With the development of deep learning, deep convolution neural networks for medical image segmentation tasks have become more and more complex in pursuit of higher accuracy. In most scenarios, medical image segmentati...
详细信息
ISBN:
(数字)9781665496209
ISBN:
(纸本)9781665496209
With the development of deep learning, deep convolution neural networks for medical image segmentation tasks have become more and more complex in pursuit of higher accuracy. In most scenarios, medical image segmentation pursues accuracy rather than speed, However, real-time performance is crucial in some scenarios, such as surgical navigation and diagnosis of acute stroke. So design of high-precision, lightweight and real-time medical image segmentation network has become an urgent need. To this end, a novel lightweight dual-domain network (LDD-Net) has been proposed in this paper. LDD-Net is comprised of two branches, learning respectively from the frequency domain and the spatial domain. In the frequency domain branch, the image spatial resolution is compressed via discrete cosine transform to have a large receptive field, so that better semantic context features can be learned. In the spatial domain branch, high-resolution feature representations with more details are learned. Finally, the learned features of these two branches are fused to yield high accuracy with low computational cost. The proposed method has been validated on two medical image segmentation datasets to yield the state-of-the-art performances with greatly reduced inference time and parameters of the learned models.
Today's Multispectral (MX) imaging systems contain multiple bands with unprecedented high spatial resolution and swath. Due to the complex mechanisms involved in image acquisition and data transmission for such sy...
详细信息
Convolutional neural network (CNN) has demonstrated powerful capabilities in many image/video processing tasks. In this paper, a low-complexity multi-model CNN in-loop filtering scheme is proposed for AVS3. Firstly, w...
详细信息
ISBN:
(纸本)9781665405409
Convolutional neural network (CNN) has demonstrated powerful capabilities in many image/video processing tasks. In this paper, a low-complexity multi-model CNN in-loop filtering scheme is proposed for AVS3. Firstly, we carefully choose simplified ResNet as the lightweight single model of our proposed network. Subsequently, based on the selected single model, the multi-model iterative training framework is proposed to train a multi-model filter, where the network depth and the number of multi-models are customized for different ranges of bit rate to achieve the trade-off between model performance and computational complexity. Experimental results show that our method achieves on average 6.06% BD-rate reduction on Y component under all intra configuration. Compared to other CNN filters with comparable performance, our proposed multi-model filter can significantly reduce the decoder complexity, and the experimental results indicate that the decoding time can be saved by 26.6% on average.
Burst Super-Resolution (BurstSR) attempts to restore a high resolution image from a misaligned, low-resolution RAW burst sequence. Current models for BurstSR are complex and require significant computational resources...
详细信息
Deep neural networks offer state-of-the-art technologies for highly nonlinear domains such as imageprocessing;yet their initial training requires large amounts of data, such that they are not directly suited for onli...
详细信息
ISBN:
(纸本)9781728190488
Deep neural networks offer state-of-the-art technologies for highly nonlinear domains such as imageprocessing;yet their initial training requires large amounts of data, such that they are not directly suited for online learning scenarios for streaming data where class distributions or class labels may change over time. In this contribution, we investigate the suitability of a combination of recent online learning technologies, which have been proposed for learning with streaming data and concept drift in simpler settings, and deep representations of image data as provided by deep networks trained in batch mode, to offer flexible learning technologies for streaming data from the image domain.
Traffic has been a major problem in recent times. Traffic management is a must for safer and faster transportation. Automatic smart signal controlling systems respond to day-to-day world traffic densities to provide p...
详细信息
The Electrocardiogram (ECG) signal is an important tool for cardiovascular diseases analysis. However, still today acquisition devices produce noisy signals that degrades the quality of information by corrupting impor...
详细信息
ISBN:
(纸本)9781665464956
The Electrocardiogram (ECG) signal is an important tool for cardiovascular diseases analysis. However, still today acquisition devices produce noisy signals that degrades the quality of information by corrupting important features. To improve the quality of the acquired data a filtering process is mandatory. Moreover, a real-time filtering of ECGs, in order to obtain a diagnosis as quickly as possible is a very interesting challenge. In this paper, we consider as denoising filter, the Savitzky-Golay method and we propose a parallel algorithm implementing it. The procedure exploits the computational power of Graphics processing Units (GPUs). Results in terms of performance and quality are provided.
Multiply and Accumulate (MAC) is an essential operation for domain-specific hardware accelerators used in the application domains such as digital signalprocessing, imageprocessing, and artificial intelligence. Moreo...
详细信息
This paper presents a study on using innovative machine learning techniques that can be applied in automotive traffic scenarios to increase a vehicle’s level of autonomy. The overtaking traffic scenario is treated fo...
This paper presents a study on using innovative machine learning techniques that can be applied in automotive traffic scenarios to increase a vehicle’s level of autonomy. The overtaking traffic scenario is treated for predicting the vehicle trajectory when overtaking another vehicle and the data is obtained by imageprocessing using a video camera. Two different methods are compared, first by using classic tracking methods and a Kalman filter (as an adaptive filter) and second by using a machine learning technique - Support Vector Machine. The present article uses as inputs the data received from the camera and focuses on tracking selected objects and estimating their position using mainly imageprocessing in automotive scenarios. The main purpose of this work is to experiment and compare different tracking modes to determine those that have the best performances in terms of runtime, memory usage and prediction accuracy.
In this paper, we propose an uncertainty-aware multi-resolution learning for point cloud segmentation, named PointRas. Most existing works for point cloud segmentation design encoder networks to obtain better represen...
详细信息
In this paper, we propose an uncertainty-aware multi-resolution learning for point cloud segmentation, named PointRas. Most existing works for point cloud segmentation design encoder networks to obtain better representation of local space in point cloud. However, few of them investigate the utilization of features in the lower resolutions produced by encoders and consider the contextual learning between various resolutions in decoder network. To address this, we propose to utilize the descriptive characteristic of point clouds in the lower resolutions. Taking reference to core steps of rasterization in 2D graphics where the properties of pixels in high density are interpolated from a few primitive shapes in rasterization rendering, we use the similar strategy where prediction maps in lower resolution are iteratively regressed and upsampled into higher resolutions. Moreover, to remedy the potential information deficiency of lower-resolution point cloud, we refine the predictions in each resolution under the criterion of uncertainty selection, which notably enhances the representation ability of the point cloud in lower resolutions. Our proposed PointRas module can be incorporated into the backbones of various point cloud segmentation frameworks, and brings only marginal computational cost. We evaluate the proposed method on challenging datasets including ScanNet, S3DIS, NPM3D, STPLS3D and ScanObjectNN, and consistently improve the performance in comparison with the state-of-the-art methods.
暂无评论