Comparative Performance Analysis of MobileNetV3, EfficientNetB0, and DenseNet121 for Android-Based Cat Skin Disease Classification

Abstract

Skin diseases in cats such as ringworm, scabies, and flea allergy dermatitis look visually similar, and many owners cannot reach a veterinarian quickly, which encourages self-diagnosis. This study compares three ImageNet-pretrained convolutional neural networks, MobileNetV3Small, EfficientNetB0, and DenseNet121, for four-class cat skin condition classification (Flea Allergy, Health, Ringworm, and Scabies) and deploys the most suitable model in an Android application. A collection of 1,362 cat images obtained from Roboflow Universe was augmented to 3,274 images and partitioned into 2,868 training, 270 validation, and 136 test images. Each network was used as a frozen feature extractor with a new classification head and trained under identical settings. Evaluation covered accuracy, precision, recall, F1-score, confusion matrices, benchmark inference latency, throughput, and model size after conversion to TensorFlow Lite (TFLite). EfficientNetB0 reached the highest test accuracy (91.18%), followed by DenseNet121 (89.71%) and MobileNetV3 (88.97%), although differences of one to three images fall within the sampling uncertainty of a 136-image test set. MobileNetV3 was the most efficient: its TFLite model occupied 1.96 MB and needed 5.26 ms per image in the benchmark (190.21 frames per second; hardware not recorded), compared with 8.34 MB and 36.40 ms for EfficientNetB0 and 13.90 MB and 135.24 ms for DenseNet121. MobileNetV3 was therefore embedded in an application named Meow Care, which passed all seven black-box test scenarios and reached 88.97% accuracy in on-device evaluation. For resource-constrained mobile screening, a small loss in accuracy can thus be traded for much lower latency and storage. The application is a decision-support aid and does not replace veterinary diagnosis.