Sparse coded histogram----a contour feature beyond hog

This paper is a CVPR2013 article from CMU, which presents a contour feature based on sparse coding, referred to as HSC (histogram of Sparse code) and surpasses hog in target detection (histogram of Gradient) This paper introduces the idea of HSC and its calculation process.

3, the HSC method uses the sparse coding principle to extract the image feature, that is, the image block patch is re-encoded according to the learned dictionary.

The algorithm mainly includes two parts, namely dictionary learning and feature extraction.

Dictionary Learning .

1. similar to dictionary learning based on the K-means method, the sparse-coded dictionary Learning room solves an optimization problem with 1-norm. The steps for dictionary learning are as follows:

2. Initialize. Suppose the length of the dictionary is m, the width of the image small block is the size, where p can go (3,5,7,... ）。

3. image Small Block gets. Collect images about a certain category of objects, and then cut them into squares of small pieces, with a wide height of 1. The image small block according to the image grayscale, the first column of the exhibition into vector form, so the collection of small pieces of the image is recorded as.

4. Dictionary Learning. According to the set, learning to get the best dictionary, the list of dictionaries, representing the length of the column vector, by solving the following equation is obtained:

Where the column vector is expressed in length. This optimization problem can be solved by some existing software tools, such as the K-svd method, by alternating to solve the optimal D and X. The paper specifies that K is 1.

5. Output Dictionary d. As shown in 2.  

 

Figure 1, image small block acquisition

Figure 2, learned dictionary

Feature Extraction

1. After learning the dictionary, you can use a dictionary to extract the features of the specified image. The extraction process is as follows:

2. Initialize, given an image, extract the characteristics of the specified position, 1, extract the characteristics of the dog, the dog into a number of small blocks, referred to as cell.

3. Sparse encoding of the cell block. 1, for each cell, recorded as Y, is sparse encoded. Known as Y and dictionary d, to solve its optimal sparse expression x: Note that x's dimensions and y are unequal, the dimensions of X are M - column vectors, X is a sparse representation of Y, and only a few of the elements of X are non-zero.

4. Smooth Sparse solution x. For each element in X, bilinear interpolation of the neighborhood field.

5. Compute the average cell block of the 16 neighborhood cell block of the X cell block, i.e. the corresponding element is added and divided by 16.

6. Normalized average cell block X. Normalized histogram method is two norm normalization, that is,

7. all normalized average cell blocks are expanded into vectors according to the order of the preceding columns, and the resulting vectors are the corresponding HSC feature histograms.

Figure 3, HSC feature extraction process

The picture is not clear, attached Word version: HSC feature extraction Cvpr2013.zip

Sparse coded histogram----a contour feature beyond hog