0:00:16thank you for introduction basic uh my name is wendy no postdoc from ghent university
0:00:24that they i'm gonna put in that use something about this classification of hyperspectral image
0:00:31oh would happen easily based on extended morphological profiles responsive reconstructions
0:00:39different from the at normal actually image hyperspectral sensor cup two dates bits hundred append
0:00:47but it's pixels
0:00:50hot very hot it's hot continuing with reflections different materials have different spectral reflections
0:01:00and we can identify different materials according to its reflections for instance here
0:01:10according to this spectrum reflects the reflections we notice these soil and the this reflections
0:01:18basis what the
0:01:21for some hyperspectral image
0:01:25the spectral reflections of some of pets are similar when only using spectral information to
0:01:35do classification we cannot get sucked is final result spatial information it's very important spatial
0:01:43information such as object size shape and texture it's very important to improve the classification
0:01:51result
0:01:52when they called by
0:01:55both spectral information and the spatial information we can improve the classification accuracy a lot
0:02:05this is all light of my presentation today first i will give a lot of
0:02:11you about feature extractions at the morphological profiles which are widely used to explore ohms
0:02:20the spectral and spatial information of hyperspectral image then i mean in that use all
0:02:26propose it can be morphological profiles we sparse reconstructions and show some experimental results of
0:02:35the last but not least it conclusion and future work
0:02:42hyperspectral image it's difficult to process due to its high dimensionality
0:02:49feature extraction
0:02:52it's use it's always used as preprocessing to reduce the dimensionality and the lead then
0:02:59there's the of the hyperspectral states
0:03:04uh the
0:03:06pretty simple components and the cannot pretty simple can note component devices are most widely
0:03:12my set used to do basic directions
0:03:18principal component analysis extracted features on which the tomb melody alliance
0:03:27it's not seem to maximize
0:03:31the first few extracted features can you percent most of a information of heidi by
0:03:38original hyperspectral states and the last are mainly noise for instance here
0:03:46the first we extracted pretty simple components can be presented ninety nine percent of Q
0:03:53melody but lines of this one hundred and three bands hyperspectral target
0:04:01and the so as the identity i dependence component and
0:04:08to achieve the same ninety nine percent
0:04:12Q melody lines can opening simple component needs poles
0:04:19twelve can uh features extracted features but this as a nonlinear mess that's kernel principal
0:04:26component the licensees can proceed that non linear distribution of that eight
0:04:33and uh you see even on the total extracted features it's have some detail features
0:04:40but for pretty simple and the linear ica in the in its twelve extracted features
0:04:48maybe the noise
0:04:54uh spatial information it's important especially for the high low solution up and hyperspectral image
0:05:01in particular morphological profiles it's one of the most popular myself which are used to
0:05:07explore the spatial information of hyperspectral image
0:05:12that's take morphological closing as example to see how morphological profiles pretty C the spatial
0:05:20information of hyperspectral date
0:05:24in morphological closings a structure it demands are moved in inside the image when there's
0:05:33the size of the dark object it's more that then these structural elements
0:05:42then the object will be removed
0:05:46at the size of this structure demands increase more and more dark
0:05:54small object uh disappear in the disappear in this image
0:06:01and then the size of a small object only like it this the size of
0:06:07the strict demons
0:06:11so we say that's morphological profiles contain yeah yes thank you
0:06:19uh yeah thank you thank you much yeah
0:06:23and the this size it's like you like it to the size of the structure
0:06:28you meant
0:06:31and the when the size of the structure immense increase
0:06:36more and more objects more dark objects disappear we can see also the shape of
0:06:44objects also chant
0:06:47for the corn the police like tango
0:06:51column is now a change to be lumped
0:07:09well um
0:07:16okay i was despising
0:07:18that's take
0:07:20one early example on how the a hyperspectral image to see the results when the
0:07:31size of structure demands increase we see a lot of
0:07:37bright
0:07:38small objects disappear i don't know what you can see these uh small disappear uh
0:07:47this one this one this one and the a small bright small objects disappear us
0:07:54the size of structuring elements increase
0:07:58but we can also see that the shape of this object at change
0:08:04you see this lack tango object now the corner it's long it long
0:08:14in order to proceed the shape of objects while always use morphological profiles these reconstructions
0:08:24this christmas that the all the site the shape of objects can be preceded by
0:08:29the then this know the constructions
0:08:33we see here this object
0:08:37the each of these black tango object it's we'll proceed how well the we see
0:08:43that some small object which are expected to disappear now
0:08:50state present in the image for this small loads point load by load which are
0:08:59disappear in here we solve the constructions but now disappear we can say that this
0:09:07reconstructions the size of object mott proceeds about the remotes proceed the well
0:09:16in order to overcome all these problems morphological profiles this posture become structures are used
0:09:25this is not set we can see us the size of the structure it immense
0:09:31increase
0:09:33more and more small point
0:09:35objects disappear corresponding to i
0:09:39structuring elements size at the same time we can proceed the age of the object
0:09:45button then result the constructions
0:09:55so we propose
0:09:57uh extended morphological profiles this posture become could decode reconstructions to combine spectral and the
0:10:04spatial information of the hyperspectral image
0:10:09this is our proposed by set we take features extracted by pca as example to
0:10:16build extended morphological profiles responsive reconstructions we first use the feature extractions delete use that
0:10:25dimensionality and the led dependency of the hyperspectral data and then repute we use all
0:10:33morphological openings and closings this partial reconstruction is to build morphological profiles on each extracted
0:10:42features
0:10:43and then we stack all these features as well
0:10:48as important as the svm classifier and then finally we got the classification maps
0:10:57that's take one example take two extracted features as pca to see we take two
0:11:06we take two openings and the closings and the we start these features us
0:11:17uh so but uh this is extended morphological profiles and the we use this as
0:11:22input to do the classification
0:11:26we can see by increasing the size of the structure immense more and more objects
0:11:32are removed
0:11:34and the extended morphological profiles county the spatial information such as the size of objects
0:11:41in the image
0:11:44the morphological profiles contain
0:11:47yeah O explained this morphological profiles contain four openings and the full closings
0:11:55and the this morphological profile with partial reconstruction is was proposed in to solve i
0:12:02it's with peak
0:12:06last row some experimental results to see if wishes all of all proposed method
0:12:14the four states we use it's problem
0:12:18university of probably i italian the states have one hundred and the fifteen bands
0:12:25some bands are removed due to the noise leaving one hundred and sleep and on
0:12:30the left this is the false color image it's
0:12:35this image has high spatial resolution is one point three meters topic cells the online
0:12:42classes in this image
0:12:47we select ten labels impose a class landlady to change svm classifier based
0:12:55a pf kernels
0:12:58and then test on the testing on the test set
0:13:02morphological features up builds on the first poll can open a simple components
0:13:11and the lease is the results we used original hyperspectral image to do the classification
0:13:17and this results it's
0:13:19we use that we the morphological profile with no reconstruction is and this is the
0:13:25morphological profile with reconstruction and this is all proposed which posture because structures
0:13:32phone to be zero maps we can see that's this more formal logical features the
0:13:39classification maps it's much smoother than results what spatial information
0:13:47this is noise
0:13:48and the compared to the morphological risk we construct use morphological profiles result reconstructions and
0:13:57vese posture because actors a much and smooth then this one
0:14:03and the full morphological profiles with no reconstruction is the shape of some object a
0:14:10chant
0:14:12for this model the corner here it's no longer
0:14:17and we can also compare the overall classification accuracy here
0:14:23here this the morphological features the overall accuracy can improve a lot even then what
0:14:31kind that sense and that all proposed mess that what's the best results
0:14:38and the beacon please see the size and shape but the then some existing methods
0:14:47let's see some results when the change simple size increase
0:14:52the extended morphological profiles built on different features because C
0:14:59the morphological features proposed you on the
0:15:03pretty simple component features
0:15:06as the simple size increase all proposed it can be morphological profiles this posture reconstructions
0:15:13got busted results
0:15:16and
0:15:18the morphological profiles you on kernel pretty simple components and lies is always get the
0:15:24possibly start
0:15:27in some cases it is button want to use the morphological profiles with reconstructions for
0:15:33here this morphological profile
0:15:38we speak objections cannot will be will be was then this no reconstructions this is
0:15:46due to the old reconstructions of morphological profiles by reconstruction
0:15:57let's see and not the plates that's a list that's it's prone city centre a
0:16:03party a
0:16:04it's has one hundred fifteen bands and some bad a limited due to the noise
0:16:09leaving one hundred and two bands on the left side it's the false color image
0:16:15is it also have high spatial resolution is one point three meters but excels yeah
0:16:22nine different classes in this image
0:16:27we select pen labels impose a class blunder the problem this training set
0:16:35and the land passed on the touch screen set of this
0:16:41morphological features appear on the top the first can cut kernel principal components
0:16:50and the this is the result from people and that's that for a while what
0:16:53dates we just use only the original hyperspectral data to do the classification and this
0:17:00is no reconstruction this is this reconstruction and this is with all proposed
0:17:06posture reconstructions because shape on that a little overall classification accuracy all postmaster what's the
0:17:15highest accuracy
0:17:22the see the performance when the change simple size increases
0:17:28but the changes impose size increase all propose
0:17:33extended morphological profiles with partial reconstructions got busted results and that this is always you
0:17:41on the kernel principal components
0:17:49so that's see the conclusion and the future work
0:17:53uh in this presentation we can see that's kernel pretty simple component and vice is
0:17:59more efficient to extracted features for comes construct in uh extended morphological profiles and the
0:18:09extended morphological profiles with partial reconstruction is more efficient competing T then lowest be snow
0:18:17constructions and the miss cost reconstructions
0:18:21and the extended morphological profiles responsive reconstructions europe on kernel pretty simple components are always
0:18:29perform fast especially before there's more training symbols
0:18:34all future work maybe explore different structure meant structure element because now in this presentation
0:18:43and the in my complex paper we just use the circle structure element and that
0:18:51we can see also this kernel principal component we need more features
0:18:57so it is but that that's it's great eight high dimensionality of the extended morphological
0:19:05profiles this posture reconstructions
0:19:10uh that's my presentation thank you
0:19:28yeah
0:19:37uh yes uh i prepared to try the linear structure elements which may efficient for
0:19:44the loans for the streets and the label of the streets they are long lines
0:19:49and the for the labels also like this one i think it's uh maybe it's
0:19:54nice feature to improve the classification accuracy
0:19:58yes
0:20:37a yes uh yes we can use some computers before the classification this may have
0:20:44some i think maybe some the same result because you smoother the original hyperspectral image
0:20:53and the this different scalars you also can smooth the image
0:20:59in different
0:21:00uh how this yeah maybe can see that that's what if devoicing possibly denoising the
0:21:07hyperspectral image and that we use that you noise image to do the classification
0:21:12yes
0:21:51so secure for suggestion i also read some level as follow your group yes uh
0:21:57the extent it's attribute profiles yes it's but uh and the is we because see
0:22:04that because uh it's sometimes it not related with this structure you meant so in
0:22:10may be our future work we can see the we use the attributes uh profiles
0:22:15to improve the classification accuracy that's good for test thank you
0:22:36uh
0:22:43yeah uh we normally we selected this according to the queue mel other human let
0:22:50steve
0:22:51cumulative but like the lions
0:22:54for pca for kernel principal components the first we extracted features three percent ninety nine
0:23:01percent of the two melodies alliance so for kind of pca how the for stole
0:23:09components a reasonable kernel principal components people are sent ninety nine
0:23:15a sense of the cumulative alliance so we just depend on the
0:23:22yeah i don't different dates that's we have different
0:23:26the different number of the extracted features we use yeah
0:23:33yeah
0:23:36yeah that's this also creates a new challenge because you also the dimensionality of the
0:23:43extended morphological profiles will be higher than other than principal component like this
0:24:00i