Bacterial colony detection method for microbiological photographic images

This paper is dedicated to the research and development of automated methods for detecting bacterial colonies on photographic images of Petri dishes. We propose a segmentation algorithm for microbiological photographic images which is based on the convolutional neural network architecture of the encoder-decoder class, trained on artificially generated image data. Our method consists of two stages: sliding window image processing by the proposed neural network to find areas containing isolated objects potentially suitable for further research, and further segmentation of the resulting binary area map into regions containing colonies by classical computer vision algorithms. According to the experimental research, the developed algorithm shows the best results in comparison with the existing neural network algorithms we adapt in this work as well as watershed segmentation algorithms. The work also proposes algorithms for microbiological photographic images artifact searching based on heuristic and neural network approaches (including flare, magnetic balls, barcodes and condensate), the use of which allows to increase the quality of the image segmentation algorithm, as well as to find errors caused by the human factor. The results of this work might be further used for classification algorithms of the bacterial colonies in the segments found, which will allow in some cases fully automated process of visual bacteria recognition on Petri dishes photographic images.