Abstract: A machine learning method includes: distinguishing foregrounds and backgrounds of a first image to generate a first mask image; cropping the first image to generate second and third images; cropping the first mask image to generate second and third mask images, wherein a position of the second mask image and a position of the third mask image correspond to a position of the second image and a position of the third image, respectively; generating a first feature vector group of the second image and a second feature vector group of the third image by a model; generating a first matrix according to the first and second feature vector groups; generating a second matrix according to the second and third mask images; generating a function according to the first and second matrices; and adjusting the model according to the function.

Abstract: A data augmentation device including a processor is disclosed. The processor is configured to capture at least one sample image from an original image, and the processor is configured to input the at least one sample image to at least two data augmentation module, so as to generate at least two augmentation image group through the at least two data augmentation module. The at least two data augmentation module include a first data augmentation module and a second data augmentation module, in which a first parameter group of the first data augmentation module is a fixed value, and a second parameter group of the second data augmentation module is not the fixed value.

Abstract: A machine learning method comprises: (a) applying a contrastive learning model to a training image and an image mask to generate a foreground feature vector pair and a background feature vector pair; (b) calculating a foreground loss and a background loss according to the foreground feature vector pair and the background feature vector pair; (c) calculating a total loss from the foreground loss and the background loss; (e) when a recursion end condition is met, using a first encoder for parameter adjustment of machine learning model; and (f) when the recursive end condition is not met, adjusting a parameter of the first encoder in the contrastive learning model using the total loss, and adjust a parameter of a second encoder in the contrastive learning model using the adjusted parameter of the first encoder and a preset multiple, thereby performing step (a) to step (d) again.