Image Caption Generation Using Deepseek API

Abstract

In the modern era of digital information, multimedia content such as images and videos dominates online communication, social media, e-commerce, and digital libraries. While humans can effortlessly perceive visual scenes and summarize them in natural language, automating this process has remained one of the most challenging tasks at the intersection of Computer Vision (CV) and Natural Language Processing (NLP). This project, titled 'Image Caption Generation Using DeepSeek API', addresses this challenge by designing, training, and implementing an intelligent web-based application capable of generating meaningful textual captions from uploaded images. The project adopts a dual-methodology framework to showcase both custom-trained models and state-of-the-art transfer learning. First, a custom Image Captioning model is developed and trained from scratch using an Encoder-Decoder architecture. A convolutional neural network (specifically, Keras InceptionV3 pre-trained on ImageNet) acts as the feature extractor, capturing high-dimensional spatial representations of input images. These visual features are then projected and combined with text token embeddings to feed a Recurrent Neural Network utilizing Long Short-Term Memory (LSTM) layers, which decodes the representation and generates captions sequentially using a custom tokenizer. Second, to achieve superior, production-ready accuracy, the web application integrates a pre-trained Salesforce BLIP (Bootstrapping Language-Image Pre-training) model. Implemented via Hugging Face Transformers in PyTorch, this model runs inference dynamically on a Flask web server, generating descriptive captions for images uploaded via a clean, user-friendly Bootstrap interface. In addition to raw caption generation, the architecture lays the foundation for a DeepSeek API post-processing layer. Keywords: Image Caption Generation, Deep Learning, Computer Vision, Natural Language Processing (NLP), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Vision-Language Models, Salesforce BLIP, DeepSeek API, Transfer Learning, Transformer Architecture, Flask Web Application, Image Feature Extraction, Hugging Face Transformers, Artificial Intelligence (AI).