Research On Object Detection And Segmentation Algorithm Of Yellow Cattle Follicle Based On Ultrasonic Image

Ultrasound imaging technology is widely used in the monitoring of dynamic changes of yellow cattle follicles because of its advantages of fast,real-time,safety and low price.By searching for the best conception time for breeding,the pregnancy rate of the yellow cattle can be effectively improved,thereby improving the reproductive capacity of the yellow cattle.Target detection and segmentation of yellow cattle follicle ultrasound images is a key step in follicular monitoring.Therefore,the target detection and segmentation of follicle ultrasound images provide good pre-conditions and basic data for subsequent three-dimensional reconstruction of follicles,volume calculation,follicle analysis,etc.,so as to better realize follicular dynamic monitoring.For the target detection task of follicle ultrasound image,according to the fact that the data set has the spatial position,size and shape of the follicle,the YOLOv3 target detection algorithm with the best detection effect and the fastest detection method is used as the basis of this experiment.An improved YOLOv3 network model,which compresses the network structure by removing the batch normalization layer in the model,reduces network parameters,thereby improving the network training speed;and then expands the network receptive field by introducing atrous convolution layer,so that the network model can make more use of the context information in the network to improve the detection accuracy of the model.The experimental results show that the improved YOLOv3 model improves the detection accuracy by 1.2%,and the detection speed increases from 63ms/img to 51ms/img.Compared with the original YOLOv3 model,the improved YOLOv3 model is more suitable for dynamic monitoring of follicular target detection tasks in the project in both performance and detection.Follicular ultrasound image segmentation is the second major work of this paper.Firstly,because of the characteristics of single-target single-class of follicular ultrasound image dataset,the semantic segmentation model based on deep learning image is very suitable for this segmentation task.In this paper,the most advanced DeepLabv3 network model is used as the experimental basis of this research.According to the specific situation of this research,an improved DeepLabv3 model is proposed.By removing the ASPP structure in the model,the parameter volume of the network is reduced,thus improving the training speed,thereby reducing the dependence of the model on the computing power of the system.Then,the atrous convolution layer of the network is moved forward to reduce the problem of the descent of segmentation precision caused by the removal of the ASPP structure.In addition,the ResNet50 network with less layers is used to further reduce the number of parameters of the network,so that the new model is accurately segmented in pixels when the rate is not obvious,the amount of calculation is greatly reduced,and the training speed is greatly improved.Experiments show that the improved DeepLabv3 only reduced the average pixel segmentation accuracy by 2.4% compared with the original model,but the model parameters are reduced by nearly 40%,and the training speed is increased from 1.78 img/sec to 2.96 img/sec,indicating that the improved model in calculating the cost and environmental dependencies with the balance of the above image segmentation accuracy,more meet the demand of practical application of dynamic monitoring follicle.In addition,for the characteristics of poor image quality,this paper also carried out a supplementary experiment of image super-resolution reconstruction.Through super-resolution reconstruction experiment,the resolution of follicle ultrasound image was improved,and the accuracy of target detection and segmentation of follicle ultrasound images was obtained.