Status Report

I talked with my mentor Tony Fountain today, along with Dr. Wang and Dr. Lu to discuss the progress on my project, the major issues, and the direction of the project.

Some key points we touched upon:

• Documentation: Things that I need to record/discuss in greater detail in my event report
• Experiments and methodology
• This includes talking about what type of validation I used (hold-out validation).
• How many samples per class, and testing/training sizes for each class
• Try to explain the results of the classifier
• Issues with project resources, and not necessarily just technical issues
• Photos of wasp wings are inconsistent from image to image, which is a major issue for not only the classifier, but pulling out new features
• This raises a difficult dilemma: To either segment the wings out, or to obtain a better database of images.
• In the former, segmenting the wings out is difficult as we need to identify precisely what the general characteristics of a wing since they will vary greatly between images (folded/unfolded, have different orientations, colors, patterns, etc.) This is a very difficult problem in itself.
• In the latter, the specimens are hand-prepared very delicately, and to get consistent wingspan in each image would require many hours of manual labor for the hundreds of samples needed for machine learning algorithms.
• It appears that I don't have enough time or experience to both process the images and build the classifier. On the other hand, there are also not enough resources to manually reorient all the wasps and take new photographs. As a result, there may be some re-scoping of the project.
• Re-scoping the project (possible ideas)
• Bee counting problem
• In previous years, PRIME students implemented (Michael Nekrasov) and improved (Michael Perry) a bee counting program.
• The program used a blob detection algorithm and did not differentiate between bees and wasps.
• Instead of pure blob detection, we could try to extract features from the insects and use those as a means for bee recognition.
• General classification of field images (as opposed to species-level of specimen images)
• Since the current images are not carefully processed for image analysis, it will be harder to distinguish different species with similar features
• Instead, we can try to work on the problem of differentiating insects that have very different features, but are from unprocessed field image
• Goals for this week
• Replace nearest neighbor classifier with Decision Tree classifier, and study the structure.
• Experiment with cross-validation
• Continue to experiment with higher level feature extraction