DenseNet201, LSTM Image Captioning: An Attention Gap Analysis
Year: 2026
Type: other
Abstract: This paper documents the design, implementation, and evaluation of an image captioning system built on a DenseNet201 convolutional encoder and an LSTM-based language decoder, trained on the Flickr8k dataset. Image captioning sits at the intersection of computer vision and natural language processing, requiring joint reasoning over visual content and language generation. While recent literature has moved decisively toward attention-based and transformer decoders, a substantial share of published implementations that use Flickr8k continue to rely on a nonattention CNN-LSTM merge architecture, either as a baseline or as a complete system. This paper situates the present implementation within that literature, reviews closely related work using the same encoder-decoder pattern, and identifies the specific architectural property responsible for its observed limitations: the fixed-length visual bottleneck between encoder and decoder. Training curves show a widening gap between training and validation loss consistent with overfitting in a small-data, non-attention regime. Qualitative caption outputs are reviewed in place of quantitative metrics, which were not computed in this implementation and are noted explicitly as a limitation. The paper closes with a literature-grounded roadmap for extending the system: attention mechanisms, larger training corpora, transformer decoders, and standard n-gram evaluation metrics.Keyword: image-captioning deep-learning computer-vision nlp densenet201 lstm encoder-decoder flickr8k tensorflow keras cnn-lstm attention-mechanism transfer-learning Source Code:GitHub: https://github.com/sohanbiswas-ds/Brain-Tumor-detection-using-transfer-learning-VGG-16- Author: Md. Sohan BiswasORCID: 0009-0000-5888-373X
Author: Biswas, Md. Sohan
Institutions: Eastern University, Eastern University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7202261149