This paper will soon look over, need prior knowledge to go over tomorrow, try to figure out the framework of this paper, as well as the relevant principles also understand its inherent, long way to go.
Today more Bo: All remaining part
Today's content is not much, since most have already been involved, and to avoid a blog about the high repetition rate cutting.
Next week will start in English, and prepare yourself a good weekend ....
(1) images and solve the augmented image field deviation : the role of our DAY1 in GAN, the expansion of data sets, reducing the gap between picture . Now and then we can use that as a gap so that generated images and determine the picture of the Euclidean distance size. The deviation prediction picture we can also use contrastive loss, more precise it is the Triplet and further quad loss . In fact, for both bias method, we can find out if a tuple in the past few days of learning, then the choice of loss function and its extensions can be resolved (of course I still cross_entropy loyal fans, because other loss function I not trained, escape); for triplets loss , that is, to find a positive sample and a negative sample , the generative close to the positive samples, increasing the (open) from negative samples, this is a clustering way . For quad loss , is the use of two negative sample and a positive sample , the purpose is to avoid the negative samples and generative pushed the distance, but instead affect the relationship between distance and generative of positive samples, it can be noted quad losses in a weak pushIn effect, is the distance between two negative samples , so it avoids the influence of the image features. Here we are using a generative as an example, but the use of contrastive loss in input any type of image is not a problem.
(2) reordering : reordering means a compiler and a processor to optimize the performance of the program and the sequence of instructions for reordering a means. The instruction sequence refers to a plurality of (or multiple giant) instruction machine code bytes as a basic unit of a linear sequence joined to form a queue. So the machine code , what is it? The machine instruction code is a set of binary control computer, and operation should be done operand address, easy understanding is the code meaning it.
(3)HR LR :
From left to right the true value HR image and a corresponding image LR trained to minimize the MSE (mean square error function) (i.e., all the pixels for the summation of squared difference between the predicted value and the true value is divided by and the number of pixels corresponding to the predicted loss of function) model.
(4) the sparse representation factor : here comes to understand mathematical concepts and mathematical formulas sparse representation, not able to directly understand, but this super-detailed speaker of Bo, so that this concept becomes more clear and understandable.
Link below: https://blog.csdn.net/Forever_pupils/article/details/88572281
(5) an idea: use an infrared camera to obtain outline in the dark, you can use an affine transformation to distinguish pedestrians do? Whether this is feasible depends on whether the human form is not the same, and block the clothes will have been affected (the good solution, using the temperature difference can also be extracted to the human body infrared map), at the same time, people of different times morphological transformation (fitness, weight loss) can also cause changes in morphology. emmm, I knocked pounding, denied that his idea of this .... When I talked about cutting blind