深度神经网络图像目标检测算法综述

随着深度卷积神经网络优异的特征提取能力被发掘, 目标检测的进程开始以一种势不可挡的姿态向前推进, 同时, 和深度学习结合的目标检测技术取得了显著的成果, 在自动驾驶、智能化交通系统、无人机场景、军事目标检测和医学导航等现实场景中得到了广泛的应用. 本文回顾了传统目标检测算法的缺点, 介绍了常用的检测数据集以及性能评估指标, 综述了基于深度学习的目标检测经典算法, 阐述了当前目标检测的以及存在的困难与挑战, 对目标检测的未来可行的研究方向进行了展望.     

基于深度学习的行人重识别综述

行人重识别主要研究在不同摄像机拍摄的图形中检索目标行人的任务,是计算机视觉领域一个极具挑战性的研究课题.传统依赖手工特征的行人重识别方法性能低且鲁棒性差,不能适应数据爆炸增长的信息时代.近年来,随着大规模行人数据集的出现和深度学习的迅速发展,行人重识别研究取得了许多突出成果.梳理了性能接近饱和的有监督学习研究方法,并探...     

Self-adaptive sampling rate assignment and image reconstruction via combination of structured sparsity and non-local total variation priors

Compressive sensing (CS) is an emerging approach for acquisition of sparse or compressible signals. For natural images, block compressive sensing (BCS) has been designed to reduce the size of sensing matrix and the complexity of sampling and reconstruction. On the other hand, image blocks with varying structures are too different to share the same sampling rate and sensing matrix. Motivated by this, a novel framework of adaptive acquisition and reconstruction is proposed to assign sampling rate adaptively. The framework contains three aspects. First, a small part of sampling rate is employed to pre-sense each block and a novel approach is proposed to estimate its compressibility only from pre-sensed measurements. Next, two assignment schemes are proposed to assign the other part of the sampling rate adaptively to each block based on its estimated compressibility. A higher sampling rate is assigned to incompressible blocks but a lower one to compressible ones. The sensing matrix is constructed based on the assigned sampling rates. The pre-sensed measurements and the adaptive ones are concatenated to form the final measurements. Finally, it is proposed that the reconstruction is modeled as a multi-objects optimization problem which involves the structured sparsity and the non-local total variation prior together. It is simplified into a 3-stage alternating optimization problem and is solved by an augmented Lagrangian method. Experiments on four categories of real natural images and medicine images demonstrate that the proposed framework captures local and nonlocal structures and outperforms the state-of-the-art methods.     

基于深度学习和Transformer的目标检测算法

目标检测是实现目标跟踪、实例分割等高级视觉任务的基础,在智慧交通、缺陷检测、智能安防等现实场景有着重要应用。现有高精度检测算法都是在深度学习的指导下实现,同时伴有锚框技术,但是锚框自身的不足对检测器性能有着较大影响,无锚点碰撞检测成为了近几年目标检测领域新的研究方向。与此同时,Transformer表现出的巨大潜力为视觉领域开辟了图像与Transformer结合这个新方向,基于Transformer的目标检测也成为一个新的研究热点。系统地总结了深度学习时代的目标检测算法,调查并研究了近五年目标检测的相关论文,重点从Anchorfree和Transformer两个角度对这些算法进行深入分析,介绍了这些算法在现实场景具体应用情况以及目标检测领域常用数据集,基于目前的研究现状对目标检测的未来可研究方向进行了展望。