Automated micro-plastic detection and classification using deep convolution neural network pre-trained models and transfer learning

Micro-plastics, which are plastic particles less than 5 mm in size, pose significant environmental threats due to their persistence and potential toxicity to marine life and humans. This work compares the effectiveness of several convolutional neural network (CNN) designs, including MobileNetV3Large, ResNet50V2, ResNet101V2, and EfficientNetB7, in identifying and categorizing microplastics into three groups: beads, fibers, and fragments. We evaluate the models using precision, recall, and F1-score criteria. The outcomes indicate that while all models perform well in identifying fiber microplastics, achieving high recall and moderate precision, they struggle significantly with bead and fragment categories. EfficientNetB7 and MobileNetV3Large exhibited the highest performance for fiber detection but failed to detect bead and fragment microplastics. The findings highlight the need for further research to enhance the classification accuracy for bead and fragment micro-plastics, suggesting that future work should focus on addressing class imbalance, utilizing advanced techniques such as transfer learning, and incorporating domain-specific knowledge to improve feature discrimination. This work provides the path for more efficient environmental monitoring systems by offering insightful information about the advantages and disadvantages of the CNN architectures now in use for the detection of microplastics.