Image Resizing: Adaptive resizing is implemented in this project. The Scale class will resize the image to match its shortest size and rescale the longest side with a particular ratio.
Image Cropping: RandomSizedCrop will crop a random area of the original size with a random aspect ration. This crop is finally resized to a fixed given size. However, this class is an advance version of image cropping. We will crop the patch in the center with a square sized out put after all the attempts trials fail.
Color Jitters: A small pertubation in the color space can leas to imaged with drastically different pixel values, yet are still perceptually meaningful. RandomColor class will implement color jitters. First, we will sample the alpha from a uniform distribution U(1-r, 1+r) and then alpha is mulitpled to the corresponding color channel. Improvment is this class: Create a lookup table, so we can look up each channel. We make sure all the pixel value is between the range 0 - 255 uing np.unit8
Random Rotate + Cropping: We can simply rotate an image uing the function. However, this function will create the black pixels in the result image. The ideal situation is that we want to avoid these black pixels.RandomRotate class is able to crop the rotated image without black pixel by applying some math - sin, cos. The goal is to get the max area rectangular within the rotated image. We will apply the math to do this implementation.
reference: stackoverflow.com/questions/16702966/rotate-image-and-crop-out-black-borders
