Semantic description of image based on I-NiC model

In order to address the problems of misprediction and object missing in semantic description of image, an improved Neural Image Caption (I-NIC) model is proposed. It primarily consists of the Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN). The model uses Inception-v4 model developed by Google to extract image features and iteratively optimizes the training process parameters through word-based loss function. Therefore, the I-NIC model can generate more relevant descriptions and improve the accuracy and efficiency of the system. Compared with NIC model, the experiment results show that the accuracy of I-NIC model is improved by 2.5% with BLEU-4 metrics, 1.2% with METEOR metrics and 7.5% with CIDEr metrics on the Microsoft COCO Caption dataset.