Model Comparison for Toxigenic Cyanobacteria Image Classification using Transfer Learning and Post-Quantization
Toxigenic cyanobacteria pose a significant threat when it comes to marine wildlife. They can produce toxins and can quickly grow out of control, directly impacting Sustainable Development Goal 14 (Life Below Water). This research supports SDG 6 (Clean Water and Sanitation) and focuses on transfer learning to classify microscopic cyanobacteria images using a publicly available dataset and quantization for edge deployment. We tested several pre-trained Convolutional Neural Networks (CNNs), including VGG16, ResNet50, InceptionV3, MobileNetV2, EfficientNetB0, and EfficientNetB1, and were fine-tuned and tested. The results indicated that EfficientNetB1 managed to achieve the highest test accuracy of the tested models with a score of 96.72% and outperforms previous research’s benchmarks. We applied post-training quantization with FP16 quantization to the model and achieved identical results as the original model while also reducing the model size by 50% and slight improvement in inference time.
Authors:
Benjamin Manafe, Marcell Kristianto, Dave Christian Thio, Gregorius Natanael Elwirehardja, Bens Pardamean
2025 International Conference on Biospheric Harmony (ICOBAR)