Generation of microbial colonies dataset with deep learning style transfer.

Area of Science:

• Microbiology
• Computer Vision
• Machine Learning

Background:

• Training deep learning models for image analysis requires large, annotated datasets.
• Acquiring and labeling extensive real-world datasets is resource-intensive and time-consuming.
• Synthetic data generation offers a potential solution to overcome data scarcity challenges.

Purpose of the Study:

• To introduce an effective strategy for generating annotated synthetic microbiological image datasets.
• To enable fully supervised training of deep learning models for microbial analysis.
• To demonstrate the feasibility of using synthetic data for microbe detection, segmentation, and classification.

Main Methods:

• Utilized traditional computer vision algorithms combined with neural style transfer for data augmentation.
• Developed a generator to synthesize realistic microbiological images of Petri dishes.
• Employed a deep learning model for localization, segmentation, and classification of microbial species.

Main Results:

• Synthesized a dataset of realistic images capable of training effective deep learning models.
• Trained models achieved comparable performance (detection mAP, counting MAE) to those trained on significantly larger real datasets.
• Demonstrated successful application in detecting and segmenting five different microbial species.

Conclusions:

• The proposed synthetic data generation strategy is resource-efficient and effective for training deep learning models.
• This method significantly reduces the need for extensive real image collection and manual annotation.
• The approach is versatile and applicable to various scientific and industrial object detection tasks.