THIS IS AN EDUCATIONAL COMMUNITY BASED SUPPORT FORUM FOR HARM MINIMIZATION ONLY.
NO SOURCING OR DEALING WILL BE TOLERATED.
TO USE THIS WEBSITE YOU MUST READ AND AGREE TO THESE RULES HERE.
MEMBERS THAT DO NOT FOLLOW THE RULES WILL BE AGGRESSIVELY BANNED.
Results 1 to 2 of 2
  1. #1
    Join Date
    Jul 2017
    Posts
    9
    Thanks
    0
    Thanked 1 Time in 1 Post

    Default Training Google Inception for ID - opinions?

    Mod - please let SWIM know if this is OT here and move elsewhere, but SWIM likes that this forum is member only...

    So, SWIM has been looking into the Google Inception image recognition system and playing around with training it to classify pictures of Psy vs. Non Psy mushrooms. Think about the "Hot Dog vs. Not Hot Dog app from the Silicon Valley show . (Also check this article). Here is the process SWIM has followed and anyone can pretty much do the same. At this point, SWIM has trained it on a small sample of about 450 images of psy and the same number of non-psy mushrooms. The images were mostly downloaded from google with scrutiny on SWIM's behalf to make sure as much as possible that the training images are exactly what they should be classified as. Then SWIM tested with some of SWIM's own images as well as downloaded ones. SWIM gets reasonable results but there are a higher than desired number of false positives just due to the fact that a lot of non-psy mushrooms look very similar to the psy ones and SWIM's training set was small.

    Once the classifier has been trained, it can be turned into a mobile app so that you can simply point the camera to a specimen and get some probabilities back on whether the classifier thinks it's a psy or non-psy. Note, SWIM hasn't made the app yet - this would be a next step to make testing easier and to see how this actually performs in the wild.

    Therefore, SWIM thinks to make this work better, SWIM needs more images for both camps and especially good quality images so the classifier can get some good feature detail, as well as some way of testing this with a small group of enthusiasts, but before SWIM continues working on it, SWIM would like to get some opinions about this approach.

    Sounds super easy but...

    It's about being responsible. SWIM totally understands that a blind reliance on this sort of classification system can be very dangerous and cannot replace expert opinion and experience, however, it can still be very useful to cut down on the level of uncertainty with obviously non-psy examples. SWIM themselves would definitely not trust this classifier - especially at the current stage and especially if one doesn't have any or much experience with ID-ing. However, in theory, the system can be continuously improved by adding more and more images and retraining the system periodically. This is pretty much what this project does for classifying any kind of plant.

    Currently, the aim of this is to experiment with the image classifier and to validate that a reasonably reliable ID can be achieved this way in this particular space. The aim is certainly not to be irresponsible and claim surefire ID - after all, lives are at stake!

    To make this better, a lot more training data is needed and the more specific the training the more accurate the classifier will be. Ideally the training set should have several thousands of images in both classes so there is a lot of image collecting to be done. Also, if the system should be geo-specific then the training set should be mostly local species found in NZ so the classifier can be more confident.

    So, SWIM is looking forward to some opinions. SWIM is happy to run some examples for you with the currently trained system if you supply some images. The best images would be in-situ before picking and at a close range so the fruit is the most prominent feature. The output is going to be:
    psychoactive (score = x)
    not psychoactive (score = y)

    A big challenge are the galerinas as they look so damn alike and there is not much one can do about this except look for the subtle details like gill colour/shape, stem colour and spore print. It would be interesting to find out if the inception algorithm can be sufficiently trained that it can distinguish this particular species. It may depend on how the image is taken or from what angle the camera is looking at the subject... again it's all in the feature detail and whether the algorithm can be trained to recognise these subtle differences.

    What do you think?

    And now for some examples!

    1. Correct classification:
    https://i.imgur.com/Ue3sAU0.jpg
    psychoactive (score = 0.99254)
    not psychoactive (score = 0.00746)

    2. Incorrect Classification - false positive but with a significantly uncertain score so you would question this further:
    https://i.imgur.com/X6MY0so.jpg
    psychoactive (score = 0.76119)
    not psychoactive (score = 0.23881)

    3. Correct Classification:
    https://i.imgur.com/q6dckZb.jpg
    not psychoactive (score = 0.98648)
    psychoactive (score = 0.01352)
    Last edited by mycophagus; 05-07-2017 at 12:56 PM.

  2. #2
    Join Date
    Jul 2017
    Posts
    9
    Thanks
    0
    Thanked 1 Time in 1 Post

    Default

    Weed in NZ
    OK, so swim just saw this app: https://itunes.apple.com/us/app/mush...n/id1135020853 that uses the same or very similar training technique as described above. It's no surprise that the reviews aren't great, but that's what swim said also - they would need a huge dataset - several K images for EACH species to make this remotely viable.

    The same crowd also has a twitter mushroom ID bot: https://twitter.com/mushroomaibot

    Anyways, swim has now made that android app for fun and still keen to get ahold of more real life images to improve the training dataset.

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •