Follicular Ultrasound Image Segmentation based on Improved Deeplabv3
Download as PDF
DOI: 10.23977/cii2019.86
Author(s)
Tianlong Zeng, Jun Liu
Corresponding Author
Tianlong Zeng
ABSTRACT
The ultrasound image segmentation technique of yellow cattle follicles plays an important role in the monitoring of the dynamic changes of follicles in yellow cattle. With the development of deep learning, the image segmentation technology based on deep learning neural network model has made great breakthroughs, such as DeepLabv3, an end-to-end semantic segmentation network model. However, the above algorithm can’t meet the requirements of follicle segmentation in cattle follicle monitoring. Because of its deep network depth, multiple operational parameters, many iterations, and huge amount of computation, it has high requirements for the operating environment of the system (memory, CPU, GPU, etc.), and cannot be easily applied to the actual production practices. Therefore, an improved DeepLabv3 model was proposed in this paper. By removing the ASPP layer and moving the atrous convolution layer forward, the improved model is more suitable for the actual production of follicle segmentation, which is more in line with the economic cost of follicle monitoring. Finally, the experimental results show that the computational resources and the running time of the improved model decrease obviously when the average segmentation precision of follicular ultrasound image does not decrease significantly.
KEYWORDS
DeepLabv3, ASPP, The ultrasound image segmentation technique of yellow cattle follicles, follicle monitoring