what algorithm to use for object detection in deep learning

I am using matlab and trying to measure the inner and outer radii of a donut in noisy background. I have thousands of images 512x512x1 (gray scale). I have been trying to train rcnn an fastrcnn on these images per the stop sign example. The classifier part works very well but the learning transfer, despite high accuracy, fails to locate and size the donuts properly. I retrainging using rcnn and fastrcnn. I often get warnings of not finding negative sample regions but even if I don't detection fails.

Should I be using a different algorithm? i.e. YOLO, SSD. is the issue the filter sizes and number of filters?

I would have thought this would be an easy task for deep learning. I train with 2000 images and retrain with 150.

My model is

1 '' Convolution 32 5x5 convolutions with stride [1 1] and padding [2 2 2 2]

2 '' ReLU ReLU

3 '' Max Pooling 3x3 max pooling with stride [2 2] and padding [0 0 0 0]

4 '' Convolution 32 5x5 convolutions with stride [1 1] and padding [2 2 2 2]

5 '' ReLU ReLU

6 '' Max Pooling 3x3 max pooling with stride [2 2] and padding [0 0 0 0]

7 '' Convolution 64 5x5 convolutions with stride [1 1] and padding [2 2 2 2]

8 '' ReLU ReLU

9 '' Max Pooling 3x3 max pooling with stride [2 2] and padding [0 0 0 0]

Final layer

1 '' Fully Connected 64 fully connected layer

2 '' ReLU ReLU

3 '' Fully Connected 10 fully connected layer

4 '' Softmax softmax

5 '' Classification Output crossentropyex

Answer 1

[Vivek Akkala]

Hi,

You can use YOLO to locate the donuts. To understand how to train object detectors with a custom dataset, you can refer to the "Object Detection Using YOLO v2 Deep Learning" example.

Object detection algorithms like YOLO typically do not provide the size of the detected objects directly. To obtain the size of the detected donuts, you may need to perform additional post-processing steps. One approach could be using the "regionprops" function, which can help measure properties of detected objects. You may also refer to the https://in.mathworks.com/help/images/ref/regionprops.html#buoogbr-1, example that demonstrates how to compute the center and radii of circular objects, which could be useful for your case.