Biometric Recognition of African Clawed Frogs

The African clawed frog (Xenopus laevis) is a commonly used model organism for cell biological, developmental, and biomedical research. For health monitoring and experimental quality control purposes, it is desirable to identify individual frogs regularly throughout their life. Current methods for identification are often invasive and associated with significant investment costs. Identification based on images of the biometric pattern on a frog’s back has been implemented in some laboratories, but so far has been performed manually and therefore is time-consuming and limited to small group sizes. This work proposes a novel pipeline for data acquisition, pre-processing, and training of a classification model based on pattern recognition. The pipeline is structured around laboratory frog colonies and smartphone usage. In order to achieve a lightweight system in our evaluation we consider a MobileNet ConvNet pre-trained on ImageNet. Two feature sets are evaluated on a new data set of 1,647 image samples collected from 160 frogs: RGB images, and 3-channel contour maps (i.e. CORF3D). The results indicate that the CORF3D feature set is favoured over RGB. CORF3D achieved the best performance of 99.94% average accuracy, while RGB had the best performance of 98.79%. Analysis of misclassifications shows that bad predictions are often caused by bad lens focus, light reflections, and positional inconsistency in pattern extraction, which can be addressed during data acquisition. The proposed methodology is, therefore, an effective solution for the recognition of Xenopus laevis