I need to build a tool for detecting where in an image certain objects are located. The objects are quite irregular – they are fruit flies – and need to be detected at close to pixel perfect boundaries, especially for the case when two flies are right next to eachother.
I have tried doing a simple convolutional neural network based approach training a softmax classifier and then running this network over every pixel in the image to obtain a heatmap.
However, I'm curious if a less complex approach would also work.
I would take all the ground truth images (x,y, r,g,b) and then find the expected eigenvectors. Then, I would try to find the best superpositions of the eigenvectors onto a given input image. In other words, convolve a filter over the image that extracts eigenvectors and then take the distance of these eigenvectors to the expected eigenvectors which represent the fly in lower dimensions. How would I analyze this idea?
When you start with transformation into Eigenspace then you need to have of correlating it to the raw data,say in terms of roi (I considered smaller rectangular areas , max in the region (because it was normalized) and indices ).
Once the eigenvalues are obtained again you need to select and them for interpretation , they are so many how would you pick the ones that are important (statistically).I used the Euclidean distance metric.
So , yes in the transformed space the 1-to-1 correlation does not exist although uniqueness (independent set of values) and lesser dimensions are achieved.
When you start with transformation into Eigenspace then you need to have of correlating it to the raw data,say in terms of roi (I considered smaller rectangular areas , max in the region (because it was normalized) and indices ).
Once the eigenvalues are obtained again you need to select and them for interpretation , they are so many how would you pick the ones that are important (statistically).I used the Euclidean distance metric.
So , yes in the transformed space the 1-to-1 correlation does not exist although uniqueness (independent set of values) and lesser dimensions are achieved.
Eigenfaces and fisherfaces are one of the earliest popular algorithms for detecting faces in images.
Your approach will work, but you need to test the number of eigenvectors needed to do a good job at feature discrimination between fly-non fly ROIs.
I would recommend looking at more general ideas (such as LLE, Laplacian Eigenmaps) because usually, images of an object span a lower dimensional structured manifold in high dim. space.
E.g. see this matlab toolbox (start with the readme.txt):
https://lvdmaaten.github.io/drtoolbox/
Why not use SIFT or SURF? They are quite robust descriptors.
For openCV both SURF and SIFT:
http://docs.opencv.org/2.4/modules/nonfree/doc/feature_detection.html
For MATLAB SURF and other descriptors too:
https://www.mathworks.com/help/vision/ref/matchfeatures.html
https://www.mathworks.com/help/vision/feature-detection-and-extraction.html
http://docs.opencv.org/2.4/modules/nonfree/doc/feature_detection.html
I think the problem solved in the following paper is similar to your problem.
Zhenzhou Wang, A New Approach for Segmentation and Quantification of Cells or Nanoparticles,Jun 2016 · IEEE Transactions on Industrial Informatics
An image contains one or more than one objects. Contour of the object not necessary in rectangular, then how we evaluate eigen-vectors? which are helpful for object extraction.
I think it is suitable, and it is also used when combining PCA with image processing
Could you post 2 or 3 examples of the images you are working with and what the expected result would be (by manually segmenting them). That might triggers other ideas using the specificity of your problem!
Bruno
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing