• Journal of Information Processing Systems
• /
• 제18권6호
• /
• pp.794-802
• /
• 2022

The recovery of reasonable depth information from different scenes is a popular topic in the field of computer vision. For generating depth maps with better details, we present an efficacious monocular depth prediction framework with coordinate attention and feature fusion. Specifically, the proposed framework contains attention, multi-scale and feature fusion modules. The attention module improves features based on coordinate attention to enhance the predicted effect, whereas the multi-scale module integrates useful low- and high-level contextual features with higher resolution. Moreover, we developed a feature fusion module to combine the heterogeneous features to generate high-quality depth outputs. We also designed a hybrid loss function that measures prediction errors from the perspective of depth and scale-invariant gradients, which contribute to preserving rich details. We conducted the experiments on public RGBD datasets, and the evaluation results show that the proposed scheme can considerably enhance the accuracy of depth prediction, achieving 0.051 for log10 and 0.992 for δ<1.25³ on the NYUv2 dataset.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제15권10호
• /
• pp.3729-3749
• /
• 2021

At present, deep convolution network-based salient object detection (SOD) has achieved impressive performance. However, it is still a challenging problem to make full use of the multi-scale information of the extracted features and which appropriate feature fusion method is adopted to process feature mapping. In this paper, we propose a new adjacency auxiliary network (AANet) based on multi-scale feature fusion for SOD. Firstly, we design the parallel connection feature enhancement module (PFEM) for each layer of feature extraction, which improves the feature density by connecting different dilated convolution branches in parallel, and add channel attention flow to fully extract the context information of features. Then the adjacent layer features with close degree of abstraction but different characteristic properties are fused through the adjacent auxiliary module (AAM) to eliminate the ambiguity and noise of the features. Besides, in order to refine the features effectively to get more accurate object boundaries, we design adjacency decoder (AAM_D) based on adjacency auxiliary module (AAM), which concatenates the features of adjacent layers, extracts their spatial attention, and then combines them with the output of AAM. The outputs of AAM_D features with semantic information and spatial detail obtained from each feature are used as salient prediction maps for multi-level feature joint supervising. Experiment results on six benchmark SOD datasets demonstrate that the proposed method outperforms similar previous methods.

• 한국인공지능학회지
• /
• 제11권3호
• /
• pp.17-22
• /
• 2023

In this paper, we propose a black ice detection system using Convolutional Neural Networks (CNNs). Black ice poses a serious threat to road safety, particularly during winter conditions. To overcome this problem, we introduce a CNN-based architecture for real-time black ice detection with an encoder-decoder network, specifically designed for real-time black ice detection using thermal images. To train the network, we establish a specialized experimental platform to capture thermal images of various black ice formations on diverse road surfaces, including cement and asphalt. This enables us to curate a comprehensive dataset of thermal road black ice images for a training and evaluation purpose. Additionally, in order to enhance the accuracy of black ice detection, we propose a multi-scale dilation convolution feature fusion (MsDC-FF) technique. This proposed technique dynamically adjusts the dilation ratios based on the input image's resolution, improving the network's ability to capture fine-grained details. Experimental results demonstrate the superior performance of our proposed network model compared to conventional image segmentation models. Our model achieved an mIoU of 95.93%, while LinkNet achieved an mIoU of 95.39%. Therefore, it is concluded that the proposed model in this paper could offer a promising solution for real-time black ice detection, thereby enhancing road safety during winter conditions.

• Current Optics and Photonics
• /
• 제8권3호
• /
• pp.215-224
• /
• 2024

In medium- and long-range optical imaging systems, atmospheric turbulence causes blurring and distortion of images, resulting in loss of image information. An image-restoration method based on an adaptive feature-fusion multi-input-multi-output (MIMO) dense U-shaped network (Unet) is proposed, to restore a single image degraded by atmospheric turbulence. The network's model is based on the MIMO-Unet framework and incorporates patch-embedding shallow-convolution modules. These modules help in extracting shallow features of images and facilitate the processing of the multi-input dense encoding modules that follow. The combination of these modules improves the model's ability to analyze and extract features effectively. An asymmetric feature-fusion module is utilized to combine encoded features at varying scales, facilitating the feature reconstruction of the subsequent multi-output decoding modules for restoration of turbulence-degraded images. Based on experimental results, the adaptive feature-fusion MIMO dense U-shaped network outperforms traditional restoration methods, CMFNet network models, and standard MIMO-Unet network models, in terms of image-quality restoration. It effectively minimizes geometric deformation and blurring of images.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제13권10호
• /
• pp.5179-5196
• /
• 2019

To explore an effective non-invasion medical imaging diagnostics approach for hepatocellular carcinoma (HCC), we propose a method based on adopting the multiple technologies with the multi-parametric data fusion, transfer learning, and multi-scale deep feature extraction. Firstly, to make full use of complementary and enhancing the contribution of different modalities viz. multi-parametric MRI images in the lesion diagnosis, we propose a data-level fusion strategy. Secondly, based on the fusion data as the input, the multi-scale residual neural network with SPP (Spatial Pyramid Pooling) is utilized for the discriminative feature representation learning. Thirdly, to mitigate the impact of the lack of training samples, we do the pre-training of the proposed multi-scale residual neural network model on the natural image dataset and the fine-tuning with the chosen multi-parametric MRI images as complementary data. The comparative experiment results on the dataset from the clinical cases show that our proposed approach by employing the multiple strategies achieves the highest accuracy of 0.847±0.023 in the classification problem on the HCC differentiation. In the problem of discriminating the HCC lesion from the non-tumor area, we achieve a good performance with accuracy, sensitivity, specificity and AUC (area under the ROC curve) being 0.981±0.002, 0.981±0.002, 0.991±0.007 and 0.999±0.0008, respectively.

• 한국전자통신학회논문지
• /
• 제19권1호
• /
• pp.301-307
• /
• 2024

본 연구는 다양한 스케일 조건에서 열화상 이미지를 향상시키기 위한 새로운 캐스케이드 융합 구조를 제안한다. 특정 스케일에 맞춰 설계된 방법들은 다중 스케일에서 열화상 이미지 처리에 한계가 있었다. 이를 극복하기 위해 본 논문에서는 다중 스케일 표현을 활용하는 캐스케이드 특징 융합 기법에 기반한 통합 프레임워크를 제시한다. 서로 다른 스케일의 신뢰도 맵을 순차적으로 융합함으로써 스케일에 제약받지 않는 학습이 가능해진다. 제안된 구조는 상호 스케일 의존성을 강화하기 위해 엔드 투 엔드 방식으로 훈련된 합성곱 신경망으로 구성되어 있다. 실험 결과, 제안된 방법은 기존의 다중 스케일 열화상 이미지 향상 방법들보다 우수한 성능을 보인다는 것을 확인할 수 있었다. 또한, 실험 데이터셋에 대한 성능 분석 결과 이미지 품질 지표가 일관되게 개선되었으며, 이는 캐스케이드 융합 설계가 스케일 간 견고한 일반화를 가능하게 하고 교차 스케일 표현 학습을 더 효율적으로 수행하는 데 기여하는 것을 보여준다.

• 한국방송∙미디어공학회:학술대회논문집
• /
• 한국방송∙미디어공학회 2022년도 추계학술대회
• /
• pp.1-3
• /
• 2022

최근 기존의 영상 압축 파이프라인 대신 신경망의 종단 간 학습을 통해 압축을 수행하는 알고리즘의 연구가 활발히 진행되고 있다. 본 논문은 종단 간 학습 기반 공간적 스케일러블 압축 기술을 제안한다. 보다 구체적으로 본 논문은 신경망의 각 계층에서 하위 계층의 학습된 특징 (feature)을 융합하여 상위 계층으로 전달하는 다중 스케일 특징 융합 (multi-scale feature fusion) 모듈을 도입해 상위 계층이 더욱 풍부한 특징 정보를 학습하고 계층 사이의 특징 중복성을 더욱 잘 제거할 수 있도록 한다. 기존 방법 대비 향상 계층(enhancement layer)에서 1.37%의 BD-rate가 향상된 결과를 볼 수 있다.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제15권12호
• /
• pp.4439-4455
• /
• 2021

Compared with two-stage object detection algorithms, one-stage algorithms provide a better trade-off between real-time performance and accuracy. However, these methods treat the intermediate features equally, which lacks the flexibility to emphasize meaningful information for classification and location. Besides, they ignore the interaction of contextual information from different scales, which is important for medium and small objects detection. To tackle these problems, we propose an image pyramid network based on dual attention mechanism (DAIPNet), which builds an image pyramid to enrich the spatial information while emphasizing multi-scale informative features based on dual attention mechanisms for one-stage object detection. Our framework utilizes a pre-trained backbone as standard detection network, where the designed image pyramid network (IPN) is used as auxiliary network to provide complementary information. Here, the dual attention mechanism is composed of the adaptive feature fusion module (AFFM) and the progressive attention fusion module (PAFM). AFFM is designed to automatically pay attention to the feature maps with different importance from the backbone and auxiliary network, while PAFM is utilized to adaptively learn the channel attentive information in the context transfer process. Furthermore, in the IPN, we build an image pyramid to extract scale-wise features from downsampled images of different scales, where the features are further fused at different states to enrich scale-wise information and learn more comprehensive feature representations. Experimental results are shown on MS COCO dataset. Our proposed detector with a 300 × 300 input achieves superior performance of 32.6% mAP on the MS COCO test-dev compared with state-of-the-art methods.

• Journal of Information Processing Systems
• /
• 제19권4호
• /
• pp.427-438
• /
• 2023

Plenty of works have indicated that single image super-resolution (SISR) models relying on synthetic datasets are difficult to be applied to real scene text image super-resolution (STISR) for its more complex degradation. The up-to-date dataset for realistic STISR is called TextZoom, while the current methods trained on this dataset have not considered the effect of multi-scale features of text images. In this paper, a multi-scale and attention fusion model for realistic STISR is proposed. The multi-scale learning mechanism is introduced to acquire sophisticated feature representations of text images; The spatial and channel attentions are introduced to capture the local information and inter-channel interaction information of text images; At last, this paper designs a multi-scale residual attention module by skillfully fusing multi-scale learning and attention mechanisms. The experiments on TextZoom demonstrate that the model proposed increases scene text recognition's (ASTER) average recognition accuracy by 1.2% compared to text super-resolution network.

• 한국인공지능학회지
• /
• 제11권4호
• /
• pp.1-8
• /
• 2023

In this paper, we propose a black ice detection platform framework using Convolutional Neural Networks (CNNs). To overcome black ice problem, we introduce a real-time based early warning platform using CNN-based architecture, and furthermore, in order to enhance the accuracy of black ice detection, we apply a multi-scale dilation convolution feature fusion (MsDC-FF) technique. Then, we establish a specialized experimental platform by using a comprehensive dataset of thermal road black ice images for a training and evaluation purpose. Experimental results of a real-time black ice detection platform show the better performance of our proposed network model compared to conventional image segmentation models. Our proposed platform have achieved real-time segmentation of road black ice areas by deploying a road black ice area segmentation network on the edge device Jetson Nano devices. This approach in parallel using multi-scale dilated convolutions with different dilation rates had faster segmentation speeds due to its smaller model parameters. The proposed MsCD-FF Net(2) model had the fastest segmentation speed at 5.53 frame per second (FPS). Thereby encouraging safe driving for motorists and providing decision support for road surface management in the road traffic monitoring department.