Inspired from the paper "Show Attend and Tell". This project's aim was to train a neural network which can provide descriptive text for a given image.
Architecture basically consists of encoder-decoder Neural Network, where encoder extracts feature from the image and decoder interprets those features to produce sentence.
Encoder consists of a pre-trained Inception-V3 model (minus the last Fully Connected layers) appended by a custom Fully Connected layer. Decoder consists of LSTM along with visual-attention. Visual attention helps LSTM in focusing on relevant image features for the prediction of a particular word in word sequence (sentence)
For word embedding, a custom word2vec model was trained and it was then intersected with Google's pre-trained word2vec model.
Google-colab and google drive combo was used for training the model. Flickr30 Dataset was used for creating train and test sets.
A number of models were trained (each trained on different no of images). Of these, version 5 model performs the best. You can access these models from "models" sub-directory inside "ImageCap" directory
InceptionV3, LSTM, visual-attention, Word2Vec
Tensorflow, Keras, Gensim, Numpy, Pandas, Sklearn, Pickle, Pillow, Matplotlib, ast, os
In order to use these pretrained models, clone/download this repository. Make sure that you have Tensorflow (>=1.15), Numpy, Matplotlib and Pillow installed on your system. Your system should also have Python 3. Now, unzip the repository and execute the "run.py" file in it with arguments <image_path> (1st argument) and <model_version> (2nd argument). Use model_version==5 to get best results.
In order to train your own model, clone/download this repository. Install all libraries mentioned above. Download content mentioned in /ImageCap/misc/Download links.txt. (1) Extract images using Extract.ipynb. (2) Run Image_Features.ipynb to obtain features of the images (3) Run Text_processing_and_embedding.ipynb to perform text preprocessing and word embedding. (4) Run Training.ipynb to train the model. (5) Run Evaluate.ipynb to get results on test images. (Note that, you will need to create some additional empty sub-directories as per paths in code to contain things like Image features, Images etc.)
Following are some of the results. You can look at some more results in the Evaluate.ipynb
