Research On Endoscopic Image Artefact Detection Algorithm Based On Improved Full Convolutional Network

Endoscopy is a clinical procedure,treatment procedure,and minimally invasive surgery(such as laparoscopy)widely used for early detection of a variety of cancers(such as nasopharyngeal cancer,esophageal adenocarcinoma,gastric cancer,colorectal cancer,bladder cancer,etc.).However,endoscopic inspection of video frames can cause artifacts(e.g.pixel saturation,motion blur,defocus blur,specular reflections,bubbles,fluids,and debris)that can cause imaging damage.Accurate detection of artifacts in clinical endoscopes can greatly accelerate the development of effective quantitative clinical endoscopic analysis across all diseases,organs and methods.With the application and development of deep convolutional neural networks in natural images,more and more researchers will apply them to medical images.In this paper,aiming at the problem that artifacts in endoscopic images are difficult to accurately detect,an object detection algorithm based on deep learning is applied to the positioning and classification of artifact images.At the same time,the effects of different target detection algorithms on endoscopic artifact detection are compared,including two-stage object detection algorithm based on anchor box and one-stage object detection algorithm.At the same time,based on the full convolutional neural target detection algorithm FCOS,detailed experiments have been performed,and a certain degree of improvement has been made in terms of detection accuracy and speed.By adding weakly supervised information for semantic segmentation,the overall spatial location information of the model is increased.Aiming at the shortcoming that IoU Loss of the original regression branch of the FCOS model cannot accurately reflect the degree of coincidence of the bounding box,this paper improves the loss function of the FCOS regression branch to GIoU Loss;At the same time,the method of center sampling is added to improve the regression quality of the model's bounding box.Through a series of improvements and enhancements to the FCOS algorithm in this paper,the detection accuracy of the FCOS algorithm exceeds that of the one-stage detection algorithm and two-stage detection algorithm based on the anchor box.Finally,in order to solve the problem that FCOS requires more computing power and a larger amount of parameters in the algorithm running process,this paper prune the channel of FCOS,which greatly reduces the amount of FCOS parameters.At the same time,the performance of the model after channel clipping will be reduced to a certain extent.Feature distillation is performed on the model.The original large model is used to supervise the training of the small model after channel clipping.The detection accuracy is improved without increasing the amount of model parameters.In this way,by improving and upgrading the FCOS series,channel clipping and feature distillation,the improved FCOS algorithm in this paper improves the detection accuracy of the algorithm,while greatly reducing the complexity of the model,making the FCOS algorithm more suitable for application in practical scenarios.Experimental results show that on the EAD2019 data set,compared with the onestage detection algorithm RetinaNet,the improved FCOS algorithm in this paper is 8.1% higher than RetinaNet on the mAP,but the amount of algorithm parameters is only about 1/8 of RetinaNet.Compared with the two-stage detection algorithm FPN,the improved FCOS algorithm in this paper is 0.3% lower than FPN on the mAP,but the parameter amount of the algorithm is only about 1/10 of FPN.At the same time,compared with the original FCOS algorithm,the improved FCOS algorithm in this paper is 1.9% higher than the original FCOS on the MAP,but the parameter amount of the algorithm is only about 1/8 of the original FCOS.The experimental results show that the improved FCOS algorithm has great advantages in both detection accuracy and model parameter quantity.