Optimizing DenseNet for Image Classification: A Synergy of Dual-Adaptive Attention and Multi-Scale Feature Extraction

초록

In this paper, we introduce a new optimization method for DenseNet. In general, the most difficult part of medical image processing is the difficulty of improving performance for certain areas, such as images of skin, and we complemented this with a multi-scale feature extraction module that captures details at various resolutions. To ensure comprehensive learning, we used an adaptive loss function and a lite moderated channel block method that fine-tuned the model with detailed mathematical corrections to optimize sensitivity and specificity. Through experiments, we were able to confirm the excellence of the model as a result of extensive testing on complex data sets, and achieved an accuracy of 96.67%, proving that it is an efficient and reliable DenseNet model. This is a superior result not only to the general DenseNet model but also to other famous models such as MobileNet and VGG19, showing that it can be used in specific areas targeting skin images.

키워드