Skip to content

Commit 065ab05

Browse files
skrisliuskrisliu
committed
x
1 parent aec33ee commit 065ab05

File tree

3 files changed

+156
-0
lines changed

3 files changed

+156
-0
lines changed

‎im/mdl4ow1.png‎

262 KB
Loading

‎im/mdl4ow2.png‎

257 KB
Loading

‎index.html‎

Lines changed: 156 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,156 @@
1+
<!DOCTYPE html>
2+
<html lang="en">
3+
4+
<head>
5+
<meta charset="UTF-8" />
6+
<meta name="viewport" content="width=device-width, initial-scale=1" />
7+
<title>MDL4OW Open-Set Hyperspectral Image Classification. Few-shot Hyperspectral Image Classification With Unknown
8+
Classes Using Multitask Deep Learning.</title>
9+
<link rel="stylesheet" href="https://skrisliu.com/css/font.css">
10+
<link rel="stylesheet" href="https://skrisliu.com/css/style.css">
11+
<style>
12+
.highlight2 {
13+
padding: 1rem;
14+
background-color: #e5e7eb;
15+
}
16+
</style>
17+
</head>
18+
19+
<body>
20+
21+
<div class="content">
22+
23+
<h2 class="content-title">
24+
MDL4OW Few-shot Hyperspectral Image Classification With Unknown
25+
Classes Using Multitask Deep Learning.
26+
</h2>
27+
28+
<h4>Open-Set Hyperspectral Image Classification. </h4>
29+
30+
<p class="content-meta">Source code and annotations for:</p>
31+
<p class="highlight2">Shengjie Liu, Qian Shi, and Liangpei Zhang. Few-shot Hyperspectral Image Classification
32+
With Unknown Classes Using Multitask Deep Learning. IEEE TGRS, 2020. <a
33+
href="https://doi.org/10.1109/TGRS.2020.3018879" target="_blank">doi:10.1109/TGRS.2020.3018879</a></p>
34+
35+
<p class="content-meta">Contact: skrisliu AT gmail.com</p>
36+
37+
<p class="content-meta" style="font-size: 1.1em; text-align: left; margin: 1.5em 0;">
38+
Code and annotations are released here, or check out <a href="https://github.com/skrisliu/MDL4OW"
39+
target="_blank">https://github.com/skrisliu/MDL4OW</a>
40+
</p>
41+
42+
<hr>
43+
<h2>Overview</h2>
44+
<h3>Ordinary: misclassify road, house, helicopter, and truck</h3>
45+
<p>
46+
Below is a normal/closed classification. If you are familiar with hyperspectral data, you will notice some
47+
of the materials are not represented in the training samples. For example, for the upper image (Salinas
48+
Valley), the road and the houses between farmlands cannot be classified into any of the known classes. But
49+
still, a deep learning model has to assign one of the labels, because it is never taught to identify an
50+
unknown instance.
51+
</p>
52+
53+
<p>
54+
<a href="im/mdl4ow1.png" target="_blank">
55+
<img src="im/mdl4ow1.png" alt="ordinary classification" width="50%">
56+
</a>
57+
</p>
58+
59+
<h3>What we do: mask out the unknown in black</h3>
60+
<p>
61+
What we do here is, by using multitask deep learning, empowering the deep learning model with the ability to
62+
identify the unknown: those masked with black color.<br>
63+
For the upper image (Salinas Valley), the roads and houses between farmlands are successfully
64+
identified.<br>
65+
For the lower image (University of Pavia Campus), helicopters and trucks are successfully identified.
66+
</p>
67+
68+
<p>
69+
<a href="im/mdl4ow2.png" target="_blank">
70+
<img src="im/mdl4ow2.png" alt="MDL4OW result" width="50%">
71+
</a>
72+
</p>
73+
74+
<hr>
75+
<h3>Key packages</h3>
76+
<pre class="highlight2">
77+
tensorflow-gpu==1.9
78+
keras==2.1.6
79+
libmr
80+
</pre>
81+
<p>Tested on Python 3.6, Windows 10</p>
82+
<p>Recommend Anaconda, Spyder</p>
83+
84+
<hr>
85+
<h2>How to use</h2>
86+
<h4>Hyperspectral satellite images</h4>
87+
<p>The input image is with size of imx*imy*channel.</p>
88+
<p>The satellite images are standard data, downloaded here: <a
89+
href="http://www.ehu.eus/ccwintco/index.php/Hyperspectral_Remote_Sensing_Scenes"
90+
target="_blank">http://www.ehu.eus/ccwintco/index.php/Hyperspectral_Remote_Sensing_Scenes</a></p>
91+
<p>The above data is in matlab format, the numpy format can be found here (recommended):<br>
92+
<a href="https://drive.google.com/file/d/1cEpTuP-trfRuphKWqKHjAaJhek5sqI3C/view?usp=sharing"
93+
target="_blank">https://drive.google.com/file/d/1cEpTuP-trfRuphKWqKHjAaJhek5sqI3C/view?usp=sharing</a>
94+
</p>
95+
96+
<h4>Quick usage</h4>
97+
<p class="highlight2">python demo_salinas.py</p>
98+
99+
<h4>Arguments</h4>
100+
<div class="highlight2">
101+
<p><strong>Command-line Arguments:</strong></p>
102+
<ul>
103+
<li>
104+
<code>--nos</code>: Number of training samples per class<br>
105+
<small>20 for few-shot learning, 200 for many-shot learning</small>
106+
</li>
107+
<li>
108+
<code>--key</code>: Dataset name<br>
109+
<small>Options: <code>'salinas'</code>, <code>'paviaU'</code>, <code>'indian'</code></small>
110+
</li>
111+
<li>
112+
<code>--gt</code>: Path to ground truth file
113+
</li>
114+
<li>
115+
<code>--closs</code>: Classification loss weight<br>
116+
<small>Default: <code>50</code> (equivalent to 0.5 in normalized scale)</small>
117+
</li>
118+
<li>
119+
<code>--patience</code>: Early stopping patience<br>
120+
<small>Stop training if loss doesn't decrease for <code>{patience}</code> consecutive epochs</small>
121+
</li>
122+
<li>
123+
<code>--output</code>: Directory path to save output files<br>
124+
<small>Includes: trained model, prediction probabilities, predicted labels, reconstruction
125+
loss</small>
126+
</li>
127+
<li>
128+
<code>--showmap</code>: Save classification map as image<br>
129+
<small>When enabled, generates and saves the predicted label map visualization</small>
130+
</li>
131+
</ul>
132+
</div>
133+
134+
<hr>
135+
<h3>Evaluation code updated on 18 May 2021</h3>
136+
<p>When using the evaluation code "<code>z20210518a_readoa.py</code>", you should change the parameter
137+
"<code>mode</code>" for different settings. The inputs are output files from the training script.</p>
138+
139+
<h4>Mode</h4>
140+
<div class="highlight2">
141+
<p><strong>Mode Selection:</strong></p>
142+
<ul>
143+
<li><code>mode == 0</code>: Closed-set classification</li>
144+
<li><code>mode == 1</code>: MDL4OW (Minimum Distance to Learned Open-set Weights)</li>
145+
<li><code>mode == 2</code>: MDL4OW/C (with confidence calibration)</li>
146+
<li><code>mode == 3</code>: Closed-set with probability output</li>
147+
<li><code>mode == 4</code>: Softmax with threshold</li>
148+
<li><code>mode == 5</code>: OpenMax (for open-set recognition)</li>
149+
</ul>
150+
</div>
151+
152+
</div>
153+
154+
</body>
155+
156+
</html>

0 commit comments

Comments
 (0)