Research Of LST Detection Based On Improved YOLO

Lateral Spreading Tumor(LST)is a relatively new subtype of colorectal cancer that is difficult to detect through routine examination and screening methods.If the high-precision target detection algorithm can help doctors find the lesion at an early stage,it can effectively prevent it from further deteriorating into bowel cancer and save the patient’s life.Therefore,this paper makes a series of improvements to YOLOv3 and YOLOv5,and finally obtains a high-precision model-YOLOv5-GDF.The YOLOv5-GDF model introduces the Efficient-RepGFPN and decoupling head structure on the basis of YOLOv5 to improve the feature extraction and semantic information transmission capabilities of the model,and uses the Focal-EIoU loss function to accurately describe the Bounding Box and Ground Truth Box The difference between,and finally build a high-precision model.The main work of this paper is as follows:To enhance the precision of the model in detecting LST,the backbone network ReXNet,Dilated Encoder module and CIoU loss function are introduced to YOLOv3,and YOLOv3-RDC is proposed,and the accuracy is increased from 81.0% to 84.9%;further,in order to obtain high precision The model introduces switchable hole convolution and BN layer channel pruning operation to improve YOLOv5 to YOLOv5-SP,and the final accuracy reaches 86.1%,which is further improved by 1.2% compared with YOLOv3-RDC.To address the issue of low confidence in YOLOv5-SP,this paper enhances YOLOv5 and introduces YOLOv5-GDF.YOLOv5-GDF uses the Efficient-RepGFPN and decoupling head structure to make it have stronger feature expression and semantic information exchange capabilities,and uses the Focal-EIoU loss function to more accurately describe the relationship between Bounding Box and Ground Truth Box,thus improving the performance and robustness of the model.In the end,YOLOv5-GDF mAP reached 87.1%,which was further improved by 1% compared with YOLOv5-SP.This shows that YOLOv5-GDF has excellent ability to detect LST lesions.