Research On Semantic Segmentation Methods Of Microbial Colonies Based On Deep Learning

Microbial colony image processing and analysis are important in a variety of disciplines,including food safety,health care,and environmental protection.Usually,the processing of microbial colony images is performed manually.Still,these images often have incredibly complex target shapes and multiple targets that overlap in position,making it extremely difficult for the staff to analyze the information contained in the images.Furthermore,manual processing and analysis methodologies have drawbacks such as low efficiency,long processing times,and high personal subjectivity,all of which have a significant impact on the accuracy of evaluation for sampled targets.In this paper,we apply deep learning techniques to study the semantic segmentation of microbial colony images and propose two methods based on deep convolutional neural networks,with the following main contributions:(1)In this paper,we propose a Deep Lab V3+-plus-HRNet network(DPHNet)based on deep convolutional neural network fusion,which combines the classical model Deep Lab V3+ in the field of semantic segmentation and the newly proposed high-resolution network(HRNet)in recent years.On the one hand,the Deep Lab V3+network module is applied to preserve the shape features of microbial colonies while identifying overlapping colonies with complexities.On the other hand,the HRNet network module is utilized to maintain the high resolution of the colony images during processing while achieving high accuracy of microbial colony classification.Then,the processing results of the Deep Lab V3+ module are grayed out to be used as the boundary image of the colonies.At the same time,the processing results of the HRNet module are used as the classification image of the colonies.Then,the boundary image and the classification image of the colony are fused by a simple and effective postprocessing method to achieve accurate classification of dense colonies and obtain clear and precise boundaries simultaneously.Finally,extensive experiments are conducted on our self-made colony dataset,and the experimental results show that DPHNet outperforms Deep Lab V3+ and HRNet for microbial colonies.(2)To further simplify the model complexity while improving the segmentation of microbial colonies,we propose a simple but effective add-on module,the cross strip attention(CSA)module.CSA captures global background features while retaining precise location information,increasing the amount of information available during colony image processing,and facilitating the segmentation of the overall shape of colonies as well as boundary details and the classification of different colonies.Specifically,there are two parallel attention modules in CSA,horizontal and vertical attention modules,which are utilized to acquire long-distance feature dependencies in the horizontal and vertical directions in colony images,respectively.In addition,the features obtained in both directions are fused,which enhances the robustness of the resulting features.Finally,CSA is inserted into the residual module of HRNet in parallel in the form of add-on modules to take advantage of HRNet’s consistently high resolution of colony images in order to generate microbial colony segmentation maps with accurate classification and clear boundaries.Extensive experiments were conducted on the self-made colony dataset,and the results showed the effectiveness of the CSA module designed in this paper.