Classification of Canine Skin Diseases Using a Custom CNN and Transfer Learning Models

Early diagnosis of canine skin diseases is important for timely treatment, yet similar clinical appearances make accurate classification challenging. This study proposes Convolutional Neural Network architecture for classifying canine skin diseases using a custom CNN and three transfer learning models, including MobileNetV2, EfficientNet-B0, and RegNetY32GF. The dataset, collected from Roboflow, contains 5,024 original images from 11 classes. Images were split before preprocessing, and class balancing and augmentation were applied exclusively to the training set. The custom CNN incorporates residual learning, multi-scale feature extraction, and CBMA attention and is trained entirely from scratch without ImageNet-pretrained weights. Experimental results show that RegNetY-32F achieved the highest test accuracy of 89.26%, while EfficientNet-B0 and MobileNetV2 showed accuracies of 88.59% and 87.67%. The custom CNN achieved an accuracy of 84.75%. The study provides a comparative investigation of independently designed custom CNN and pretrained architectures for broader multi-class canine skin disease classification. The custom CNN was further implemented in an Android-based real-time application, demonstrating its practical applicability for image-based canine skin disease classification.