Multilabel Image Annotation Based on Double-Layer PLSA Model

Algorithm 1

The image concept detection algorithm.

(1) Input: a new untagged image 𝐼

(2) Get the BoW representation of 𝐼, a 𝑁_𝑣 dimensions vector:

   𝑉(𝐼)= 𝑛𝐼,𝑣₁,…,𝑛𝐼,𝑣_𝑖,…,𝑛𝐼,𝑣_𝑁𝑣

(3) According to the training parameter 𝑃(𝑣/𝑧^𝑣), applying the folding-in algorithm to get

  the visual latent topic distribution:

   𝑍^𝑣=𝑛𝐼,𝑣₁,…,𝑛𝐼,𝑣_𝑖,…,𝑛𝐼,𝑣_𝑁𝑣

(4) The same as the step (3), according to 𝑃(𝑧^𝑣/𝑧^top), using the folding-in algorithm to get

  the top-layer latent topic distribution of image 𝐼:

   𝑍^top=𝑃_𝑧^top₁,…,𝑃_𝑧^top_𝑖,…,𝑃_𝑧^top_𝐾

(5) Combine the training parameter 𝑃(𝑧^𝑡/𝑧^top) and the top-layer latent topic distribution

  gained from the step (4) to get the label latent topic distribution of image 𝐼:

   𝑍^𝑡=𝑃_𝑧^𝑡₁,…,𝑃_𝑧^𝑡_𝑖,…,𝑃_𝑧^𝑡_𝐾𝑡

(6) According the training parameter 𝑃(𝑡/𝑧^𝑡) and the distribution of 𝑧^𝑡, the probability of

  each image semantic label appearing on the image 𝐼 can be calculated:

   𝑇(𝐼)=𝑃_𝑡1,…,𝑃_𝑡𝑖,…,𝑃_𝑡𝑁𝑡

  then choose 𝑛 (𝑛 can be selected as need, in this paper 𝑛=5) labels with the largest

  probabilities to construct the label set of image 𝐼

(7) Output: the 𝑛 labels of image 𝐼