Neural Architecture Search Survey: A Computer Vision Perspective

A lightweight large receptive field network LrfSR for image super-resolution

Deep convolutional neural networks have demonstrated excellent performance in the field of single-image super-resolution (SISR) reconstruction. However, existing methods often suffer from issues such as large number of parameters, intensive computation, and high latency, which limit the application of deep convolutional neural networks on devices with low computational resources. To solve these problems, this paper proposes a lightweight large receptive field network for image super-resolution (LrfSR). The innovations of this paper mainly include the following aspects. Firstly, we design an information distillation module based on large receptive field (LrfDM). The module achieves large receptive field by dilated convolution, and the enlarged receptive field facilitates the network to capture more pixel-to-pixel relationships and fuse multi-scale information in the feature distillation stage. This design effectively extracts the high-frequency features of the image, which can be demonstrated by the feature map. Secondly, a more efficient attention mechanism is introduced into the network, designed as ECCA and SESA, respectively, which achieves an improvement in super-resolution image quality with fewer network parameters. Experiments on Set5, Set14, B100, Urban100 and Manga109 datasets show that the LrfSR model achieves 4-fold super-resolution Peak Signal-to-Noise Ratio (PSNR) values of 32.23 dB, 28.65 dB, 27.59 dB, 26.36 dB and 30.53 dB, which is better than the existing model LKDN etc. Meanwhile, both qualitative and quantitative experimental results show that the LrfSR model explores the potential of large receptive fields in lightweight image super-resolution networks and successfully achieves a balance between high-quality image reconstruction and limited resources. The code and models are available at https://github.com/wanqin557/LrfSR .

DDEvENet: Evidence-based ensemble learning for uncertainty-aware brain parcellation using diffusion MRI

In this study, we developed an Evidential Ensemble Neural Network based on Deep learning and Diffusion MRI, namely DDEvENet, for anatomical brain parcellation. The key innovation of DDEvENet is the design of an evidential deep learning framework to quantify predictive uncertainty at each voxel during a single inference. To do so, we design an evidence-based ensemble learning framework for uncertainty-aware parcellation to leverage the multiple dMRI parameters derived from diffusion MRI. Using DDEvENet, we obtained accurate parcellation and uncertainty estimates across different datasets from healthy and clinical populations and with different imaging acquisitions. The overall network includes five parallel subnetworks, where each is dedicated to learning the FreeSurfer parcellation for a certain diffusion MRI parameter. An evidence-based ensemble methodology is then proposed to fuse the individual outputs. We perform experimental evaluations on large-scale datasets from multiple imaging sources, including high-quality diffusion MRI data from healthy adults and clinically diffusion MRI data from participants with various brain diseases (schizophrenia, bipolar disorder, attention-deficit/hyperactivity disorder, Parkinson's disease, cerebral small vessel disease, and neurosurgical patients with brain tumors). Compared to several state-of-the-art methods, our experimental results demonstrate highly improved parcellation accuracy across the multiple testing datasets despite the differences in dMRI acquisition protocols and health conditions. Furthermore, thanks to the uncertainty estimation, our DDEvENet approach demonstrates a good ability to detect abnormal brain regions in patients with lesions that are consistent with expert-drawn results, enhancing the interpretability and reliability of the segmentation results.

TUH-NAS: A Triple-Unit NAS Network for Hyperspectral Image Classification

Over the last few years, neural architecture search (NAS) technology has achieved good results in hyperspectral image classification. Nevertheless, existing NAS-based classification methods have not specifically focused on the complex connection between spectral and spatial data. Strengthening the integration of spatial and spectral features is crucial to boosting the overall classification efficacy of hyperspectral images. In this paper, a triple-unit hyperspectral NAS network (TUH-NAS) aimed at hyperspectral image classification is introduced, where the fusion unit emphasizes the enhancement of the intrinsic relationship between spatial and spectral information. We designed a new hyperspectral image attention mechanism module to increase the focus on critical regions and enhance sensitivity to priority areas. We also adopted a composite loss function to enhance the model's focus on hard-to-classify samples. Experimental evaluations on three publicly accessible hyperspectral datasets demonstrated that, despite utilizing a limited number of samples, TUH-NAS outperforms existing NAS classification methods in recognizing object boundaries.

Bridging the Gap: Exploring Opportunities, Challenges, and Problems in Integrating Assistive Technologies, Robotics, and Automated Machines into the Health Domain

The field of healthcare is continually evolving and advancing due to new technologies and innovations [...].

Hardware-Software Partitioning for Real-Time Object Detection Using Dynamic Parameter Optimization

Computer vision algorithms implementations, especially for real-time applications, are present in a variety of devices that we are currently using (from smartphones or automotive applications to monitoring/security applications) and pose specific challenges, memory bandwidth or energy consumption (e.g., for mobility) being the most notable ones. This paper aims at providing a solution to improve the overall quality of real-time object detection computer vision algorithms using a hybrid hardware-software implementation. To this end, we explore the methods for a proper allocation of algorithm components towards hardware (as IP Cores) and the interfacing between hardware and software. Addressing specific design constraints, the relationship between the above components allows embedded artificial intelligence to select the operating hardware blocks (IP cores)-in the configuration phase-and to dynamically change the parameters of the aggregated hardware resources-in the instantiation phase, similar to the concretization of a class into a software object. The conclusions show the benefits of using hybrid hardware-software implementations, as well as major gains from using IP Cores, managed by artificial intelligence, for an object detection use-case, implemented on a FPGA demonstrator built around a Xilinx Zynq-7000 SoC Mini-ITX sub-system.

AutoML with Bayesian Optimizations for Big Data Management

The field of automated machine learning (AutoML) has gained significant attention in recent years due to its ability to automate the process of building and optimizing machine learning models. However, the increasing amount of big data being generated has presented new challenges for AutoML systems in terms of big data management. In this paper, we introduce Fabolas and learning curve extrapolation as two methods for accelerating hyperparameter optimization. Four methods for quickening training were presented including Bag of Little Bootstraps, k-means clustering for Support Vector Machines, subsample size selection for gradient descent, and subsampling for logistic regression. Additionally, we also discuss the use of Markov Chain Monte Carlo (MCMC) methods and other stochastic optimization techniques to improve the efficiency of AutoML systems in managing big data. These methods enhance various facets of the training process, making it feasible to combine them in diverse ways to gain further speedups. We review several combinations that have potential and provide a comprehensive understanding of the current state of AutoML and its potential for managing big data in various industries. Furthermore, we also mention the importance of parallel computing and distributed systems to improve the scalability of the AutoML systems while working with big data.