HSI-DeNet：深度卷积网络驱动的高光谱图像恢复技术

PDF格式 | 1.63MB | 更新于2024-07-14 | 184 浏览量 | 举报

"HSI-DeNet是利用卷积神经网络进行高光谱图像恢复的研究论文，发表在2019年2月的IEEE Transactions on Geoscience and Remote Sensing期刊上。该研究主要关注如何结合高光谱图像的光谱和空间信息进行噪声去除，包括随机噪声、结构条纹噪声和死像素/线。HSI-DeNet通过深度卷积神经网络实现这一目标，学习的滤波器能有效提取局部接收域内的空间信息，同时多个滤波器通道可以描绘光谱相关性。与传统方法相比，HSI-DeNet有三个主要优势：首先，它被视为一种基于张量的方法，直接学习各层滤波器而不破坏光谱-空间结构；其次，HSI-DeNet采用端到端训练，优化了恢复过程；最后，其能够处理不同类型和复杂程度的噪声。" 高光谱图像（HSI）包含了丰富的光谱和空间信息，这对遥感、环境监测、物体识别等领域至关重要。然而，HSI常受到各种噪声的影响，如随机噪声、结构噪声和硬件缺陷导致的死像素或死行。传统的噪声去除方法可能无法充分兼顾光谱信息和空间信息的保留。 HSI-DeNet提出的解决方案是采用深度卷积神经网络（CNN），CNN在图像处理中已经展现出了强大的特征提取能力。HSI-DeNet的网络结构设计使得滤波器能够在其局部感受野内提取空间信息，同时，通过多通道滤波器表示光谱相关性。这种方法避免了对光谱-空间结构的破坏，确保了图像恢复的准确性和完整性。 HSI-DeNet的第二个优势在于其端到端的训练策略。传统的算法通常需要分步处理，而HSI-DeNet将整个恢复过程作为一个整体进行优化，这使得模型能够更好地学习和适应复杂的噪声模式。最后，HSI-DeNet的适应性体现在它可以处理多种类型的噪声，包括非高斯噪声和结构噪声。这种灵活性使其在实际应用中更具优势，能够应对更广泛的高光谱图像恢复任务。 HSI-DeNet通过创新的CNN架构和训练策略，有效地融合了光谱和空间信息，提高了高光谱图像噪声去除的效果，对于高光谱成像领域的图像恢复技术是一次重要的进步。

670 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 57, NO. 2, FEBRUARY 2019

Fig. 1. Architecture of the proposed HSI-DeNet.

as ADMM [38], by introducing the auxiliary variable A to

solve (2)

(k+1)

= shrink

(DX

(k)

+ αJ

(k)

) (3)

where D is the sparse transformation operator, A is the

auxiliary variable that can be approximately equivalent to X,

J can be regarded as the compensating variation, α is the

regularization parameter, shrink

is the soft shrinkage operator,

and k is the iteration number.

We can observe that (1) and (3) are very similar to each

other. Both of them obtain the desired solution gradually via

a linear transformation and then nonlinear activation function.

The number of the recursion depends on the depth of the

deep model and the iteration of the optimization method. This

intrinsic similarity can partially explain why the deep model is

also suitable for the image restoration task. However, the trans-

formation parameters in the CNN model are adaptively learned

to implicitly ﬁt the distribution of the training data set, which

makes them more professional for a speciﬁc task.

B. Problem Formulation

The noise degradation model in this paper is formulated as

Y = X + N (4)

where Y ∈ R

R×C×B

is the measured HSI, R, C,andB stand

for the numbers of the row, column, and band respectively,

X is the desired clear HSI, and N is the noise in HSI, which

includes various noise components. Note that the goal of this

paper is to estimate the residual noise component N, not the

clear image, from the degraded image Y. The main reason is

that we adopt the residual learning idea from [19] and [22] to

train a residual mapping F (Y) = N. The restoration problem

is formulated as a regression task as follows:

Recon

F (Y ) − N

(5)

where F is the composite network mapping function of S.

Fig. 2. Illustration of one block. Each block contains the convolution, BN,

and nonlinear response.

C. Architecture of HSI-DeNet

In the proposed HSI-DeNet, we use a very deep con-

volutional network followed by [19], [50], and [53]. Each

convolutional layer consists of M

ﬁlters with the size of

3×3× N , except the ﬁrst and the last output layer. The channel

of the ﬁrst and the last output layer has to match the spectral

dimension of the input HSI. We use a 3×3 ﬁlters throughout

the network with stride 1, which has been demonstrated that

the decomposition of larger size ﬁlters into small-size ﬁlters

with deeper layers would make the model more discriminative

[19], [20], [22], [50], [53]. To avoid the boundary effect and

preserve the spatial size, we pad each layer with the same size

as the original image.

The architecture of the HSI-DeNet is shown in Table I.

Each block contains three components: convolutional, batch

normalization (BN), and RELU, as shown in Fig. 2. We denote

the Convolutional(C) + Batch normalization(B) + RELU(R)

block as CBR. The depth D of HSI-DeNet is 19 (including the

loss layer). The main reasons for us to choose the depth as

19 are threefold. On the one hand, the depth in the CNN model

is similar to the iteration number in optimization-based meth-

ods. Many works [20], [54] have discussed their relationship

and design their deep architecture based on the optimization

solvers. Since the iteration number of the nonconvex problem

剩余15页未读，继续阅读

weixin_38750406

粉丝: 6

HSI-DeNet：深度卷积网络驱动的高光谱图像恢复技术

hsimatlab代码-HSI-detection:用于高光谱目标/异常检测的matlab代码

hsi图像分割matlab代码-CNN_HSIC_MRF:这是用于高光谱图像分类的卷积神经网络的TensorFlow实现

MCNN-based_HSI_Classification:MCNN-CP：使用混合卷积和协方差合并的高光谱图像分类（TGARS 2021）； MCNN-PS和Oct-MCNN-PS：使用混合3D八度和2D-亚像素卷积神经网络的高光谱图像分类（已提交TGARS）

神经网络 高光谱图像

一维卷积神经网络高光谱

matlab高光谱图像处理

HSI-82C54RH-Q

高光谱成像图像python

python显示高光谱图像

最新资源

神经网络高光谱图像