Deep Learning Project
Image classification is a common problem in both computer vision and machine learning to mapping an image with a suitable label(s) or tag(s). This project used supervised Image Classification via Deep Convolutional Neural Network to classify images.
1 Environment
•Windows 10 operating
•System Intel Core i7-7700HQ CPU
• NVIDIA GeForce GTX1070
• The latest TensorFlow in version 1.8
2 Data Preprocessing
(1) image resizing
Image resizing is to resize all the images into the same size to facilitate following steps.
(2) pixel value rescaling
Pixel value rescaling is to use min-max scaling method to unify all pixel values in an image between 0 and 1.
(3) feature standardization
Feature normalization is to subtract pixel values with the mean of all pixels on each dimension and then divide the results with the standard deviation of all pixels.
3 Training Process
For parameters of the training process, the input images are in a 254-by-254 size with three channels.
The batch size is 32 according to the input size, and the learning rate is 0.0002 to receive a feasible drop of the loss value.
4 RESULTS
With the learning rate of 0.0002 and other parameters mentioned before, the loss value is basically stabilized around 3.2 after six epochs but still fluctuates within a visible small range as the following figure.
