Recently, deep learning algorithm "U-Net" started to be used in several fields. Image segmentation and Image super resolution is included in that several fields. But I don't understand how one algorithm can be used for both image segmentation and image super resolution.
If we google about U-Net most of result will show result of image segmentation. But if you look at this website, we can see that U-Net is also used for Image super resolution.
https://medium.com/datatype/superresolution-by-unet-and-customized-feature-loss-79b469aa5bd3
How can this be possible? I mean segmentation and super resolution for me is completely different field.
Related
I want to implement an application, that is able to recognize pictures from camera input. I don't mean classification of objects, but rather detecting the exact single image from given set of images. So if I for example have an album with 500 pictures, then if I point a camera to one of them, then application will be able to tell it's filename. Most of tutorials I find about CoreML is strictly for image classification (recognizing class of object) and not about recognizing exact image name in camera. This needs to work from different angles as well, and all I can have for training the network is this album with many different pictures (single picture for single object). Can this be somehow achieved? I can't use ARKit Image Tracking, because there will be about 500 of these images, and I need to find at least a list of similar ones first with CoreML / Vision.
I am not sure, but I guess perceptual hashing might be able to help you.
It works in a way that it makes some fingerprint from the reference images, and for a given image, it extracts the fingerprints as well, and then you can find the most similar fingerprints.
in this way, even if the new image is not 100% as the image in the dataset, you still can detect it.
It is actually not very hard to implement. but if you would like, i think phash library is a good one to use.
I'm very new to AEM and I'm trying to reduce the image size by using named transform image servlet in AEM 6.3. After reducing the image size using resize property, the image quality goes down terribly.
I've also tried using attributes sharpen and quality however I'm not able to understand their proper usage even after going through the adobe docs.
This is the original image resolution
1600 x 530
This is the image configuration that I've tried so far:
bounded-resize:width=1423&height=471&upscale=true (using this only reduces the image quality a lot)
quality:quality=82 (this changed nothing)
sharpen:op_usm=3.5,20(this distorted the image completely, sort of oil paint effect)
Above configuration maintains the aspect ration but not the quality.
Need to know what wrong I've done in this and how do I REDUCE THE IMAGE SIZE WHILE MAINTAINING THE ASPECT RATIO AND IMAGE QUALITY.
You should check the order of Image Transformers in your transformer configuration. The documentation states:
Order matters when defining your image transformation rules. For example, a resize then crop can yield significantly different results than a crop then resize.
I am not a 100% sure which makes more sense: Resize and Crop or Crop and Resize. But that would be easy to validate:
Create two transformer configurations. Both should just contain the resize and crop setting but in a different order. Call both of them (with the right URL) and then compare the resulting images.
how do I REDUCE THE IMAGE SIZE WHILE MAINTAINING THE ASPECT RATIO AND
IMAGE QUALITY.
As a best practice, to achieve a high image quality and small file size, start with the following combination of parameters:
fmt=jpg&qlt=85,0&resMode=sharp2&op_usm=1.75,0.3,2,0
This combination of settings products excellent results under most circumstances.
For further reading and learning, Refer to this
Closed. This question does not meet Stack Overflow guidelines. It is not currently accepting answers.
This question does not appear to be about programming within the scope defined in the help center.
Closed 8 years ago.
Improve this question
I'm working on ways of collecting farm aerial images (images collected from a helicopter in a perpendicular fashion) that I'd want to stitch them together to build the whole photo of the area that's being covered and then I wanted to run analytics.
I'm assuming the images will come with [latitudue, longitude] coordinates, to help me determine the spots to place the images.
To understand the issues with this technology, I tried manually stitching pictures taken from my phone of some sample area in my back yard. I experienced that the edges don't usually look the same because they are being seen by the camera from different sides or angles. I guess this is a distortion in image that could potentially be fixed by ortho-rectification (not completely sure).
I quickly created the following picture to help explain my problem.
My question to you:
What are the algorithms/techniques used to do ortho-rectification?
What tools would best suit my needs: opencv, or processing or matlab or any other tool that could easily help in rectification of images and creating a mosaic photo?
What other issues should be considered in doing aerial imagery mosaic-ing and analytics?
Thank you!
Image stitching usually assumes that the camera center is fixed across all photos, and uses homographies to transform the images so that they seem continuous. When the fixed camera center assumption is not strictly valid, artifacts/distortions may appear due to the 3D of the scene. If the camera center moved by a small distance compared to the relief of the scene, "seamless image blending" techniques may be sufficient to blur out the distortions.
In more extreme cases, ortho-rectification is required. Ortho-rectification (Wikipedia entry) is the task of transforming an image observed from a given perspective camera into an orthographic (Wikipedia entry) and usually vertical point of view. The orthographic property is interesting because it makes the stitching of several images much easier. The following picture from Wikipedia is particularly clear (left is an orthographic or directional projection, right is a perspective or central projection):
The task of ortho-rectification usually requires having a 3D model of the scene, in order to map appropriately intensities observed by the perspective camera to their location with respect to the orthographic camera. In the context of aerial/satellite images, Digital Elevation Models (DEM) are often used for that purpose, but generally have the serious drawback of not including man-made structures (only Earth relief). The NASA provides freely the DEM acquired by the SRTM missions (DEM link).
Another approach, if you have two images acquired at different positions, you could try to do a 3D reconstruction using one of the stereo matching technique, and then to generate the ortho-rectified image by mapping the two images as seen by a third orthographic and vertical camera.
OpenCV has several interesting function for that purpose (e.g. stereo reconstruction, image mapping functions, etc) and might be more appropriate for intensive usage. Matlab probably has interesting functions as well, and might be more appropriate for quick tests.
First, rectification is some kind of warping, but not the one that you need. Regular rectification is used in stereo to ensure that matching points lie on the same row - not your case. Ortho-rectification warps perspective projection into orthographic - again not your case. Not only you lack a 3D model for this warping to calculate but also you dont need it since your perspective distortions are negligible and you already have image pretty close to ortho ( that is when the size of the objects is small compared to the viewing distance perspective effects are small).
You problems in aligning two images stem from small camera rotations between shots. To start fixing the problem you need to ensure that your images actually overlap by say 30%. To read about this see chapter 9 of this book.
What you need is to review a regular image stitching techniques that use homography to map two images. Note that doing so assumes that images are essentially flat. To find homography you can first manually select 4 points in one image and 4 matching points in another image and run openCV function findHomography(). Note that overlap is required to find the matches (in your picture there is no overlap). warpPerspective() can warp images for you after homogrphy is found.
"What are the algorithms/techniques used to do ortho-rectification?
If you want a good overview of the techniques, then the book "Multiple View Geometry in Computer Vision" by Hartley & Zisserman might be a good place to start: http://www.robots.ox.ac.uk:5000/~vgg/hzbook/
Andrew Zisserman also has some tutorials available here at www.robots.ox.ac.uk/~az/tutorials/ which might be more accessible/make it easier for you to find the particular technique you want to use.
"What tools would best suit my needs: opencv, or processing or matlab or any other tool that could easily help in rectification of images and creating a mosaic photo?"
OpenCV has a fair few number of tools available - take a look at Images stitching for starters. There's also a lot available for correcting distortion. However, it doesn't have to be the tool you use, there are others!
I am trying to detect exact silhouette of human body in this dataset using background subtraction. After doing some thresholding I was getting split blobs so I looked at this tutorial by Steve but now I am getting blob other that human body as shown below
So here is the original
After Subtracting it from background, background was considered as the first frame of the video, so after subtracting it from orignal image I get the following image
so I did basic thresholding and I get the following image, which is split from further areas
and using Steve's method I get this
But this contains a lot of area which is not a part of human body, any suggestion if somehow or using edges I can get good blob of human body.
EDIT
As #lennon310 asked me to upload color image so here it is
and as #NKN asked me to upload edge information of the same image so here it is
Instead of literally subtracting the background, try using the vision.ForegroundDetector object, which is part of the Computer Vision System Toolbox. It implements the mixture-of-gaussians adaptive background modeling, and it may give you a cleaner segmentation.
Having said that, it is very unlikely that you will get the "exact" silhouette. Some error is inevitable.
In your result image, you have tow types of black regions. one is moving and the other is stationary.
So when you you want to fill the human body, you have to choose only the moving region, for this purpose, I suggest to segment your image by adding optical flow technique to know where the moving regions are.
This is an interesting tutorial doing what you need to do:
http://docs.opencv.org/trunk/doc/py_tutorials/py_video/py_lucas_kanade/py_lucas_kanade.html
I am currently studying image processing and learning matlab for my project.
I needed to know that if there is any method to detect a car from traffic image or parking lot image and then segment it out from it.
I have googled a lot but mostly the content is video based and I dont know anything about image processing.
language prefered : MATLAB
I am supposed to do this on images only not videos.
It's a very difficult problem in general. I'd suggest the easier way is to constrain the problem as much as possible - control lighting, size orientation of cars to detect, no occlusions.
This constraining has been the philosophy image processing has followed up until recently. Now the trend is that instead of constraining your problem, obtain as a massive amount of example data to train a suppervised learning algorithm. In fact it's possible that you can use a pre-trained model that would let you detect cars as it has been suggested in a previous answer.
There has been recently massive progress in the area of object detection in images and here are a few of the state of the art approaches based on neural network based approaches:
OverFeat
Rich feature hierarchies for accurate object detection and semantic segmentation (R-CNN paper)
Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition (paper)
Framework that you could use include:
Caffe: http://caffe.berkeleyvision.org/
Theano
Torch
You can use the detection by parts method:
http://www.cs.berkeley.edu/~rbg/latent/
It contains a trained model for "car" which you can use to detect cars, surround then with a bounding box and then extract them from the images.