Performance Analysis of Deep Neural Networks UsingComputer Vision

EAI Endorsed Transactions on Industrial Networks and Intelligent Systems

Nidhi Sindhwani¹  Vikash Yadav²  Rohit Anand³  Meivel S.⁴  Rati Shukla⁵  Mahendra Yadav⁶ 

[1] Amity University, Noida, India;Department of Technical Education, Uttar Pradesh, India;G.B.Pant DSEU Okhla – 1 Campus, New Delhi, India;M.Kumarasamy College of Engineering, Karur, Tamil Nadu, India;MNNIT Prayagraj, Allahabad, Uttar Pradesh, India;VIT Bhopal University, Madhya Pradesh, India;

关键词: computervision;    objectdetection;    deeplearning;    deepneuralnetwork;    imageclassification;    artificial intelligence;    machine learning;    convolutional neural network;    accuracy;   

DOI  :  10.4108/eai.13-10-2021.171318

来源: DOAJ

【 摘 要 】

INTRODUCTION:Inrecentyears,deeplearningtechniqueshavebeenmadetooutperformtheearlierstate-of-the-art machine learning techniques in many areas, with one of the most notable cases being computer vision. Deep learning is also employedtotraintheneuralnetworkswiththeimagesandtoperformthevarioustaskssuchasclassificationandsegmentationusingseveraldifferentmodels.Thesizeanddepthofcurrentdeeplearningmodelshaveincreasedtosolvecertaintasksasthesemodelsprovidebetteraccuracy.Aspre-trainedweightsmaybeusedforfurthertrainingandpreventcostlycomputing,transferlearningisthereforeofvitalimportance.Abriefaccountisgivenoftheirhistory, structure, benefits,anddrawbacks,followedbya descriptionoftheirapplicationsinthedifferenttasksofcomputervision,suchasobject detection, face recognition etc. OBJECTIVE: Thepurposeofthispaperistotrainadeepneuralnetworktoproperlyclassifytheimagesthatithasnever seenbefore,definetechniquestoenhancetheefficiencyofdeeplearninganddeploydeepneuralnetworksinvarious applications. METHOD:Theproposedapproachrepresentsthatafterthereadingofimages,256x256pixelimage’srandompartsare extracted and noise, distortion, flip, or rotation transforms are applied. Multiple convolution and pooling steps are applied by controlling the stride lengths. RESULT: Data analysis and research findings showed that DNN models have been implemented in three main configurations of deep learning: CNTK, MXNet and TensorFlow. The proposed work outperforms the previous techniques in predicting the dependentvariables,learningrate,imagecount,imagemean,performanceanalysisoflossrateandlearningrateduring training, performance Analysis of Loss with respect to Epoch for Training, Validation and Accuracy. CONCLUSION:Thisresearchencompassesalargevarietyofcomputerapplications,fromimagerecognitionandmachine translation to enhanced learning. DNN models have been implemented in three main configurations of deep learning: CNTK, MXNetandTensorFlow.ExtensiveresearchhasbeenconductedusingthevariousdeeparchitecturessuchasAlexNet, InceptionNet,etc.Tothebestofauthors’knowledge,thisisthefirstworkthatpresentsaquantitativeanalysisofthedeep architectures mentioned above.

【 授权许可】

Unknown