Captioning Image Using Convolutional Neural Network (CNN) and Long-Short Term Memory (LSTM)

Internet access in Indonesia brings both negative and positive impact. The data shows that there is a significant raise the number of children who access porn image. The researcher tried to implement the Convolutional Neural Network (CNN) algorithm and Long-Short Term Memory (LSTM) algorithm using learning transfer to predict the caption of a pornographic image. The input space given is a pornographic image that would be extracted using the bottleneck layer of CNN encoder. The weight values in the extraction were obtained from the pre-trained model using Inception V3. Image captioning was processed using one hot encoding technique and produced one hot vector. The results of feature extraction by bottleneck layer and one hot vector are used as training datasets by the LSTM network. The accuracy of the model was calculated using BLEU, METEOR and CIDEr matrices with a score range of 0 to 100. The results of BLEU-1 to BLEU-4 shows the score of 55.84; 24.01; 10.57; 4.59 respectively, METEOR at 12.75 and CIDEr at 43.84.