| Abstract |
The use of deep learning enabled computer vision applications gained momentum after AlexNet convolutional neural network architecture won the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) in 2012. This led implementation of such methods in various application fields including defense industry, smart cities, intelligent video analytics, industrial inspection, and healthcare. Success of these applications enabled agriculture and livestock management to benefit from such implementations in recent years. In this work, we proposed an end-to-end workflow for deep learning-based computer vision applications to facilitate the AI model training, model optimization, model deployment, metadata streaming, metadata recording, result filtering/searching and data visualization. The source data of the visualizations as result of the analysis are stored in a high-performance database presented within the scope of the study. On the other hand, using the state-of-the-art deep learning algorithms and infrastructures, both model training and inference performance-oriented classification model comparison were carried out. In order to compare the developed models with the models developed in other studies in the literature, the training and inference process were tested on the same open-source data set. The workflow is proposed as a hybrid solution that can be deployed both on streamline x86 computer systems and ARM64 embedded devices. Therefore, deployment benchmarks for servers and NVIDIA Jetson devices are also provided within the paper. © 2024 IEEE. |
