In the first step, we obtain the image representations with the help of a CNN
which was supervised pre-trained on the ImageNet 2012 dataset and fine-tuned
on target dataset. The CNN model contains five convolutional layers, two fully
connected layers, and a softmax classifier (a classifier for feature extraction
andparameter separation). This model incorporates a huge amount of semantic
information, since it was trainedon ImageNet 2012 classification dataset which
consists ofmore than 1 million images and it is able to classify into 1.000
categories. This huge amount of semantic information and the large number of
categories makes suitable this model with slight modifications to our task. Note
that the input of this pre-trained CNN is a fixed-size, mean-subtracted 224 _ 224
RGB image.
Unlike, the semantic feature is derived from the last fully-connected layer
directly, instead of inserting a new hash layer. The output of the last fully-
connected layer is splitted into two ways. One part eventuates in an n-ways
softmax classifier where n stands for the number of categories of the target
dataset. The other part makes up a hash-like function which composes the
features obtained by CNN to hash codes. The so-called mid-level features are
obtained from the last fully-connected layer, and softmax classifiers are trained
for each semantic label concurrently. We interpret the output of the softmax
