Multi-Class Detection of Skin Disease: Detection Using HOG and CNN Hybrid Feature Extraction

Abstract

AbstractIt is essential to monitor and analyse skin problems early on in order to prevent them from spreading and turning into deadly skin cancers. Due to artefacts, poor contrast and similar imaging of scars, moles and other skin lesions, it is difficult to distinguish skin diseases from skin lesions. As a consequence, automated skin lesion identification is performed using lesion detection methods that have been optimised for accuracy, efficiency and performance. Photographs of skin lesions are used to illustrate the suggested technique. To assist in the early detection of skin lesions, the proposed method uses CNN, GLCM and HOG feature extraction. The files include many skin lesions of various kinds. The suggested work includes a pre-processing step that aims to improve the quality and clarity of the skin lesion and to remove artefacts, skin colour and hair, amongst other things. Then, using geodesic active contours, segmentation is done (GAC). Skin lesions may be separated separately during the segmentation step, which is beneficial for subsequent feature extraction. The proposed system detects skin lesions via the use of feature extraction methods such as CNN, GLCM and HOG. Score features are extracted using the CNN technique, whilst texture features are extracted using the GLCM and HOG methods. After collecting characteristics, a multi-class SVM classifier is utilised to categorise skin lesions. Using ResNet-18 transfer learning for feature extraction, many skin diseases, including malignant lesions, may be rapidly classified.KeywordsGeodesic active contourGrey-level co-occurrence matrixHistogram of oriented gradientsConvolution neural networkResNet-18SVM classifier