[Deep learning network-based recognition and localization of diatom images against complex background].

Objective We propose a deep learning network-based method for recognizing and locating diatom targets with interference by complex background in autopsy. Method The system consisted of two modules: the preliminary positioning module and the accurate positioning module. In preliminary positioning, ZFNet convolution and pooling were utilized to extract the high-level features, and Regional Proposal Network (RPN) was applied to generate the regions where the diatoms may exist. In accurate positioning, Fast R-CNN was used to modify the position information and identify the types of the diatoms. Results We compared the proposed method with conventional machine learning methods using a self-built database of images with interference by simple, moderate and complex backgrounds. The conventional methods showed a recognition rate of diatoms against partial background interference of about 60%, and failed to recognize or locate the diatom objects in the datasets with complex background interference. The deep learning network-based method effectively recognized and located the diatom targets against complex background interference with an average recognition rate reaching 85%. Conclusion The proposed method can be applied for recognition and location of diatom targets against complex background interference in autopsy.