Second lesson-data-driven APPROACH:KNN and linear classifier category pictures

2017-08-12

1. Image classification is the basis of many CV tasks;

2. Picture classification to face a lot of problems, than the film is obscured, the same animal has many kinds of colors, shapes and so on, the algorithm needs to be strong enough;

3. It is very difficult to write the program directly for image classification, the common method is the data driven method:

4.KNN: The focus is to select the value of K, you can take the cross-validation method to find the best K value;

At the same time, the distance representation also has a concentration method, such as Euclidean distance, Manhattan distance:

5. Then from KNN, the general function model is introduced, that is, to map a picture into several categories of possible numerical scores, the highest is the category of the picture belongs to:

Note: The x here is a 3072-dimensional vector, function f is to map 3072 dimensions to 10-dimensional functions, 10 represents the last possible category of 10, of course, sometimes need to consider the offset b,bias;

Example:

6. Next time focus on the loss function loss functions, as well as the optimization process, that is, to find the minimum parameter value of the loss function W, finally extended to other classifiers, neural networks, convolutional neural networks;

Second lesson-data-driven APPROACH:KNN and linear classifier category pictures