CBABasketball	−5.24%	−6.01%	−4.36%	−5.20%	0.07%	−0.19%
Average		−4.12%	−6.96%	−1.62%	−4.54%	0.01%	−0.18%

lists the BDBR performances of the basic and additional views of the texture atlas. Since FECNN operates as a decoder-side post-filter, the bitstream and bitrate remain unchanged. Therefore, the reported BD-rate reductions stem from decreased distortion at identical rate points. Compared with TMIV as an anchor, the proposed method achieved average BD-rate reductions in basic and additional views to 4.12% and 6.96%, respectively. Particularly, Fan sequence exhibited a significant BDBR reduction in both basic and additional views. As the additional views composed the combinations of patches, resulting in numerous boundaries between patches, compression artifacts primarily occur at the boundaries of the additional view. Consistent with our modeling, additional views exhibit larger gains than basic views because they typically contain a higher density of patch/object boundaries where boundary-localized distortions accumulate. Sequences with stronger boundary activity or high-frequency textures show the most pronounced improvements. Thus, the proposed method achieved good performance in artifact removal at these boundaries as well as BDBR improvement. Existing CNN-based post-filters [20], ref. [22] apply single-stream sequential architectures from conventional natural video coding to reconstructed frames. These methods do not sufficiently consider the many patch and object boundaries present in atlases. Consistent with this difference, ref. [20] attains modest, content-dependent gains by using QP information, and these gains tend to be larger on additional views than on basic views, yet they remain smaller than those achieved by the proposed method. In contrast, ref. [22] uses only the texture input and shows limited or inconsistent improvements, with little or no gain in some cases. Collectively, these outcomes indicate that designs that explicitly reflect atlas characteristics are advantageous.

In addition to BD-rate, we report Structure Similarity Index Measure (SSIM) as a perceptual indicator on the reconstructed texture luma, averaged over frames for each sequence and reported separately for basic and additional views. As shown in

Table 6. Comparison of SSIM between the proposed method and CNN-based post-filters.

Class	Sequences	Proposed Method		Ref. [20]
		Basic	Additional	Basic	Additional
B02	Chess	0.9744	0.9636	0.9741	0.9628
B03	Guitarist	0.9704	0.9689	0.9699	0.9685
J02	Cadillac	0.9540	0.9421	0.9494	0.9365
J04	Fan	0.8796	0.9226	0.8683	0.9149
W01	Group	0.8785	0.8829	0.8763	0.8804
D01	Painter	0.9136	0.9204	0.9084	0.9176
E01	Frog	0.8674	0.8434	0.8652	0.8415
L02	CBABasketball	0.9641	0.9587	0.9636	0.9580
Average		0.9252	0.9253	0.9219	0.9225

, the proposed method attains higher SSIM than ref. [20] on all sequences and on both view types. The average gains are +0.0033 on basic views and +0.0028 on additional views, with the largest improvements on Fan and Cadillac additional views. Higher SSIM indicates better preservation of structural content near patch and object boundaries.

Table 7. Comparison of computation complexity the network.

Methods	Parameters	FLOPs	Inference Time
Proposed Method	5.01 MB	31.62 T	22,861.44 ms
[20]	8.45 MB	39.19 T	14,364.18 ms
[22]	0.34 MB	1.54 T	422.27 ms

summarizes model size, FLOPs, and inference time measured on an RTX 4070 Ti at the native atlas resolution. The proposed model processes four inputs per frame, namely two maps at 1920 × 4608 for the texture atlas and its QP map and two maps at 960 × 2304 for the depth atlas and its QP map. The proposed model uses 5.01 MB of parameters and 31.62 T FLOPs, which are both smaller than ref. [20] with 8.45 MB and 39.19 T. Its inference time is 22.86 s, while ref. [20] records 14.36 s. This latency difference arises because FLOPs count arithmetic operations only, whereas wall-clock time is also affected by memory traffic from multiple inputs and the movement of intermediate feature maps. Method ref. [22] has the smallest complexity and the fastest runtime, and its coding gains are correspondingly limited. Taken together, these results indicate that the proposed method achieves substantially better performance at a suitable level of complexity compared with generic natural content post-filters.

Both BD-rates are computed on the rendered views: Y-PSNR measures pixel-by-pixel distortion on the luma channel, whereas Immersive Video (IV)-PSNR [47] uses a local search window with a global color-difference compensation and is less sensitive to small spatial shifts.

Table 8. Comparison of BD-rate on rendered views using Y-PSNR and IV-PSNR for MIV-CTC. Negative values indicate bitrate reduction at equal quality.

Class	Sequences	BD-Rate Y-PSNR	BD-Rate IV-PSNR
B02	Chess	−0.7%	2.7%
B03	Guitarist	−2.9%	5.1%
J02	Cadillac	−0.9%	3.5%
J04	Fan	3.3%	2.7%
W01	Group	0.3%	1.3%
D01	Painter	−3.2%	4.4%
E01	Frog	−1.9%	5.1%
L02	CBABasketball	−1.4%	2.3%
Average		−0.9%	3.4%

shows that our post-filter attains an average BD-rate of −0.9% on Y-PSNR but +3.4% on IV-PSNR. This gap mainly occurs because the network sharpens texture edges while the depth atlas, which drives view warping, is unchanged, so small mis-registrations around depth discontinuities persist after rendering and the local matching in IV-PSNR cannot fully compensate, which can lower the score even when pixel-wise distortion decreases. The impact is further diluted because most synthesized pixels come from interior regions, so boundary-focused gains contribute less to the rendered average. To better align atlas enhancement with downstream quality, a renderer-aware integrated network that jointly processes texture and depth atlases and enforces cross-modal consistency at patch and object boundaries will be explored as future work.

4.3. Visual Comparisons

Figure 7 and Figure 8 show the visual quality comparisons between VVC compressed content and the proposed method at the maximum QP value. For a sophisticated comparison of test datasets, the figures show the zoom-in for the areas in the red boxes. Each row presents the ground truth, the reconstructed result of the TMIV, and that of the proposed method. As a result, the proposed network effectively reduced artifacts occurring between objects and patch boundaries of the atlas videos. In Figure 7a the boxed region around the sphere shoulder and the distant chess pieces exhibits reduced ringing and smoother curvature and the silhouettes remain sharp. As show in Figure 7b the inset over the fan grille shows thin curved lines that stay continuous and straight with bending and staircase artifacts suppressed. As depicted in Figure 8a the boxed patch junction indicates that the visible seam and brightness jump disappear and texture continuity is restored across the packing cut. As shown in Figure 8b under a high QP setting, the boxed region near a patch boundary shows reduced blocking and ringing while the diagonal object edge remains crisp. Overall, the proposed method proves effective for atlas videos by directly targeting patch seams and object boundaries with depth guidance and QP adaptation.

4.4. Ablation Studies

To investigate the optimal network architecture of FECNN, a variety of ablation studies were conducted as follows:

• Performance analysis according to the use of chroma components as an input
• Determination of a suitable up-sampling method
• Optimal number of SFE and DFE blocks

As shown in

Table 9. Coding performance between w/chroma and w/o chroma.

Method		w/Chroma		w/o Chroma
Class	Sequences	Basic	Additional	Basic	Additional
		BDBR-YUV	BDBR-YUV	BDBR-YUV	BDBR-YUV
B02	Chess	−1.64%	−3.41%	−1.46%	−3.90%
B03	Guitarist	−0.98%	−2.63%	−1.71%	−3.14%
J02	Cadillac	−1.28%	−5.27%	−3.36%	−6.63%
J04	Fan	−6.91%	−12.95%	−8.18%	−12.36%
W01	Group	−1.85%	−5.93%	−1.96%	−6.45%
D01	Painter	−3.19%	−7.82%	−4.71%	−8.47%
E01	Frog	−3.14%	−4.49%	−3.81%	−4.73%
L02	CBABasketball	−4.13%	−4.87%	−4.64%	−5.44%
Average		−2.89%	−5.92%	−3.73%	−6.39%

, excluding chroma yields better averages. Without chroma, the averages are −3.73% on basic views and −6.39% on additional views. With chroma, the averages are −2.89% and −5.92%. Since adding chroma increases memory use and inference time due to extra inputs, the final model does not include chroma.

Table 10. Coding performance between de-convolution and pixel shuffle.

Method		De-Convolution		Pixel Shuffle
Class	Sequences	Basic	Additional	Basic	Additional
		BDBR-YUV	BDBR-YUV	BDBR-YUV	BDBR-YUV
B02	Chess	−1.46%	−3.90%	−2.01%	−4.19%
B03	Guitarist	−1.71%	−3.14%	−1.64%	−3.00%
J02	Cadillac	−3.36%	−6.63%	−3.05%	−6.56%
J04	Fan	−8.18%	−12.36%	−8.55%	−12.58%
W01	Group	−1.96%	−6.45%	−1.80%	−6.38%
D01	Painter	−4.71%	−8.47%	−4.92%	−8.57%
E01	Frog	−3.81%	−4.73%	−3.81%	−4.63%
L02	CBABasketball	−4.64%	−5.44%	−4.62%	−5.43%
Average		−3.73%	−6.39%	−3.80%	−6.42%

compares transposed convolution and PixelShuffle. PixelShuffle achieves slightly better averages, −3.80% on basic views and −6.42% on additional views, while transposed convolution achieves −3.73% and −6.39%. PixelShuffle also performs channel-to-space rearrangement that avoids interpolation blur and checkerboard artifacts near patch or object boundaries, so we adopt PixelShuffle.

We swept the number of blocks while keeping other settings fixed. As reported in

Table 11. Comparisons of the number of SFE and DEF blocks. Bold indicates the best performance.

Test		Basic	Additional
SFE Block	DFE Block	BDBR-YUV	BDBR-YUV
3	3	−3.80%	−6.42%
4	3	−3.79%	−6.51%
5	3	−3.93%	−6.67%
3	4	−3.93%	−6.71%
3	5	−3.49%	−6.42%
4	4	−4.05%	−6.78%
4	5	−4.12%	−6.96%
5	4	−3.82%	−6.64%
5	5	−3.75%	−6.56%

, performance improves up to 4 SFE and 5 DFE, reaching −4.12% on basic views and −6.96% on additional views, and then plateaus or slightly regresses. We therefore adopt 4 SFE/5 DFE as the final configuration and use this setting in all main experiments.

5. Conclusions

This study proposes an FECNN to enhance both visual quality and coding performance compared to the conventional TMIV as an anchor. The proposed network employs newly designed SFE blocks that extract boundary and patch aware features from the texture atlas, depth atlas, and QP maps, followed by DFE blocks that aggregate wide spatial context conditioned on depth and QP to restore texture atlas details while suppressing blocking and ringing near patch and object boundaries. A separable convolution method is deployed to reduce the complexity of the FECNN. The performance of FECNN was objectively evaluated under MIV CTC environments and subjectively identified through visual quality comparisons. A variety of ablation studies were conducted to determine the optimal FECNN structure. Experimental results show that the proposed method can improve the BDBR by −4.12% and −6.96% on average in the basic and additional views of the texture atlas, respectively. Compared with conventional TMIV decoding and a generic CNN post filter, the proposed method reduces blocking, ringing, and patch seams near object boundaries through atlas aware depth guidance and QP adaptive restoration.

For future work, we will extend the decoder side restoration to a joint model that enhances both the texture atlas and the depth atlas. Our observation is that improving only the texture atlas can leave inconsistencies with the depth atlas, which limits rendering quality near patch and object boundaries. The planned model will share features across modalities and will use geometry aware constraints to enforce consistency between texture edges and depth discontinuities while remaining QP adaptive. We will also study temporal consistency and lightweight modules so that the method remains real time and preserves full compatibility with TMIV and the VVC bitstream.