Utilizing Deep Learning Technique for Arabic Image Captioning

The Arabic image captioning process offers a valuable avenue for delving into the components within images and the intricate interconnections they exhibit. While this subject holds significant importance in the English language realm, it faces unique challenges in the Arabic context due to the scarcity of readily available databases. In contrast, English benefits from an abundance of resources, whereas Arabic often resorts to Google translation as an imperfect alternative, leading to the introduction of errors. To mitigate this issue, a crucial step involves pre-processing the textual data. In this paper, a novel model based on encoder-decoder techniques was presented; the proposed approach involves conducting a series of experiments, categorized into two models - VGG19 and Inception-ResNet-v2. These models play a pivotal role in feature extraction from the image during the encryption phase. Additionally, a cutting-edge model named BILSTM is introduced, which capitalizes on processing word sequences to predict text; this model has demonstrated superior performance compared to LSTM and GRU models in the decoding stage. The findings of this study, as measured by the Bleu performance scale, suggested notable improvements, with scores ranging from 33 to 37.