Enhancing the classification of Finnish bird images through deep learning : a comparison and optimisation study of CNN architectures

Abstract

Birds are an important indicator of the ecological environment, which is gradually deteriorating due to industrial development and other factors. This study focuses on comparing different Convolutional Neural Networks (CNNs) and attempts to optimise the models to improve their performance in image classification for a wider range of applications. Specifically, it focuses on image classification of bird species inhabiting Finland and improves the accuracy of bird image classification by comparing different convolutional neural network models and attempting to optimise them in as diverse a way as possible. The study also explores the potential of applying these techniques to biodiversity monitoring and their scientific and practical implications.

The study employs a variety of CNN architectures and has trained and tested them through a transfer learning approach. The experimental results show that the deep learning models selected in this study can effectively process and classify highly complex bird image data, among which DenseNet-121 and EfficientNet_B0 exhibit high accuracy and stability under multiple testing conditions. Meanwhile, the optimised CNN models can effectively improve the performance of the models on isolated test sets, with the models using data enhancement techniques showing better generalisation ability. Models using the attention mechanism are more accurate in capturing details and extracting features and show higher accuracy and stability on the test set. In addition, the study explores the efficacy of the models in processing bird images with complex backgrounds, pointing to the importance of optimising deep learning models in ecological research and practical applications.