Using convolutional neural networks for tick image recognition - a preliminary exploration.

Area of Science:

• Computer Vision
• Machine Learning
• Entomology

Background:

• Citizen science initiatives increasingly use smartphone images for tick data collection.
• Deep learning models offer potential for automated tick image classification.
• No current mobile or web applications provide automated tick image classification.

Purpose of the Study:

• To compare the accuracy of a pre-trained deep learning model (ResNet-50) against a custom-built shallow convolutional neural network (CNN) for tick classification.
• To evaluate model performance on common hard ticks in northeastern USA anthropic areas.

Main Methods:

• A dataset of approximately 2000 images of four tick species, two sexes, and two life stages was created.
• Two CNN models were trained: ResNet-50 and a shallow custom-built model.
• Model performance was evaluated on an independent test set of tick images.

Main Results:

• The shallow custom-built CNN achieved higher training (99.7%) and validation (99.1%) accuracies than ResNet-50.
• The shallow model demonstrated superior performance on new data, with 80% mean prediction accuracy and 88.7% true detection confidence.
• The shallow custom-built model exhibited a faster mean response time (3.64 seconds).

Conclusions:

• A simple, shallow custom-built CNN model is effective for classifying common hard ticks in northeastern USA.
• This model shows significant potential for application in mobile or web-based tools for citizen scientists, even with limited training data.
• Shallow CNNs offer a viable alternative to complex deep learning models for specific image classification tasks with constrained datasets.