In this work, the color segmentation and morphological transformations were used for ROI. The color segmentation identify the white regions on the image, then a binarized image of these regions is then created and a sequence of dilation and erosion are performed, as shows in Figure the ROI are extracted and used in Deep Neural Network (DNN) Input to classify whether is ball image.
To classify the images of ball and no ball, you can use a simple Deep Neural Network. The Deep Neural Network architecture with only 2 convolution layer and 2 fully-connect classify images of ball and no ball with high accuracy (highest than 99%).
| Type | Kernel | Stride | Pad | Filters |
|---|---|---|---|---|
| Convolution 1 | 5x5 | 1 | 0 | 20 |
| Max Pooling 1 | 3x3 | 2 | 0 | 20 |
| Convolution 1 | 5x5 | 1 | 2 | 32 |
| Max Pooling 1 | 3x3 | 2 | 0 | 32 |
| Fully-connect 1 | 128 | |||
| Fully-connect 2 | 2 |
The input image is 32x32x3.
The image can be extract using the imaggetaguer https://imagetagger.bit-bots.de or can be extracted using your program that select a rectangular region of interest (ROI) in images, but the images should be tag manually.
If the images was tag with imaggetaguer, you should be use a script to crop images from rectangular region of interest (ROI), the scripts is available in folder scripts.
This DNN was trainned using NVIDIA DIGITS.
The model used to train is available in folder net.
To testing the net, you can use the NVIDIA DIGITS.
There are a script in ipython to test your net to detect ball in a image, available in folder ball detect.
A simple C++ code is available to classify images using C++ program in folder cplusplus.
These results used the dataset available in Dataset_ball.
| Ball | no Ball | Accuracy per class | |
|---|---|---|---|
| Ball | 998 | 2 | 99.80% |
| no Ball | 0 | 3000 | 100.0% |