The human skin plays a huge part in a person's physical appearance, and it is the biggest organ of the human body. It regulates body temperature, sensing from touching heat and cold. However, there are a number of risks that affect the skin, one of which is a disease. Fungus, bacteria, allergies, enzyme, and viruses cause most skin diseases. Identify the disease on the basis of manual feature extractions or based on the symptoms is time consuming and it requires extensive knowledge for perfect identification. Diagnosing, detection and classification of skin diseases is done by researchers previously. However, the recognition rate is still not enough and is dependent on feature selection, filtering and segmentation methods. In the previous work tinea pedies and tinea capitis are not identified. In this thesis, we developed a model using Convolutional Neural Network (CNN) for feature extraction and Support Vector Machine (SVM) for classification. CNN is state of the art for deep feature extraction, hence we used it for feature extraction. The model used to detect and classify human skin diseases such as tinea corpories, tinea pedies, and tinea capitis. The dataset is collected from patients at Bahir Dar Tibebe Giyon specialized Hospital, Bahir Dar Felege Hiywet General specialized Hospital and from skin disease image repository. The amounts of collected images are 65, 58, and 62 for tinea capitis, tinea corpories, and tinea pedies respectively. From the total 185 images 80 images are from medical image repository. After collecting datasets, Image augmentation, Image Preprocessing, and Image Segmentation techniques are applied to increase the performance of human skin disease identification. In image preprocessing, we normalize the image to 3 image sizes which are 200x200, 224x224, and 300x300. We also apply median filtering to remove noise and we used Histogram equalization to balance the intensity of image. In segmentation, we used integration of threshold and region based segmentation methods. From 1196 image dataset, we used 80% of image for training and 20% for testing. After the model is evaluated, we have achieved 95.6% test accuracy and 95.6% training accuracy which is 11.6% better than AlexNet model.