Injung Kim

Han Gyel Sun, Hyunjae Ahn, HyunGyu Lee et al.

In this paper, we propose a new adapter network for adapting a pre-trained deep neural network to a target domain with minimal computation. The proposed model, unidirectional thin adapter (UDTA), helps the classifier adapt to new data by providing auxiliary features that complement the backbone network. UDTA takes outputs from multiple layers of the backbone as input features but does not transmit any feature to the backbone. As a result, UDTA can learn without computing the gradient of the backbone, which saves computation for training significantly. In addition, since UDTA learns the target task without modifying the backbone, a single backbone can adapt to multiple tasks by learning only UDTAs separately. In experiments on five fine-grained classification datasets consisting of a small number of samples, UDTA significantly reduced computation and training time required for backpropagation while showing comparable or even improved accuracy compared with conventional adapter models.

Jinyoung Choi, Youngchae Kwon, Injung Kim

Fashion style classification is a challenging task because of the large visual variation within the same style and the existence of visually similar styles. Styles are expressed not only by the global appearance, but also by the attributes of individual items and their combinations. In this study, we propose an item region-based fashion style classification network (IRSN) to effectively classify fashion styles by analyzing item-specific features and their combinations in addition to global features. IRSN extracts features of each item region using item region pooling (IRP), analyzes them separately, and combines them using gated feature fusion (GFF). In addition, we improve the feature extractor by applying a dual-backbone architecture that combines a domain-specific feature extractor and a general feature extractor pre-trained with a large-scale image-text dataset. In experiments, applying IRSN to six widely-used backbones, including EfficientNet, ConvNeXt, and Swin Transformer, improved style classification accuracy by an average of 6.9% and a maximum of 14.5% on the FashionStyle14 dataset and by an average of 7.6% and a maximum of 15.1% on the ShowniqV3 dataset. Visualization analysis also supports that the IRSN models are better than the baseline models at capturing differences between similar style classes.

Youngchae Kwon, Jinyoung Choi, Injung Kim

Fashion item detection is challenging due to the ambiguities introduced by the highly diverse appearances of fashion items and the similarities among item subcategories. To address this challenge, we propose a novel Holistic Detection Transformer (Holi-DETR) that detects fashion items in outfit images holistically, by leveraging contextual information. Fashion items often have meaningful relationships as they are combined to create specific styles. Unlike conventional detectors that detect each item independently, Holi-DETR detects multiple items while reducing ambiguities by leveraging three distinct types of contextual information: (1) the co-occurrence relationship between fashion items, (2) the relative position and size based on inter-item spatial arrangements, and (3) the spatial relationships between items and human body key-points. %Holi-DETR explicitly incorporates three types of contextual information: (1) the co-occurrence probability between fashion items, (2) the relative position and size based on inter-item spatial arrangements, and (3) the spatial relationships between items and human body key-points. To this end, we propose a novel architecture that integrates these three types of heterogeneous contextual information into the Detection Transformer (DETR) and its subsequent models. In experiments, the proposed methods improved the performance of the vanilla DETR and the more recently developed Co-DETR by 3.6 percent points (pp) and 1.1 pp, respectively, in terms of average precision (AP).

Dougho Park, Younghun Kim, Harim Kang et al.

Bolus segmentation is crucial for the automated detection of swallowing disorders in videofluoroscopic swallowing studies (VFSS). However, it is difficult for the model to accurately segment a bolus region in a VFSS image because VFSS images are translucent, have low contrast and unclear region boundaries, and lack color information. To overcome these challenges, we propose PECI-Net, a network architecture for VFSS image analysis that combines two novel techniques: the preprocessing ensemble network (PEN) and the cascaded inference network (CIN). PEN enhances the sharpness and contrast of the VFSS image by combining multiple preprocessing algorithms in a learnable way. CIN reduces ambiguity in bolus segmentation by using context from other regions through cascaded inference. Moreover, CIN prevents undesirable side effects from unreliably segmented regions by referring to the context in an asymmetric way. In experiments, PECI-Net exhibited higher performance than four recently developed baseline models, outperforming TernausNet, the best among the baseline models, by 4.54\% and the widely used UNet by 10.83\%. The results of the ablation studies confirm that CIN and PEN are effective in improving bolus segmentation performance.

Kihyun Na, Junseok Oh, Youngkwan Cho et al.

License plate recognition (LPR) is important for traffic law enforcement, crime investigation, and surveillance. However, license plate areas in dash cam images often suffer from low resolution, motion blur, and glare, which make accurate recognition challenging. Existing generative models that rely on pretrained priors cannot reliably restore such poor-quality images, frequently introducing severe artifacts and distortions. To address this issue, we propose a novel multi-frame license plate restoration and recognition framework, MF-LPR$^2$, which addresses ambiguities in poor-quality images by aligning and aggregating neighboring frames instead of relying on pretrained knowledge. To achieve accurate frame alignment, we employ a state-of-the-art optical flow estimator in conjunction with carefully designed algorithms that detect and correct erroneous optical flow estimations by leveraging the spatio-temporal consistency inherent in license plate image sequences. Our approach enhances both image quality and recognition accuracy while preserving the evidential content of the input images. In addition, we constructed a novel Realistic LPR (RLPR) dataset to evaluate MF-LPR$^2$. The RLPR dataset contains 200 pairs of low-quality license plate image sequences and high-quality pseudo ground-truth images, reflecting the complexities of real-world scenarios. In experiments, MF-LPR$^2$ outperformed eight recent restoration models in terms of PSNR, SSIM, and LPIPS by significant margins. In recognition, MF-LPR$^2$ achieved an accuracy of 86.44%, outperforming both the best single-frame LPR (14.04%) and the multi-frame LPR (82.55%) among the eleven baseline models. The results of ablation studies confirm that our filtering and refinement algorithms significantly contribute to these improvements.

Sungjae Kim, Kihyun Na, Jinyoung Choi et al.

Automatic Pitch Correction (APC) enhances vocal recordings by aligning pitch deviations with the intended musical notes. However, existing APC systems either rely on reference pitches, which limits their practical applicability, or employ simple pitch estimation algorithms that often fail to preserve expressiveness and naturalness. We propose BERT-APC, a novel reference-free APC framework that corrects pitch errors while maintaining the natural expressiveness of vocal performances. In BERT-APC, a novel stationary pitch predictor first estimates the perceived pitch of each note from the detuned singing voice. A context-aware note pitch predictor estimates the intended pitch sequence by leveraging a music language model repurposed to incorporate musical context. Finally, a note-level correction algorithm fixes pitch errors while preserving intentional pitch deviations for emotional expression. In addition, we introduce a learnable data augmentation strategy that improves the robustness of the music language model by simulating realistic detuning patterns. Compared to two recent singing voice transcription models, BERT-APC demonstrated superior performance in note pitch prediction, outperforming the second-best model, ROSVOT, by 10.49%p on highly detuned samples in terms of the raw pitch accuracy. In the MOS test, BERT-APC achieved the highest score of $4.32 \pm 0.15$, which is significantly higher than those of the widely-used commercial APC tools, AutoTune ($3.22 \pm 0.18$) and Melodyne ($3.08 \pm 0.18$), while maintaining a comparable ability to preserve expressive nuances. To the best of our knowledge, this is the first APC model that leverages a music language model to achieve reference-free pitch correction with symbolic musical context. The corrected audio samples of BERT-APC are available online.

Kihyun Na, Jinyoung Choi, Injung Kim

Virtual Try-On (VTON) is the task of synthesizing an image of a person wearing a target garment, conditioned on a person image and a garment image. While diffusion-based VTON models featuring a Dual UNet architecture demonstrate superior fidelity compared to single UNet models, they incur substantial computational and memory overhead due to their heavy structure. In this study, through visualization analysis and theoretical analysis, we derived three hypotheses regarding the learning of context features to condition the denoising process. Based on these hypotheses, we developed Re-CatVTON, an efficient single UNet model that achieves high performance. We further enhance the model by introducing a modified classifier-free guidance strategy tailored for VTON's spatial concatenation conditioning, and by directly injecting the ground-truth garment latent derived from the clean garment latent to prevent the accumulation of prediction error. The proposed Re-CatVTON significantly improves performance compared to its predecessor (CatVTON) and requires less computation and memory than the high-performance Dual UNet model, Leffa. Our results demonstrate improved FID, KID, and LPIPS scores, with only a marginal decrease in SSIM, establishing a new efficiency-performance trade-off for single UNet VTON models.

Gyuhwan Park, Kihyun Na, Injung Kim

The significance of license plate image restoration goes beyond the preprocessing stage of License Plate Recognition (LPR) systems, as it also serves various purposes, including increasing evidential value, enhancing the clarity of visual interface, and facilitating further utilization of license plate images. We propose a novel diffusion-based framework with character-level guidance, CharDiff, which effectively restores and recognizes severely degraded license plate images captured under realistic conditions. CharDiff leverages fine-grained character-level priors extracted through external segmentation and Optical Character Recognition (OCR) modules tailored for low-quality license plate images. For precise and focused guidance, CharDiff incorporates a novel Character-guided Attention through Region-wise Masking (CHARM) module, which ensures that each character's guidance is restricted to its own region, thereby avoiding interference with other regions. In experiments, CharDiff significantly outperformed the baseline restoration models in both restoration quality and recognition accuracy, achieving a 28% relative reduction in CER on the Roboflow-LP dataset, compared to the best-performing baseline model. These results indicate that the structured character-guided conditioning effectively enhances the robustness of diffusion-based license plate restoration and recognition in practical deployment scenarios.

Junseok Oh, Donghwee Yoon, Injung Kim

We propose a one-shot ultra-high-resolution generative adversarial network (OUR-GAN) framework that generates non-repetitive 16K (16, 384 x 8, 640) images from a single training image and is trainable on a single consumer GPU. OUR-GAN generates an initial image that is visually plausible and varied in shape at low resolution, and then gradually increases the resolution by adding detail through super-resolution. Since OUR-GAN learns from a real ultra-high-resolution (UHR) image, it can synthesize large shapes with fine details and long-range coherence, which is difficult to achieve with conventional generative models that rely on the patch distribution learned from relatively small images. OUR-GAN can synthesize high-quality 16K images with 12.5 GB of GPU memory and 4K images with only 4.29 GB as it synthesizes a UHR image part by part through seamless subregion-wise super-resolution. Additionally, OUR-GAN improves visual coherence while maintaining diversity by applying vertical positional convolution. In experiments on the ST4K and RAISE datasets, OUR-GAN exhibited improved fidelity, visual coherency, and diversity compared with the baseline one-shot synthesis models. To the best of our knowledge, OUR-GAN is the first one-shot image synthesizer that generates non-repetitive UHR images on a single consumer GPU. The synthesized image samples are presented at https://our-gan.github.io.

Minsu Kang, Sungjae Kim, Injung Kim

We propose a novel high-fidelity expressive speech synthesis model, UniTTS, that learns and controls overlapping style attributes avoiding interference. UniTTS represents multiple style attributes in a single unified embedding space by the residuals between the phoneme embeddings before and after applying the attributes. The proposed method is especially effective in controlling multiple attributes that are difficult to separate cleanly, such as speaker ID and emotion, because it minimizes redundancy when adding variance in speaker ID and emotion, and additionally, predicts duration, pitch, and energy based on the speaker ID and emotion. In experiments, the visualization results exhibit that the proposed methods learned multiple attributes harmoniously in a manner that can be easily separated again. As well, UniTTS synthesized high-fidelity speech signals controlling multiple style attributes. The synthesized speech samples are presented at https://anonymous-authors2022.github.io/paper_works/UniTTS/demos/.

Minsu Kang, Jihyun Lee, Simin Kim et al.

We propose an end-to-end speech synthesizer, Fast DCTTS, that synthesizes speech in real time on a single CPU thread. The proposed model is composed of a carefully-tuned lightweight network designed by applying multiple network reduction and fidelity improvement techniques. In addition, we propose a novel group highway activation that can compromise between computational efficiency and the regularization effect of the gating mechanism. As well, we introduce a new metric called Elastic mel-cepstral distortion (EMCD) to measure the fidelity of the output mel-spectrogram. In experiments, we analyze the effect of the acceleration techniques on speed and speech quality. Compared with the baseline model, the proposed model exhibits improved MOS from 2.62 to 2.74 with only 1.76% computation and 2.75% parameters. The speed on a single CPU thread was improved by 7.45 times, which is fast enough to produce mel-spectrogram in real time without GPU.

Jinhyeok Yang, Junmo Lee, Youngik Kim et al.

We present a novel high-fidelity real-time neural vocoder called VocGAN. A recently developed GAN-based vocoder, MelGAN, produces speech waveforms in real-time. However, it often produces a waveform that is insufficient in quality or inconsistent with acoustic characteristics of the input mel spectrogram. VocGAN is nearly as fast as MelGAN, but it significantly improves the quality and consistency of the output waveform. VocGAN applies a multi-scale waveform generator and a hierarchically-nested discriminator to learn multiple levels of acoustic properties in a balanced way. It also applies the joint conditional and unconditional objective, which has shown successful results in high-resolution image synthesis. In experiments, VocGAN synthesizes speech waveforms 416.7x faster on a GTX 1080Ti GPU and 3.24x faster on a CPU than real-time. Compared with MelGAN, it also exhibits significantly improved quality in multiple evaluation metrics including mean opinion score (MOS) with minimal additional overhead. Additionally, compared with Parallel WaveGAN, another recently developed high-fidelity vocoder, VocGAN is 6.98x faster on a CPU and exhibits higher MOS.

Inyoung Paik, Taeyeong Kwak, Injung Kim

In capsule networks, the routing algorithm connects capsules in consecutive layers, enabling the upper-level capsules to learn higher-level concepts by combining the concepts of the lower-level capsules. Capsule networks are known to have a few advantages over conventional neural networks, including robustness to 3D viewpoint changes and generalization capability. However, some studies have reported negative experimental results. Nevertheless, the reason for this phenomenon has not been analyzed yet. We empirically analyzed the effect of five different routing algorithms. The experimental results show that the routing algorithms do not behave as expected and often produce results that are worse than simple baseline algorithms that assign the connection strengths uniformly or randomly. We also show that, in most cases, the routing algorithms do not change the classification result but polarize the link strengths, and the polarization can be extreme when they continue to repeat without stopping. In order to realize the true potential of the capsule network, it is essential to develop an improved routing algorithm.

Inyoung Paik, Sangjun Oh, Tae-Yeong Kwak et al.

To address the issue of catastrophic forgetting in neural networks, we propose a novel, simple, and effective solution called neuron-level plasticity control (NPC). While learning a new task, the proposed method preserves the knowledge for the previous tasks by controlling the plasticity of the network at the neuron level. NPC estimates the importance value of each neuron and consolidates important \textit{neurons} by applying lower learning rates, rather than restricting individual connection weights to stay close to certain values. The experimental results on the incremental MNIST (iMNIST) and incremental CIFAR100 (iCIFAR100) datasets show that neuron-level consolidation is substantially more effective compared to the connection-level consolidation approaches.