Generating image captions using Xception Network and Beam Search
view notebook on nbviewer:link
Dataset:
The dataset consist of 8,000 images that are each paired with five different captions which provide clear descriptions of the salient entities and events. The images were chosen from six different Flickr groups, and tend not to contain any well-known people or locations, but were manually selected to depict a variety of scenes and situations. The images are divided into train set (6000 images), validation set(1000 images), and test set (1000 images).
You can download the data from here: https://academictorrents.com/details/9dea07ba660a722ae1008c4c8afdd303b6f6e53b or here: https://github.com/jbrownlee/Datasets/releases
Model:
I utilized Encoder-Decoder architecture for the task. The Encoder network is a pre-trained Xception without the last two fully connected layers and it operates as a feature extractor. The Decoder network consists of two layers of GRU units with 256d hidden state. For regularization purposes, I used dropout with a rate of 0.4 between two GRU layers. Extracted features by the Encoder are 2048-d vectors for each image and they are fed to the Decoder alongside the input and also as the hidden state of the first GRU cell in the decoder.
Metric:
I used BLEU metric and it is calculated by comparing n-grams of the candidate with the n-grams of the reference translation and count the number of matches. These matches are position-independent. The more the matches, the better the candidate translation is.
Results:
The model reaches the BLEU accuracy of 61% for uni-grams on the test set, you can improve the score by training the model for longer durations or using a more sophisticated RNN with more layers. If you look at the examples below, you observe that the model is pretty good at recognizing the actions but makes some mistakes at recognizing the colors. Beam search with beams of 1,3, and 5 have been tested. Also, I tried using sum of the log of probabilities in beam search and result improved a little bit for some samples of the test set as shown below. I have included the weight of the trained model, feel free to use the trained model in your own projects.
More on image captioning:
I have recorded a Farsi tutorial explaining this code. You can find it here: http://sariab.ir/Home/Roadmap/8