Leveraging spectral information for enhanced multispectral image retrieval using deep neural network

Abstract

Similar multispectral images for a given query image can be retrieved from large databases of multispectral images. These images provide more detail due to multiple bands, but the high dimensionality creates a challenging, inefficient environment for retrieving similar images. This research introduces a method for multispectral image retrieval based on the DenseNet deep learning framework. DenseNet's use of dense connections between network layers enables the reuse of previously learned features and ensures that network layers support better representation of both spectral and spatial attributes of images, reducing the likelihood of important attributes being lost during data transfer through the DenseNet architecture. Deep features will be compared using two methods of measuring similarity: Euclidean distance and Cosine distance. The research finds that the DenseNet multispectral image retrieval approach provides a Mean Average Precision (MAP) score of 0.73 for the top 100 retrieved images based on Euclidean distance, which exceeds the MAP@100 score of 0.68 obtained from prior retrieval methods on the same dataset. When DenseNet is compared to state-of-the-art hash-based image retrieval methods, it consistently provides the highest MAP score for all tested code lengths, demonstrating that the DenseNet approach preserves the most critical spectral and spatial information. Overall, these results demonstrate that the DenseNet multispectral image retrieval framework is highly effective and efficient, rendering it an applicable solution for managing large multispectral image collections for actual remote sensing applications such as land use studies, environmental monitoring, or geo-spatial data management.