Research On Object Feature Extraction And Action Recognition In Chemical Analysis

In chemical analysis experiments,the use of intelligent equipment such as robots or robotic arms to replace human beings to complete some experiments with high risk and high precision requirements has become a trend of intelligent chemical analysis experiments.How to quickly teach a robot to complete a set of complex chemical analysis actions,and use deep learning to identify various chemical analysis experimental instruments and behaviors is the key to the intelligence of chemical analysis experiments.This paper explores the object feature extraction and recognition algorithm and action recognition algorithm in the laboratory operating environment,and combines the chemical analysis experiment with deep learning to provide a certain research foundation for realizing the intellectualization of chemical analysis.The specific research contents of this paper are as follows:（1）Self-made experimental instrument image data sets and video action recognition data sets in a laboratory environment,pre-process the pictures of experimental instruments,and make VOC data sets to label experimental instruments;pre-process video data and cut them into action samples with the same number of frames And decompose it into image frames of the same size,centrally crop and enhance.（2）Improvement of object feature extraction and recognition algorithm: Aiming at the problem of object recognition in the chemical experiment environment,this paper improves the yolov5 target detection algorithm.When improving the feature extraction network,efficientnetv2 model and hole convolution are used.At the same time,in order to further improve the accuracy of the model,the original loss function is improved.Its complete loss function includes rectangular box prediction loss function Confidence prediction loss function and category prediction loss function are three parts.For rectangular box prediction loss function,CIO loss function is used,and category prediction uses binary cross entropy function,in which adjustment factors are added.The improved yolov5 model is trained by transfer learning.The experiment shows that the accuracy rate on the self-made data set reaches 93.7%.（3）Improvement of behavior recognition algorithm: starting from the direction of improving network learning,improving classification accuracy and reducing network computation,this paper uses efficientnetv2 algorithm to replace the dual stream convolution network,adds void convolution in the network,and explores the fusion of RGB image flow and optical flow in the network,using the joint loss function.The improved two stream convolution neural network is trained by transfer learning.The experiment shows that the accuracy of the improved two stream convolution neural network model in the self-made chemical experiment behavior data set can reach 92.4%.