Research On Machine Learning Based Sequential Ultrasound Bovine Follicular Image Set Segmentation Algorithm

Ultrasound image segmentation technology plays an important role in the monitoring of follicles in cattle.By monitoring the follicles of cattle and searching for the best timing of pregnancy,breeding can effectively improve the breeding ability of cattle.With the maturation of the cattle industry,in order to pursue greater economic benefits,higher requirements have been put forward for ultrasonic image segmentation technology.In this paper,according to the characteristics of continuous ultrasound follicular image sets,some commonly used image segmentation algorithms and their applications in ultrasonic image segmentation are analyzed and compared.Ultrasonic images contain a large amount of speckle noise and blurry edges.Traditional image segmentation algorithms are difficult to obtain good segmentation results on ultrasound images.As a supervised machine learning method,support vector machine(SVM)can learn the characteristics of samples well through the training of sample sets,so as to obtain higher segmentation accuracy.However,when the traditional SVM model processes continuous data sets,it can only train the segmentation model for each image.Therefore,sample extraction is required for each image,which takes time and effort.This paper improves the traditional SVM model and combines a feature that can characterize image continuity with the features of the image pixels to train a model that can segment the entire image set.This model greatly reduces the workload of extracting samples while ensuring certain segmentation accuracy.This paper also carries out some research on the automatic selection of ROI.The image segmentation model Mask R-CNN based on deep learning combines target monitoring with segmentation,and does not require artificial selection of ROI region;however,the RPN layer in this model is trained Candidate frames are not suitable for image segmentation tasks.This paper also focuses on the characteristics of candidate frames required for image segmentation tasks,and improves the RPN layer of Mask R-CNN,making the training candidate frame more suitable for image segmentation tasks.The experimental results show that the Mask R-CNN model with improved RPN layer is more accurate than the original model when the appropriate parameters are selected.