Detection of Hindwing Landmarks Using Transfer Learning and High-Resolution Networks

Simple Summary The landmark annotation of hindwing venation is one of the most important methods in the hindwing morphological, functional, and evolutionary analysis of beetles, and the number of the landmark samples greatly affects the effectiveness of these analysis. However, large-scale manual landmark annotation is a time-consuming task that hinders the progress of wing morphology research. Some machine learning techniques have been applied to beetle image recognition, but the lack of data for beetle hindwing landmarks limits the use of machine learning for beetle hindwing landmark detection. In this study, we propose a new approach to solve the problem of insufficient training samples for beetle hindwing landmark detection, by fine-tuning a new deep high-resolution convolutional neural network pretrained on a natural image database to transfer it to the domain of beetle hindwings. The results of experiments shows the effectiveness of this new approach as it demonstrated small error in the detection of leaf beetle hindwing landmarks and required a very low number of samples. Hindwing venation is one of the most important morphological features for the functional and evolutionary analysis of beetles, as it is one of the key features used for the analysis of beetle flight performance and the design of beetle-like flapping wing micro aerial vehicles. However, manual landmark annotation for hindwing morphological analysis is a time-consuming process hindering the development of wing morphology research. In this paper, we present a novel approach for the detection of landmarks on the hindwings of leaf beetles (Coleoptera, Chrysomelidae) using a limited number of samples. The proposed method entails the transfer of a pre-existing model, trained on a large natural image dataset, to the specific domain of leaf beetle hindwings. This is achieved by using a deep high-resolution network as the backbone. The low-stage network parameters are frozen, while the high-stage parameters are re-trained to construct a leaf beetle hindwing landmark detection model. A leaf beetle hindwing landmark dataset was constructed, and the network was trained on varying numbers of randomly selected hindwing samples. The results demonstrate that the average detection normalized mean error for specific landmarks of leaf beetle hindwings (100 samples) remains below 0.02 and only reached 0.045 when using a mere three samples for training. Comparative analyses reveal that the proposed approach out-performs a prevalently used method (i.e., a deep residual network). This study showcases the practicability of employing natural images-specifically, those in ImageNet-for the purpose of pre-training leaf beetle hindwing landmark detection models in particular, providing a promising approach for insect wing venation digitization.