Research On Application Of Target Recognition Technology In Robot Vision System

Target recognition is an important topic in the field of artificial intelligence.It has been widely applied to medicine,robot vision,intelligent transportation and remote sensing.In robot vision,the target recognition requires more emphasis on high precision,real-time performance and practicality.It is required to be able to adapt to the harsh noise environment at the scene.This paper uses the improved CNN framework to classify and identify the Caltech dataset and self-built office supplies dataset.It also analyzes the advantages and disadvantages of the classical algorithm.At the same time,combining the improved CNN framework with wavelet adaptive threshold denoising to build intelligence,the office robot target recognition system is proposed.This article proposes three target identification methods.For the problem that long time and low efficiency of traditional classification methods of large-scale image set recognition,the classical framework AlexNet was introduced.A target recognition method based on AlexNet activation function layer feature extraction is proposed.The main idea is to train the AlexNet model,select the ReLU activation function layer of the trained model for feature extraction,and send the feature vector to a multi-class SVM for training and recognition.The experimental results show that the recognition rate of this method is obviously improved compared with both the classical AlexNet framework and the method of selecting the fully connected layer as the feature extraction layer.For the CNN framework,the ReLU activation function gradient death leads to the problem of low recognition accuracy.LeakyReLU and PReLU activation functions are introduced,and a CNN framework based on LeakyReLU and PReLU activation functions is proposed.LeakyReLU combines with the PReLU activation function to construct PL-ActNet and weight the negative data.A PL-ActNet based activation function layer feature extraction target recognition method is proposed.Experimental results show that the framework performs negative value data.Weighted optimization effectively improves the recognition accuracy.For the problem that the robot target recognition system is affected by the noise environment when real-time images are captured.The wavelet adaptive threshold denoising method is introduced.Different wavelet functions and threshold functions are used to denoise the image.PL-ActNet based wavelet adaptive threshold noise reduction target recognition method is proposed.This method integrates the above work and further increases the recognition accuracy and robustness.The experimental results show that the noise reduction method solves the problem of hard threshold discontinuity and soft threshold decomposition of the fixed image generated by the reconstructed image,which effectively improves the image quality and the target recognition accuracy.The accuracy rate of this method reached 88.91% in the Caltech101 dataset,reached 96.67% in self-built office supplies data set.This article has important contributions to the research of target recognition technology.