Intellectual Property India Publishes Patent Application for System And Method For Automatic Hindi Image Caption Generation Using A Four Model Ensemble Feature Extractor, A Stacked Lstm Decoder, And Beam Search

MUMBAI, India, July 24 -- Intellectual Property India has published a patent application (202611065325 A) filed by Vishal Jayaswal on May 23, 2026, for System And Method For Automatic Hindi Image Caption Generation Using A Four Model Ensemble Feature Extractor, A Stacked Lstm Decoder, And Beam Search.

Inventors include Vishal Jayaswal; Rajneesh Rani; and Jagdeep Kaur.

The application for the patent was published on July 17, 2026, under issue no. 29/2026.

Abstract: System and Method for Automatic Hindi Image Caption Generation Using a Four Model Ensemble Feature Extractor, a Stacked LSTM Decoder, and Beam Search A system and method for generating Hindi captions from digital images. The system comprises a four-model ensemble feature extractor that processes an input image in parallel through four pre-trained Convolutional Neural Networks: InceptionV3, DenseNet121, EfficientNetB0, and ResNet101. Each network applies global average pooling to its final feature map, producing vectors of dimensions 2048, 1024, 1280, and 2048 respectively, which are concatenated into a 6400-dimensional combined feature vector. A stacked Long Short-Term Memory (LSTM) decoder, comprising a first LSTM layer with 512 units returning sequences and a second LSTM layer with 256 units returning only the final state, receives the combined feature vector together with embedded and padded Hindi text sequences (maximum length 26 words) and generates a descriptive Hindi caption. During inference, a beam search decoder with a beam width of 3 refines the output using log-probability scoring, improving the BLEU-4 score by 33% over greedy decoding. A progressive feature extraction module processes images in chunks of 500, saves intermediate features to disk, and clears GPU memory after each chunk, enabling processing of up to 31,783 images on a standard 8 GB GPU. Experimental results on the Flickr30K-Hindi dataset achieve a BLEU-4 score of 15.31, and on a custom Hindi dataset 43.27, significantly outperforming all prior art on PASCAL 1K-Hindi and Flickr8K-Hindi benchmarks. The system finds industrial application in assistive technology for the visually impaired, automatic image indexing, e-commerce, and educational content creation.

Disclaimer: Curated by HT Syndication.