I need to classify image to a certain category. I did extract features using SURF then clustered using k-means in Matlab then transfer the data in csv files for classification in Weka. I managed to get 97% for the train data when classified using SMO and LIBSVM. However, when I test to my unlabelled data, Weka give an error of un-compatible data. I need to test for new images without labeling them. Can you guide me on this issues? Thanks in advanced.
I think SURF is for 3-D colored surface plot. how did you clustered your data? please clear this doubt.
Hi Siti,
Your train and test sets have to be compatible--attributes need to have same name, order, and type.
HTH.
Samer
Salam Samer,
I did check all the name, order and type and there are all same because I simply copy the train name and paste it to the test name. It just that I did not put the class label (category)for each data in the test datasets. I just put '?' this instead.
Mr Sanjay Gaur,
1. I convert coloured image into gray scale
2. Using this code: detectSURFFeatures(imagesGray);, I extract the key interest point
3.clustered the extraction point using kmeans
Here is the link which I referred:
https://www.mathworks.com/help/vision/ref/detectsurffeatures.html
Salam Siti,
I'm not sure that your process is correct.
The header is important, so kindly also make sure that the order of the class attribute labels (last attribute/column) is identical in the training and the test sets. Moreover, for the test set data, make sure that you've replaced the missing labels with "?" in the class attribute (last coloumn), while your training set needs to have the labels.
However, if you still have same problem, please send me your data to fix.
HTH.
Samer
Salam Samer,
Thank you so much for your guidance. I've managed to get the result but unfortunately the result showed a lower accuracy (66%). Need to figure out for the next problem by the way thank you again...
Hi Aparna,
Does fcm means fuzzy c-means? Can you share in what way it is better than kmean? I just used kmeans because it is widely used as bag of features model for image classification. Thank you for your feedback, I will try to search more input on that since my accuracy is so low.
Salam Siti,
Happy to hear that. However, to enhance your accuracy, you could normalize your data and/or remove the outliers (or perform other preprocessing techniques). You could also apply attribute selection approach. Another option you can consider is adding additional instances to your data.
When building the original predictive model that gets evaluated using cross-validation (if selected as an evaluation method), you may look at another evaluation metrics (in addition to the accuracy), such as ROC area, confusion matrix, etc. Once getting a satisfactory performance, you can apply your model to classify the new unlabeled instances.
Finally, you can also rely on K-means algorithm to perform feature extraction.
HTH.
Samer
Hi Siti,
You are in a right direction. That time, I was not sure about the matlab function detectSurf features... This function is newly added in matlab 2017.
No doubt FCM works better than K mean. You can use this clustering approach to retrieval system to get better results.
Siti Suhaila
Hi Siti, in the end, what did you do to solve that problem? I am having the exact same problem as you and dont know what to do. Thanks!
- Badges
- Science topic
- Similar topics
- Filing