Performance Evaluation of Pre-Trained Deep Learning Models for Bird Species Identification

The goal of this work was to measure the performance of pre-trained deep-learning models for bird species identification. A flatten layer, three dense layers, two dropout layers, and an output layer with a softmax activation function were utilized in the study's CNN classification models. Precision, recall, and f1-score were some of the evaluation measures used to gauge how well the pre-trained models performed. The results show that the ResNet50v2 and Inceptionv3 models successfully identified bird species with high levels of accuracy, with ResNet50v2 achieving an accuracy of 97.32%. VGG19, on the other hand, has a comparably lower accuracy of 67.82%. These results reveal that pre-trained models have great promise to be effective tools for identifying bird species, with ResNet50v2 and Inceptionv3 providing the most promising results.

I. Introduction

Birds are an essential part of nature, and they play a vital role in the ecosystem. However, when they are present on farms, In some cases, birds view crops as a valuable food source and cause significant damage by pecking at fruits and vegetables, eating seeds, and defecating on crops. Additionally, birds also pose a risk to farm animals, potentially causing disease transmission or stressing animals out. Birds such as crows, sparrows, and starlings, cause significant damage to crops by pecking at fruits and vegetables, eating seeds, and defecating on crops [1]. This results in lower yields, increased costs, and lost revenue for the farmers. [2] Birds that feed on fruits and vegetables cause physical damage to the crops, making them unappealing. Additionally, birds that defecate on crops introduce harmful bacteria and fungi, making them unsafe for human consumption [3].