LoRA -CNN: Parameter Efficient Convolutional Neural Networks with FAISS -Based Pixel -Level Feature Retrieval

This paper proposes a novel framework that integrates Low-Rank Adaptation (LoRA) modules into Convolutional Neural Networks (CNNs) to achieve parameter-efficient training, while simultaneously incorporating FAISS-based retrieval mechanisms for fast, scalable pixel-level feature lookup. By leveraging LoRA, the number of trainable parameters is significantly reduced without compromising the model’s representational capacity. The integration of FAISS enables efficient similarity searches, making the framework particularly effective for fine-grained feature matching. This method supports computationally demanding tasks such as image patch matching and few-shot image classification with high accuracy and enhanced efficiency. The proposed approach is evaluated on the CIFAR-۱۰ dataset, demonstrating strong performance in both retrieval and classification contexts. Experimental results indicate that the method maintains accuracy comparable to full-parameter models, while achieving substantial improvements in training efficiency and retrieval speed. This makes the framework well-suited for deployment in resource-constrained environments where both performance and efficiency are critical.

کلیدواژه ها:

LoRA ، CNN ، FAISS ، Image Retrieval ، Feature Extraction ، Parameter Efficiency ، Few-shot Learning ، Deep Learning