Hybrid Relative Attributes Based on Sparse Coding for Zero-Shot Image Classification

Algorithm 1

Nonsemantic relative attribute learning based on sparse coding.

Input: the low level features of the training samples and the testing samples, the weight attenuation coefficient .

Output: the non-semantic relative attributes of the training samples and the testing samples.

Step 1: bring the low level features of the training samples into Equation (1), and initialize the basis vector ;

Repeat;

Fix the basis vector in the last step, update the activation value to minimize Equation (1).

Fix the activation value in the last step, update the basis vector to minimize Equation (1).

Repeat the above procedure till convergence is reached, obtain the basis vector to represent the sample

features well.

Step 2: bring the basis vector and the low level feature of the testing samples into Equation (2).

Step 3: adjust the activation value to minimize Equation (2), the adjusted activation value is the non-semantic relative

attributes of the testing sample.