Analysis of RepVGG on Small Sized Dandelion Images Dataset in terms of Transfer Learning, Regularization, Spatial Attention as well as Squeeze and Excitation Blocks

Abstract

The automated weed detection is an important research field in terms of agricultural productivity and economy. This study aims to apply RepVGG which is a new deep learning architecture developed on PyTorch framework and has promising results when trained and tested on ImageNet1K dataset. 920 images of the small sized Dandelion Images dataset is used for this study. Pretrained vanilla, pretrained and dropout regularized, squeeze and excitation block added and spatial attention block added versions of RepVGG are tested on the dataset. VGG16 method is also applied to the dataset and the results of the MobileNetV2 method is taken from the Kaggle Competition to get an insight about the baseline results of the classical state of the art models. The proposed RepVGG modifications could not outperform the state of the art methods on this dataset but the effect of the modifications are deeply analyzed and the best configuration is obtained by Squeeze and Excitation block added RepVGG-A0 architecture which is trained from scratch for 5 epochs and provided results of 0,875, 0,665, 0,89 and 0,74 for Accuracy, Recall, Precision and F1 metrics respectively.