Skip to content

Latest commit

 

History

History
31 lines (25 loc) · 2.14 KB

README.md

File metadata and controls

31 lines (25 loc) · 2.14 KB
Raw

CZ3004_image_recognition

image recognition for MDP. It uses Keras/Tensorflow

Preparing the Dataset:

  1. Using the arduino camera on the robot we took pictures of the maze.
  2. We took images of both symbols & non-symbols inside the maze.
  3. Next, we crop the images to only include locations where the image is expected to appear. Which is 3 locations. Therefore, we have 3 cropped images for every image we have taken. In total we had about 60 croppped images
  4. To improve the image recognition, we distorted the quality of these 60 images even further. We added blur effects, slightly tilted the image, reduced colors to black and white, etc.
  5. Finally, we trained the model on this dataset.

1st round of training:

  1. For the first layer of the model, we taught the model to different between symbols & non-symbols images.
  2. It outputs a yes if a symbol is detected and a no if not detected.

2nd round of training:

  1. We froze the previously trained layer to ensure it remembers what it learned. Which is how to detect symbols or non-symbols images
  2. Next, we added another layer for the model to train.
  3. If a symbol is detected, we taught the model to differentiate the symbol between 15 possible symbols.
  4. The model tells us which of the symbol that the image is likely to be.

Implementation:

  1. On startup, the robot loads the model into memory. This is to remove the need to load from hard drive when the robot is exploring the maze.
  2. The robot takes an image whenever it detects a block in the maze.
  3. The image is sent to our trained model (that is already loaded in memory). The model identifies the symbols if it is present.
  4. Finally, the robot designates the block as having a symbol and this is reflected in the android screen.

Final Results: The robot can detect images 100% accurately in less than a second. We were confident of getting 100% for image recognition. However, the android app declared that the robot had finished exploring before getting the last image as the last image was beside the finish line. We got 4th place for image recognition.

Final Trained Model: JiaChin_Martyn_MDP22_19S1_.h5