Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 6 years ago.
Improve this question
I have some images i'm using for face recognition.
Some of the images are very dark.
I don't want to use Histogram equalisation on all the images only on the dark ones.
How can i determine if an image is dark?
I'm using opencv in python.
I would like to understand the theory and the implementation.
Thanks
To determin if an image is dark, simply calculate the average intensity and judge it.
The problem for the recognition although is not that the image is dark, but that it has a low contrast. A bright image with the same contrast would yield the same bad results.
Histogram equalization is a method that is used to improve images for human vision. Humans have difficulties to distinguish between very similar intensity values. A problem that a computer does not have, unless your algorithm is somehow made to mimic human vision with all its flaws.
A low contrast image bears little information. There is no image enhancement algorithm in the world that will add any further information.
I won't get into too much detail about image characterization. You'll find plenty of resources online or in text books.
A simple measure would be to calculate the standard deviation of image regions you are interested in.
You can try using CLAHE instead of the using histogram equalization. I am not sure if it will work for all your images (if you post some images I can try), however, conceptually, it should work better than "global" histogram equalization.
Related
Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 5 days ago.
Improve this question
I have a stock of photos, of a spinning disk of varying angles. I wish to find the edge of the top of the disk. The top is colored in a distinct black color in comparison to the rest of the photo.
A
B
I first tried using canny edge detection which does a decent job, but also identifies the bottom half of the disk, which I wish to avoid.
Next idea was to use the distinct black color- perhaps by dividing the photo into domains characterized by colors\intensities- and by choosing the largest domain\most black domain or some other parameter, perhaps to isolate that black circle, and only then to use canny edge detector.
Is there any existing function that can divide a greyscale image into domains? Transferring from matlab to python, so I'm new to it's syntax and functions.
Thanks
The Canny disaster goes on !
People playing with image processing keep willing to rely on edge detection when they have beautifully segmenting scenes. With a careful selection of a binarization threshold, you can extract the ellipse as a single piece.
Closed. This question needs details or clarity. It is not currently accepting answers.
Want to improve this question? Add details and clarify the problem by editing this post.
Closed 1 year ago.
Improve this question
I have an analog speedometer image, with the needle pointing to the current speed. I am trying to find a way to get the speed that the needle is pointing out to. I tried using HoughCircles() from OpenCV, but it is throwing an error as the image contains only speedometer and which is a semi-circle. Any resources to help me move forward will be appreciated.
Assuming the needle will have a different colour to the rest of the speedometer OR its size is distinctly larger than the rest of the elements on the speedometer (which is often the case), I'll do something like below.
Convert the image to grayscale.
Apply colour thresholding (or size-based thresholding) to detect the pixel area representing the needle.
Use HoughLines() or HoughLinesP() functions in OpenCV to fit a line to the shape you detected in Step 2.
Now it's a matter of measuring the angle of the line you generated in Step 3 (example provided here: How can I determine the angle a line found by HoughLines function using OpenCV?)
You can then map the angle of the line to the speed through a simple equation (Will need to see an image of the speedometer to generate this).
let me know how it went.
Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 2 years ago.
Improve this question
I’m trying to make sky detection, I already try on opencv blue color detection, it didn’t work well because of clouds and also because of different color depends on time, the final try was the most useful method, I used a canny detection that need adjusted by user, then I filled the sky with white pixels, and other with black.
The question is,
Is it possible to make sky detection automatically without interface from the user ?!
I can give a recommendation from an AI perspective. Read to the end for a non-AI recommendation.
Actually, there does exist a "simple" way given that you are willing to annotate about a few thousand images manually. You can have some of your coworkers/classmates to help you out with this. I worked with the YOLO-V3 program which does give you a decent GUI to manually annotate your images. YOLO-V3 however only works with bounding boxes, so my next suggestion will work with identifying the whole sky and segmenting the image pixel by pixel.
But if you already have an annotated dataset, There's a neural network architecture called Mask RCNN which overlays your given image with a mask of any color you choose over a given object or setting you to indicate. This one, from my experience, does take a LOT of annotated data to train on for a decent result. But for something as generalizable as a sky-detector and overlay, it should work well with only 1-3k annotated pictures. If you chose to go down this route, here is an article that describes how you can make your own annotated pictures.
Non-AI recommendations:
using the blue color index in the RGB tuple, play around with what thresholds you can use for each color, and you can then do some random sampling of these points and go from there.
But seriously, based on everything I researched on this, looking at other people's repositories, it seems the best method is via AI. Here's an example. The reason is detecting a sky takes a lot of spacial "awareness". For example, how will the computer discriminate between the sky and an ocean? Both are blue. But you can see from the waves that it is an ocean. Basic spatial reasoning can really be done by AI or a crap ton of manual coding.
Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 3 years ago.
Improve this question
I have a couple of standard ways of detecting a modified image such as
Luminance gradient
Copy move detection
Metadata Extraction
Histogram analysis
ELA(Error level analysis)
Quantization matrix analysis
Thumbnail analysis
are there any other standard ways of detecting a modified image?
Tried out
Finding the EXIF of the image to check the created and modified date and check for modification. I also had some rules for EXIF camera make and make note validation along with checking for the software used such as photoshop, Shotwell, etc.
Was able to segment the image and use SLIC(simple linear iterative clustering) to find out the similar cluster regions in an image
Find the largest contour with less pixel inconsistency with luminance gradient to mark that as a potential modified region
Largest contour with ELA as a potential modified region
Check for inconsistencies in histogram graph and mark it as a potential editted image.
Here are my questions
Are there any standard logics to verify the image with metadata such as using the created and modified dates, the camera make or maker note, etc. As these details are not consistent for any given image.
Finding out the least pixel inconsistency contour in the Luminance gradient would always give me an image that is modified?
If the histogram graph has a regular interval fluctuation could it be considered a modified image?
How could I use Quantization matrices to find out image anomalies
what is the best way to compare the thumbnail image to the original image to check for inconsistencies?
The answer to this question needs more detailed so, I will give some references to the subject itself and I will share with you the code of every part of your question :
You need to use exif to verify the image with metadata
For Anomaly Detection in Images see here
To compare the thumbnail image to the original image read this. where it showed you how to compare two images using Python.
References :
ccse.kfupm.edu.sa
github.com/redaelli
github.com/Ghirensics
www.amazon.com/Learning
books.google.com.tw
hal.inria.fr/
Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 8 years ago.
Improve this question
I work with image processing in OpenCV in python.
My main problem is light deflection. Can these deflections be removed with some method?
I implemented a lot of code here, but cant find this particular lights deflections effect. 1)I implemented grayscale, sobel filter, median blur, histogram analysis for plates detections, but this deflections cause that my histogram is bad for edges from sobel filtering, removing these flashes cause that it should works good.
An input image:
Use a colorspace transformation. For instance, if you transform your image to the HSV space, you'll see the "light" components in the V("value") channel:
This is the HSV image:
This is the V channel:
This is the regions of the V channel above a certain level (i.e. a thresholded image):
Now, you can use this kind of stuff to get things done by removing the high values of this V channel, then merging the channels back again. Good luck!
NOTE: as you see, I'm not giving you the code. I think that this should be easy to program if you search the documentation on OpenCV's cvtColor, split/merge or threshold methods ;)