Estimating human pose in complex multi-frame situations is a challenging task and has attracted intensive research by many researchers. Although 3D human pose estimation methods have achieved remarkable results in sce...
详细信息
Federated Learning (FL) has been a promising paradigm in distributed machine learning that enables in-situ model training and global model aggregation. While it can well preserve private data for end users, to apply i...
详细信息
ISBN:
(纸本)9781450397339
Federated Learning (FL) has been a promising paradigm in distributed machine learning that enables in-situ model training and global model aggregation. While it can well preserve private data for end users, to apply it efficiently on IoT devices yet suffer from their inherent variants: their available computing resources are typically constrained, heterogeneous, and changing dynamically. Existing works deploy FL on IoT devices by pruning a sparse model or adopting a tiny counterpart, which alleviates the workload but may have negative impacts on model accuracy. To address these issues, we propose Eco-FL, a novel Edge Collaborative pipeline based Federated Learning framework. On the client side, each IoT device collaborates with trusted available devices in proximity to perform pipeline training, enabling local training acceleration with efficient augmented resource orchestration. On the server side, Eco-FL adopts a novel grouping-based hierarchical architecture that combines synchronous intra-group aggregation and asynchronous inter-group aggregation, where a heterogeneity-aware dynamic grouping strategy that jointly considers response latency and data distribution is developed. To tackle the resource fluctuation during the runtime, Eco-FL further applies an adaptive scheduling policy to judiciously adjust workload allocation and client grouping at different levels. Extensive experimental results using both prototype and simulation show that, compared to state-of-the-art methods, Eco-FL can upgrade the training accuracy by up to 26.3%, reduce the local training time by up to 61.5%, and improve the local training throughput by up to 2.6x.
Sensitive data leakage has become an urgent problem to be solved as more images based functionalities are being developed in vehicles. However, there is a scarcity of evaluation for on-board videos data desensitizatio...
详细信息
ISBN:
(数字)9798350352719
ISBN:
(纸本)9798350352726
Sensitive data leakage has become an urgent problem to be solved as more images based functionalities are being developed in vehicles. However, there is a scarcity of evaluation for on-board videos data desensitization. This research analyzes on-board video desensitization process including the image pre-processing stage, sensitive area localization stage and sensitive area desensitization stage. Considering that, this paper presents several evaluation methods and metrics of sensitive target detection performance, privacy-utility evaluation of video file metadata and image, so as to provide reference for the related research on the desensitization evaluation of on-board video.
Proteins are complex biological information granules that play a crucial role in various cellular processes within living organisms. processing 3D protein structures, which are the most informative from the biological...
详细信息
While petabytes of data are generated each day by a number of independent computing devices, only a few of them can be finally collected and used for deep learning (DL) due to the apprehension of data security and pri...
详细信息
While petabytes of data are generated each day by a number of independent computing devices, only a few of them can be finally collected and used for deep learning (DL) due to the apprehension of data security and privacy leakage, thus seriously retarding the extension of DL. In such a circumstance, federated learning (FL) was proposed to perform model training by multiple clients' combined data without the dataset sharing within the cluster. Nevertheless, federated learning with periodic model averaging (FedAvg) introduced massive communication overhead as the synchronized data in each iteration is about the same size as the model, and thereby leading to a low communication efficiency. Consequently, variant proposals focusing on the communication rounds reduction and data compression were proposed to decrease the communication overhead of FL. In this article, we propose Overlap-FedAvg, an innovative framework that loosed the chain-like constraint of federated learning and paralleled the model training phase with the model communication phase (i.e., uploading local models and downloading the global model), so that the latter phase could be totally covered by the former phase. Compared to vanilla FedAvg, Overlap-FedAvg was further developed with a hierarchical computing strategy, a data compensation mechanism, and a nesterov accelerated gradients (NAG) algorithm. In Particular, Overlap-FedAvg is orthogonal to many other compression methods so that they could be applied together to maximize the utilization of the cluster. Besides, the theoretical analysis is provided to prove the convergence of the proposed framework. Extensive experiments conducting on both image classification and natural language processing tasks with multiple models and datasets also demonstrate that the proposed framework substantially reduced the communication overhead and boosted the federated learning process.
Nowadays many applications of computer vision and imageprocessing are based on Integral image (IIM) algorithm. Since the first use of IIM by Viola-Jones in the face detection algorithm, its computation is still a rea...
详细信息
ISBN:
(纸本)9781665401050
Nowadays many applications of computer vision and imageprocessing are based on Integral image (IIM) algorithm. Since the first use of IIM by Viola-Jones in the face detection algorithm, its computation is still a real challenge and receive a great attention in the research area especially in the field of real time applications. The IIM is greedy in hardware resources where the total number of additions increases with the image size and requires frequent memory accesses which result in the bottleneck effect when realized with CPU. To improve the processing efficiency, we propose a generic parallel/pipeline architectures with and without memory based on dual direction (by columns then by rows) or two levels of data flow oriented in IIM computing architecture. An interesting characteristic of the proposed architectures is the genericity of the calculation, according to the degree of parallelism (the maximum column size and the maximum row size that can be calculated in a clock cycle), the quantization of image values, on the one hand and on the other hand these architectures can be used in two functional modes. Based on the synthesis results of the implementation using Altera Quartus prime lite edition targeting an Intel/Altera Cyclone V - (FPGA), the proposed architecture achieves a high-throughput and low-area, when compared to the state-of-theart methods. More particularly, for 480x 640 image size, the proposed IIM architecture involves 765 logic registers, 643 slice LUT, and just 13,7 kbits and it operates at a maximum frequency of 182.28 MHz. These results show that this approach is one of the best candidates for portable applications that require high speed processing and low hardware resources usage.
In this paper, we have designed an innovative deep learning model using U-net with PatchGAN discriminator aiming at automatic colouring of sketch images, which provides a novel solution for sketch colouring. Firstly, ...
详细信息
ISBN:
(数字)9798350352719
ISBN:
(纸本)9798350352726
In this paper, we have designed an innovative deep learning model using U-net with PatchGAN discriminator aiming at automatic colouring of sketch images, which provides a novel solution for sketch colouring. Firstly, the image is passed through the dataset to the generator, which receives the input image and tries to convert it into the expected output image. Subsequently, in each iteration, the loss functions of the generator and the discriminator are calculated and the model parameters are updated by the optimiser. Finally, predictions are made on the test dataset and the results are plotted. By passing new unseen data to the trained generator, it was observed that the generation results were as expected and the deep learning coloured images were very close to the actual coloured images. This research brings important implications for the fields of digital art and computer vision. Firstly, it provides a completely new approach to processing sketch images, opening up more possibilities for artists and designers. Second, the model expands new ideas and examples for the application of deep learning techniques in the field of imageprocessing. Finally, in practical applications, this auto-colouring technique is expected to greatly improve productivity and can be widely used in animation, comics and other fields.
Generating realistic sketches of human faces is an important research direction in the fields of computer vision and computer graphics. However, generating high-quality sketch faces remains challenging, especially whe...
详细信息
Nowadays, it is common in many disciplines and application fields to collect large volumes of data characterized by a high number of features. Such datasets are at the basis of modern applications of supervised Machin...
详细信息
The proceedings contain 27 papers. The topics discussed include: Tetris: automatic UAF exploit generation by manipulating layout based on reactivated paths;artificial neural network for processing fingerprint image no...
ISBN:
(纸本)9798350396379
The proceedings contain 27 papers. The topics discussed include: Tetris: automatic UAF exploit generation by manipulating layout based on reactivated paths;artificial neural network for processing fingerprint image noise;facial expression intensity estimation considering change characteristic of facial feature values for each facial expression;developing a gamification method based on motivation subscales for lifelogging applications;preliminary study of reasoning existing projects' descriptions based on classname word elements;construction and evaluation of a speech emotion classifier using LSTM;concurrency control program generation in genetic programming considering depth of the program tree;reconfiguration cost for reconfigurable computing architectures;and parallel binary search tree construction inspired by thread-level speculation.
暂无评论