Thesis Step (Implementation of Deep Learning Using Convolutional Neural Network For Stationary)

This is my thesis Step, about Implementation of Deep Learning Using Convolutional Neural Network For Stationary Classification(Study Case: Ballpoint, Eraser and Ruler Images). The benefits of this research are to classify the stationery in store, to motor development of children and promote writing for children. We use Convolutional Neural Networks with Keras in Rstudio for create training model. We create Dataset from Crawling google image  with python and javascript program.

1. First step we crawling images in google image search, crawling image can use python and javascript program. Why crawling?? If you want to download manual images is ok, but I think it’s make me bored. Time is very important so we must very fast to download images and we use crawling image technique.
2. The next step is create dataset, with choosing image with 1 object and background white. The quality of images is very important  for accuracy model. Why 1 object and background white?? At the training model, images will be easy to recognize because there is no other object in the picture.
3. In preprocessing we have 5 step :
   a. Create train and test data
   From the dataset we divided into train and test data with comparison 80:20. Train data used to create the model and test for evaluate this model.
   b. Resize Image
   Image size for input shape should form a symmetrical size (100 x 100, 64x64, 32 x 32). Researchers are using a resolution of 32 x 32 as some consideration i.e. If you use 64x64 then it will make the process more weight due to the limited RAM and GPU used computer then when using 16 x 16 resolution then the information from the image will be too small and will have an effect on the accuracy of the model
   c. Combine Image
   At the stage of training that will read it first on the structure of the data is the number of images used, so we need to combine all the images into one and change the structure of the dimension
   d. Reorder Dimension
   On the training model, the dimension will be read in advance are the total of images, pixels, and format RGB. So we need to exchange the position  of the dimensions in train and test data.
   e. Create Label  
   After reorder dimension we create labels from the train and test data used for labeling image.
4. Training Model

In this model we use 4 convolutional layer and 2 pooling with relu activation in feature extraction layer. We can change this model about filters, kernel size , activation, learning rate etc. If you want to see some result of running in this model you can open my GitHub, but In my github you can use tensorboard for to see  it. I use learning rate manipulation for create the model good.

5. We have the best accuracy in data train is 100% and test 95%, with loss 0.54 in test data.

We can save model and use model for new images, but we will do until preprocessing step for new images. After that we can predict new images with CNN model that we make. This is output predict test images and we have 3 object wrong classification. Note (0 = ballpoint, 1 = eraser, 2 = ruler)

I’m sorry for my bad English language and Let's learn together.