DKFD: Optimizing Common Pediatric Dermatoses Detection with Novel Loss Function and Post-Processing

Using appropriate classification and recognition technology can help physicians make clinical diagnoses decisions more effectively as a result of the ongoing development artificial intelligence in medical field. There are currently number issues with detection common pediatric dermatoses, including challenge image collection, low resolution some collected images, intra-class variability inter-class similarity disease symptoms, mixing symptom results. To resolve these problems, we first introduced Random Online Data Augmentation Selective Image Super-Resolution Reconstruction (RDA-SSR) method, which successfully avoids overfitting training, to address issue small dataset increase improve quality. Second, for an imbalance between difficult simple samples, is brought on by variation within classes signs during distinct phases. By increasing loss contribution hard samples basis cross-entropy, propose DK_Loss function two-stage object detection, allowing model concentrate learning samples. Third, order reduce redundancy precision, Fliter_nms post-processing method intermingling results based NMS algorithm. We created CPD-10 dermatoses used Faster R-CNN network training findings benchmark. The experimental show that RDA-SSR technique, while needing similar collection parameters, mAP than 4%. Furthermore, experiments were conducted over PASCAL VOC2007 evaluate effectiveness algorithm, cross-entropy loss-function-based baselines. demonstrated that, taken into account, its 1–2% above baseline. confirmed also precision.