Publication

Abstract : This research study proposed a novel methodology that enhances image classification accuracy by incorporating spatial analysis-driven confidence scores. The proposed pipeline is a synergy of image segmentation, spatial analysis, and state-of-the-art machine learning techniques, aiming to improve classification outcomes by leveraging finer contextual information. The Dermnet dataset, a comprehensive collection of dermatology images encompassing a diverse range of skin conditions, is utilized for evaluation. The proposed approach is evaluated across several prevalent neural network architectures to assess its applicability in real-world scenarios. The first step of the proposed pipeline employs the U-Net architecture for image segmentation, effectively identifying regions of interest within the images. These segmented regions form the foundation for subsequent spatial analysis. The spatial analysis stage capitalizes on the insights derived from the segmented regions, calculating density maps to capture object distribution patterns within the images. By integrating this spatial information, the proposed pipeline augments classifier confidence scores, enabling enhanced discrimination between different classes. To validate our methodology, we conduct extensive experiments using the Dermnet dataset. Notably, this study employs a selection of widely adopted neural network architectures, including ResNet34, VGG16, DenseNet121, InceptionV3,and EfficientNet. The results showcase substantial improvements in classification accuracy across the evaluated models, thereby affirming the effectiveness of the spatial analysis-driven confidence scores. Specifically, the accuracies obtained are as follows: ResNet34 (0.918), VGG16 (0.876), DenseNet121 (0.942), InceptionV3 (0.971), and EfficientNet (0.906).