Image-Caption-Generation

InceptionV3-Multi-layer GRU (Keras and TensorFlow)

Requirements:

• Python 3.6
• TensorFlow 1.13.1 (Install tensorflow-1.13.1-cp36-cp36m-win_amd64.whl)
• Keras 2.2.4
• Joblib 1.0.1
• Matplotlib 3.3.4
• Open-cv 4.5.1
• Pandas 1.1.5
• Nltk 3.5
• Download Flickr30k dataset images and captions.

Procedure to Train Model:

1. Clone the Repository to preserve directory structure.
2. For Flickr30k put results_20130124.token and Flickr30k images in flickr30k-images folder.
3. Create pre-trained InceptionV3 TensorFlow graph by running:

python keras2tensorflow.py

4. Generate features(features.npy) corresponding to the images in the dataset folder by running:

python convfeatures.py --data_path Dataset/MSCOCO-images --inception_path ConvNets/InceptionV3_keras.pb

5. Train the model by running:

python main.py --mode train --caption_path ./Dataset/results_20130124.token --feature_path ./Dataset/features.npy --resume

Procedure to Embedded Model:

1. Save both encoder and decoder graphs by running:

python main.py --mode test --image_path ANY_TEST_IMAGE.jpg/png --saveencoder --savedecoder

2. Freeze encoder by running:

python utils/save_graph.py --mode encoder --model_folder model/Encoder/

2. Freeze decoder by running:

python utils/save_graph.py --mode decoder --model_folder model/Decoder/

3. Merge encoder and decoder by running:

python utils/merge_graphs.py --encpb ./model/Trained_Graphs/encoder_frozen_model.pb --decpb ./model/Trained_Graphs/decoder_frozen_model.pb

4. Generate idmap file by running:

python utils/singleProtoBuf.py

Procedure to Optimization (Post-Training Quantization) Model:

Evaluation:

python main.py --mode eval --image_path Dataset/val2017 --inception_path ConvNets/InceptionV3_keras.pb

Results:

• 32-bit floating-point trained model size: 207,167 KB

• 8-bit integers quantized model size: 52,711 KB

• Performance metrics results of proposed Inception v3 + 3-layer GRU language model-based image captioning system on MS COCO dataset:

CIDEr	Bleu-4	Bleu-3	Bleu-2	Bleu-1	ROUGE-L	METEOR	SPICE
0.0462	0.03738	0.0769	0.1818	0.3884	0.2971	0.0961	0.0228

• Python code-generated captions:

• Embedded model-generated captions on the application: