CUDA加速的自适应图像类比方法：基于纹理特征的高效合成

49 浏览量更新于2024-08-26 收藏 1.35MB PDF 举报

本文主要探讨了一种改进的图像类比方法，它基于自适应CUDA加速的邻域匹配框架。图像类比框架在处理非均匀输入时具有显著优势，因为它允许用户对合成结果进行创新性的控制。然而，传统的图像类比框架在处理邻域匹配时未能根据纹理内容的差异灵活运用搜索策略，导致合成速度受到限制。作者提出的新算法针对这一问题，利用CUDA技术，能够在GPU上高效执行。CUDA是NVIDIA开发的一种并行计算平台和编程模型，特别适合处理并行密集型任务，如图像处理中的大量计算。新算法在合成过程中，能够根据图像的不同纹理特征动态地结合精确的L2最近邻搜索和k-coherence搜索策略。L2距离是衡量像素间差异的标准，而k-coherence搜索则考虑了邻域内的局部一致性，这对于处理非均匀纹理区域非常关键。为了实现在GPU上的无缝切换和优化，算法采用了基于CUDA的快速k最近邻搜索算法。这样做的好处在于显著减少了预处理时间以及全局搜索过程，使得调整合成参数变得更加高效。此外，通过引入合成放大倍数，算法能够生成高分辨率的合成图像，进一步提升了运行效率。实验结果显示，该CUDA加速的图像类比方法在各种应用场景中表现出色，有效利用了GPU的并行计算能力，实现了合成速度的显著提升。这不仅提高了整体性能，还确保了生成的合成结果质量令人满意。这项工作革新了图像类比技术，使之在保持创意控制的同时，提升了合成任务的处理速度和效率。

An improved image analogy method based on adaptive CUDA-accelerated neighborhood matching 745

graph-cut synthesis approach into a global constraint min-

imization problem and supports the image analogy frame-

work.

Many efforts have been made to accelerate texture syn-

thesis. The core of pixel/patch-based texture synthesis is

the nearest neighborhood searching process. In general, the

acceleration methods can be classiﬁed into two categories,

of which one adopts tree-based search for a point in high-

dimensional space and the other utilizes image consistency

to accelerate synthesis. For the ﬁrst kind, it achieves acceler-

ation by considering the neighborhood as a point in a high-

dimensional space and regarding the neighborhood match-

ing process as a nearest point searching problem. The tree

generated by tree-structured vector quantization (TSVQ) [7]

can be used as a data structure for efﬁcient nearest point

query, but it needs large memory space for data structure.

Since querying the nearest neighbor accurately in a high-

dimensional data set is time-consuming, ANN [19] calcu-

lates the approximate most similar neighbor with relatively

long time. The latest released FLANN library [20]isre-

ported to have an order of magnitude speedup than previ-

ous ANN methods. The PCA is employed to reduce the

point’s dimensionality [3, 21] to speed the searching pro-

cess, but it leads to more feature information loss which is

proved to be detrimental to synthesis quality [4]. The precise

nearest neighborhood search performing exhaustive search-

ing is regarded to generate the superior matching results but

to be slow. The precise k nearest neighbor searching [22]

based on GPU architecture is capable to search the exact

nearest point with fast speed, which is adopted in our al-

gorithm to accelerate the pixel-based texture synthesis. An

efﬁcient exact nearest patch matching algorithm for image

processing and editing is proposed [23]. Xu et al. propose

the scheme for edit propagation by exploring adaptive clus-

tering in the afﬁnity space, which reduces both memory us-

age and time signiﬁcantly without sacriﬁcing visual ﬁdelity

[24]. As for the coherence search method, the time of k-

coherence searching [25] during synthesis is constant and

fast, the drawback is that it requires long pre-processing to

calculate the candidate set. The recently proposed Patch-

Match algorithm [26] is a fast randomized approximation

algorithm for computing the nearest neighbor ﬁeld between

two disjoint image regions, however, it may get stuck in lo-

cal minima. In recent years, the improvements of GPUs of-

fer a powerful processing platform for parallel computing

and graphic applications [27]. Lefebvre and Hoppe propose

a parallel controllable texture synthesis algorithm based on

GPU using HLSL [21], but it applies only to homogeneous

texture. They further propose appearance space texture syn-

thesis [28], using appearance vectors including non-local in-

formation rather than point-wise color to improve synthesis

quality. Our work draws ideas from parallel synthesis to ac-

celerate synthesis process and is extended to synthesize non-

homogeneous textures.

3 The improved image analogy method based on

adaptive neighborhood matching

For image analogy presented by Hertzmann [3] the neigh-

borhood matching process is performed by adopting both

approximate nearest neighbor search and coherence search

to compute the smallest distance for pixels to be synthesized,

with the parameters tuned to favor coherence over accuracy.

Since two searching strategies are employed simultaneously

for one pixel, their computation time is long. And the pa-

rameters cannot be modiﬁed once it has been set. We have

mentioned that for non-homogeneous textures, regions with

large texton structural features need to be matched within

a search space as large as possible, while for smaller tex-

ton structures we need to consider image consistency more.

Thus different textural regions need to be processed by

applying different searching strategies. In this section we

present the new neighborhood matching algorithm which

adaptively selects the appropriate searching strategies based

on textural characters of different regions.

3.1 Image synthesis algorithm based on adaptive

neighborhood matching

Before synthesis, we ﬁrst generate a user-speciﬁed label

map labelB (Fig. 1) for the synthesized result. It is a binary

image corresponding to the expected synthesis result B



with different colors representing different searching strate-

gies. As Fig. 1 shows, the black area corresponds to struc-

tural rocks region with large textons, and the white areas

correspond to random grasses region. The labelB map can

be speciﬁed by users manually according to B. Since B may

have different color regions, for these regions, we ﬁnd their

corresponding regions in A



. According to the textons’ sizes

in A, we assign different searching strategies to the regions

in B and generate the binary image labelB. Once labelB is

given, along with the input image A, A



and B, we use our

searching strategy to synthesize B



Our algorithm employs hierarchical image pyramids dur-

ing synthesis. The inputs images are A, A



, B and labelB,

and the output result is B



.ForA, A



, B, Gaussian pyramid

is generated. For labelB, the Gaussian pyramid is not appli-

cable, which makes the pixels values blur. Instead, we adopt

voting stack [15] to obtain its pyramid which keeps the bi-

nary values of each pixel. We give the pseudo-code of our al-

Fig. 1 Label map

剩余10页未读，继续阅读

weixin_38633475

粉丝: 3
资源: 946

CUDA加速的自适应图像类比方法：基于纹理特征的高效合成

自适应学习最优邻域核方法

极化SAR图像分割：自适应邻域马尔可夫场方法

基于自适应大邻域启发式搜索的无人机动态规划路径模型

基于照明自适应的邻域嵌入的幻觉

基于分层视觉语义和自适应邻域多项式潜在模型的SAR图像分割

一种基于强化学习的自适应多邻域人工蜂群算法

一种基于自适应最近邻的聚类融合方法

基于自适应最优邻域的散乱点云降噪技术研究 (2010年)

基于自适应大邻域搜索GLNS的求解旅行商问题附matlab代码+仿真结果和运行方法.zip

论文研究-一种自适应空间邻域的显著图获取方法研究.pdf

最新资源