Soumyajit Roy

This paper presents our submission for Task 2 of the Bangla Language Processing (BLP) Workshop, which focuses on generating Python code from Bangla programming prompts in a low-resource setting. We address this challenge by fine-tuning the gemma-2-9b instruction-tuned model using parameter-efficient fine-tuning (PEFT) with QLoRA. We propose an iterative self-improvement strategy that augments the extremely limited training data (74 examples) by reusing verified correct predictions from the development set, alongside LoRA rank experiments (8, 16, 32), observing a clear correlation between rank and accuracy, with rank 32 delivering the best results. Compared to translation-based and retrieval-augmented baselines, our approach achieves significantly higher accuracy, with a pass rate of 47% on the development set and 37% on the hidden test set. These results highlight the effectiveness of combining iterative data augmentation with rank optimisation for specialised, low-resource code generation tasks.

In several fields, including marine research, environmental monitoring, and undersea exploration, underwater imaging is essential. Unfortunately, the complicated interactions between light and water molecules make it difficult to take high-quality underwater pictures, leading to problems including low contrast, colour distortion, and poor visibility. Even with improvements in imaging technology, these problems are frequently not adequately addressed by current enhancement methods, particularly in deeper water or under unfavorable circumstances. To address these shortcomings in existing enhancement strategies, this work presents the method of Natural-based Colour Enhancement of Underwater Images (NUCE). NUCE seeks to improve underwater photos' authenticity, boost contrast, and reduce the colour cast, by using a four-step technique that includes colour cast neutralization, fusion of dual-intensity images using the mean and median as averages, mean equalization based on swarm intelligence, and unsharp masking. NUCE's ability to enhance item visibility and detail in underwater scenes by restoring real colours, boosting contrast, and minimizing colour cast has been demonstrated empirically using the EUVP dataset.

Plant diseases cause 70–80% of crop losses globally, making automated detection critical for food security. Traditional visual inspection methods are impractical for large-scale farming. PlantDiseaseNet-RT50 addresses this by employing a fine-tuned ResNet50 architecture with strategically unfrozen layers, custom classification components, and dynamic learning rate scheduling. The model achieves approximately 98% accuracy, precision, and recall across multiple crop species and disease categories. Key improvements include targeted layer unfreezing, batch normalization, dropout regularization, and advanced training techniques. This computationally efficient solution enables rapid, accurate plant disease diagnosis suitable for real-world farming applications.