A Study On Improvement Of Small Object Detection Models For Incompletely Annotated Cervical Cancer Cells Dataset

Traditional object detectors usually require fully annotated target instances for training to achieve satisfactory detection results.However,for sparsely labeled image instances,especially in medical images where sample labeling is difficult,their detection results are still far from satisfactory.In this regard,Co-mining self-supervised learning sparse annotation object detectors perform well,but they exhibit significant errors for smaller targets.Therefore,based on this model,this paper proposes the CO-FCOS model to address the issues of small targets and incomplete annotations,aiming to design a better cervical cancer cell target detection algorithm to assist doctors in diagnosis.The main research contents are as follows:To address the issue of incomplete labeling,this paper proposes a label-level consistency strategy,adopts more efficient data augmentation strategies,and introduces the OTA dynamic label assignment strategy.Breaking the limitations of the original model on another branch of data augmentation methods,it generates more pseudo-labels,fully utilizes the position and category information of existing objects in images,enriches training data,and improves the performance of the model and its adaptability and generalization ability in different scenarios and data distributions.Experimental results show that on the cervical cancer TCT dataset,compared with the Co-mining detector,the proposed CO-FCOS model has a performance improvement of 4.3 m AP,and the accuracy is improved by 1.7m AP on the challenging public COCO sparse labeling dataset COCO-miss50.2)In response to the problem of small objects,this paper added a CBAM module to the detection head to implement attention mechanism and replaced all ordinary convolutions in the model with deformable convolutions.By introducing attention mechanism,the model can calculate attention weights,which enables the model to automatically learn which parts or channels are more critical for the current task,thus highlighting important information and enhancing the feature representation ability of the model for small objects.At the same time,introducing DCNv2 deformable convolution layer helps the model to better capture the features of small objects and model the features of different regions more finely within the receptive field,thereby improving the accuracy of detecting small objects.Through verification on the cervical cancer TCT dataset,the detection accuracy of small objects was improved by 0.9 m AP and 1.8 m AP,respectively.3)Developed a doctor-side cell image detection system.In order to promote the practicality of the CO-FCOS model proposed in this paper,a cell diagnosis target detection system was developed for medical users based on both B/S and C/S architectures,on the basis of this model.The system includes a web end and a client end.