Research On Detection Of Multiple Lesion Targets In Wireless Capsule Endoscopy Images

The wireless capsule endoscopy(WCE)system has made rapid progress in the last decade.Optical endoscopes can be placed in small,pharyngeal capsules to wirelessly transmit color images.However,the intelligent detection technology of lesions cannot meet the actual needs of medical treatment.As a new type of endoscopic image lesion detection technology,its research and application development is not yet mature.The traditional pattern recognition technology is difficult to extract effective universal feature vectors for lesion images,so it has extremely important practical significance for its research.The existing single texture analysis methods have achieved good results,verifying the feasibility of texture features in the detection of endoscopic lesion images,but the existing texture analysis methods generally perform endoscopic abnormal image detection and classification.Only for a single lesion.The lack of extraction of multiple texture features is still insufficient,so the paper does the following work.In this paper,the research background and significance of the intelligent identification of lesions are first introduced.For the texture analysis methods,two methods of image texture description,LM filter bank and Local Binary Pattern(CS-LBP),are focused on.Mirror image features,preprocessing steps such as image enhancement,image denoising,normalization processing,and image segmentation are applied to the acquired endoscopic image to provide assistance for extracting accurate texture features.Then,combined with the filter bank and LBP operator,texture information is extracted from the determined lesion blocks and normal image blocks.The KNN method is used to cluster the image block vectors to obtain a texture primitive dictionary.Next,the pre-processed image blocks are processed in the same manner to form an image block dictionary.Finally,based on the image block dictionary,a classifier is designed using the K-nearest neighbor method to study the effect of different texture feature extraction method combinations on the recognition results.The experiment shows that the LM filter bank(48 filters)is combined with CS-LBP.The classification effect is best,and the computational complexity is relatively low and the algorithm time is short.On the other hand,the EDBTC(Error Diffusion Block Truncation Coding)implementation of the color endoscope image is described in detail.The definition of the EDBTC and the use of the LBG(Linde,Buzo,Gray)algorithm are introduced.The overall endoscopic image retrieval is performed.The framework is described and a complete solution is provided for this key point of feature extraction,including vector quantization(VQ),color histogram features(CHF)and bit pattern histogram features(BHF).Finally,this method was compared with the LM+CS-LBP method.The results were simulated by matlab software.The results showed that the accuracy of the method was improved,but the classification time became longer.In summary,this study focuses on lesion detection of endoscopic images,uses a variety of image texture description methods to extract the texture features of the lesio n image,forms a lesion image dictionary and design classifiers,and compares the clas sification effect of different feature extraction methods.Compared with the existing me thods,the method has greatly improved the detection accuracy,and can detect multiple lesions,and can effectively provide the doctor with clinical diagnosis.