Cell/color detection

Hello,

I’m working on biological cell detection in images. There are many cells in the image and some cells are stained with different color (e.g. brown, red, etc). I created a training data set containing one cell type (>150; distinct dark brown). I trained the model based on the “halolens” example; except the lowest loss was large in my opinion (>15). My detection failed to detect any of those cells, while some detected “objects” was light messy background.

Any advice what i did wrong? Do I have to include background or other objects?

Thank you,
Art

PS. I am new, but I managed successfully to run ImageAI example/tutorials.

2 Likes

I too have same problem on different dataset

How many images do you have in your training dataset? You might need to collect 2X, 3X or more pictures to ensure the model have enough examples to train.

Hi,

So, I used 5 large images that have many cells inside (thousands). I have annotated hundreds of cells inside of them: all cells of interest (20-40 per image; brown) and some background cells (50-100 per image; light bluish). I have inserted an example that is small cropped area to zoom in to show details. I tried to increase the number of cells but no success. I wander what i’m doing wrong…

.

You have done a lot of work annotating the images. However, considering the number of images you have is few (5), you will need more number images to train an effective model.

ImageAI uses deep learning models which require large numbers of training instances. In the case of detection, you will most likely need many images as opposed to many annotations with few images.