Study On SEM Image Instance Segmentation And Improvement Of The Whale Algorithm In Material Design

Titanium dioxide nanotubes,as the main raw material for dye-sensitized solar cells,can greatly improve the photoelectric conversion efficiency of solar cells.Titanium dioxide nanotubes prepared under different process conditions have different material morphologies,which directly affect the performance of the material.However,the relationship between the process and the material morphology is usually verified through continuous experiments in the material design,which wastes a lot of manpower and material resources.At the same time,scanning electron microscope is an important material characterization method,which can extract the tube length,tube diameter and wall thickness of Ti O₂ nanotubes by reading images,but for the dense Ti O₂ nanotube arrays,manual calculation is often inaccurate and time-consuming.Considering the above problems and application scenarios based on SEM images of titanium dioxide nanotubes,we studied the target detection algorithm and size measurement system based on deep learning to measure the size of titanium dioxide nanotubes.A whale optimization algorithm is introduced to train a back propagation network that predicts the morphology of materials prepared under different process parameters,establishing a prediction system for predicting material morphology through the process parameters of the material.The main work is as follows.1.Through a comparative analysis of mainstream target detection algorithms,we chose Mask R-CNN to detect titanium dioxide nanotubes.2.We introduced a capsule network as an image classifier and designed a new capsule network architecture.The image classification accuracy of the capsule network has been improved and applied to Mask R-CNN.3.By reading the scale of the image and the ratio of pixels,we have established a size measurement system for titanium dioxide nanotubes.4.We used the improved whale algorithm instead of the gradient descent algorithm to train the network to improve the accuracy of the neural network’s prediction performance of the material.The capsule network we designed is based on Ge Force GTX 1080Ti and 16GB of RAM.The experimental environment framework is based on Pytorch 1.3.Our model achieved 92.96%accuracy on the CIFAR-10 data set and we introduced it into Mask R-CNN.For Mask R-CNN,we used Tesla K80 GPU and four RAM with 12GB.The experimental environment framework is based on Pytorch 1.5.Our model achieved an accuracy rate of92.3%on SEM images in the target detection experiment.The average error rate of the tube diameter measurement was 5.22%,and the average error rate of the tube length measurement was 1.29%.The neural network trained by the improved whale optimization algorithm predicted the material morphology with an average error rate of 6.65%,which basically realizes the prediction of the material morphology for different process parameters.