VOLUME 13 NUMBER 2 (July to December 2020)

Anton Louise P. De Ocampo*¹ and Elmer P. Dadios²

¹Electronics and Communications Engineering Department,
      De La Salle University, Manila, Philippines
²Manufacturing Engineering and Management Department
      De La Salle University, Manila, Philippines

In UAV-based monitoring for smart farms, one of the core tasks in the detection of farmers is the extraction of object-specific features sufficient to emphasize them from the vegetative environment. This article presents the evaluation of recent deep neural network architectures that can be assimilated as automated feature extractors for binary classifiers instead of being standalone. Seven classifiers are first evaluated to find the best candidate for the hybrid network which consists of cascaded DNN and binary classifier models. Support Vector Machine performed best on the farmer dataset (FDS) among the others based on F1-score: naïve bayes, tree classifier, linear discriminant, logistic regression, kNN, and various ensembles. For the selection of feature extractor, a Pareto analysis based on the following selection criteria is conducted: accuracy, precision, recall, F-measure, number of parameters, and the model size. From the 11 deep neural networks considered, only 10 are integrated with the SVM classifier. GoogleNet is excluded from the Pareto analysis since prior experimentation proved it to be performing better as a standalone classifier than as a feature extractor. But, the rest of the group has improved the classifier performance significantly when used as a feature extractor for the SVM classifier. ResNet-18 is identified as the most balanced feature extractor for linear SVM in terms of feature-length, performance on both INRIA and farmer dataset (FDS), the number of layers and parameters, and model size.