the pictures I trained are not very good. I don't know what the problem is.can anyone help me？ thank you！！

[FFjiahao]

the pictures I trained are not very good. I don't know what the problem is.

[FFjiahao]

@costapt @diazandr3s Whether it's related to the data set？Looking forward to your reply.thank you.

[AliSaeed86]

Hi @FFjiahao,
Which dataset have you used please?
i ran the code using DRIVE dataset with 16 images for training and 4 images for validation.

Regards

[FFjiahao]

@AliSaeed86 .I use messidor dataset.How does your code work?Can you send your training results?

[AliSaeed86]

@FFjiahao
Actually i am not sure about the performance of my code. i just wanted to make it run then i thought to change the dataset with larger size one.
using the 16 images for training and 4 images for validation i got this atob.png image as this :

and the result.png as this

it seems that it is memorizing the dataset and generating same image for all, and the vessel tree is not generated.

i guess this issue is because i have used very few number of images for training and validating and not because of the code problem.

Please, how did you get the vessel masks of Messidor dataset?
have you excluded those images of grades 3 and 4 of diabetic retinopathy? and how did you recognize them from grades 0,1 and 2 ?

Thanks

[FFjiahao]

@AliSaeed86 I've had the same results a few times before.This problem is not caused by too few pictures.See if you're using tif for the image suffix.You can use UNET network to segment blood vessels in datasets.i dont exclude the images of grades 3 and 4 of diabetic retinopathy.There are corresponding documents to mark the pictures of retinopathy，you can use the documents to classify pictures.

[FFjiahao]

@AliSaeed86 can you tell me your version of keras,tensorflow and python? thank you.

[AliSaeed86]

@FFjiahao
yes i am using .tif images, do i have to change the extension to another type of images?
As i know, images in Messidor dataset aslo have .tif extension.

i am using keras keras 1.2.2 , tensorflow 1.15 and python 3.7.11
i couldn't modify the code to work on tensorflow 2.x so that i built new envo with same configurations as Costa used.
if you could modify the code to earlier version of tensorflow and keras, please let me know.

Thanks

[AliSaeed86]

if you don't mind, please send me you email on [email protected] for further discussion.
Thanks