Real-Time Classification of Dog Breeds Using the M5StickV

The accurate classification of dog breeds is a challenging task due to the subtle differences between breeds and the computational limitations of portable devices. Despite advancements in deep learning, there is a gap in the literature regarding the deployment of efficient and accurate models on resource-constrained embedded systems. This study addresses this gap by investigating the classification of nine dog breeds using a dataset of 1,080 images, processed through the MobileNetV1 architecture. The focus is on implementing this model within the M5StickV embedded system, a compact and portable device. The images underwent pre-processing, including resizing to 224 by 224 pixels and background removal. Transfer learning with Sipeed was utilized to fine-tune the model, which was subsequently converted to a Keras-compatible format for deployment. The resulting model demonstrated high accuracy (0.9833) and precision (0.9840), underscoring its robust performance. However, a slight delay in the classification process suggests the need for further optimization. This study highlights the potential of integrating advanced deep learning models with embedded systems for real-time, portable image classification, offering applications in pet identification and other fields.