تشخيص جسم برجسته در تصاوير با روش‌هاي يادگيري عميق

چكيده انگليسي :

Salient Object Detection using Deep Learning Methods Navid Ziaei n ziaei@ec iut ac ir June 16 2019 Department of Electrical and Computer Engineering Isfahan University of Technology Isfahan 84156 83111 Iran Degree M Sc Language Farsi Supervisor Prof Saeid Sadri sadri@cc iut ac ir Supervisor Dr Behzad Nazari nazari@cc iut ac ir Abstract One of the most active fields in Image processing and machine vision is saliency detection While Human beings caneasily perceive the distinction between parts of an image and concentrate on a specific part of it this is a hard and complicatedtask for computers Considering saliency as a psychological phenomenon this issue was firstly studied by researchers incognitive science and psychology But nowadays this subject has attracted interest in computational sciences such asmachine learning The reason of this interest is its application in such domains as object detection image compression andtarget detecting and tracking One of the saliency detection areas is salient object detection The purpose of salient objectdetection is to determine and segment the first meaningful object at first glance by majority of people At first researchesused basic features such as color light intensity color contrast and other low level features in their classifying methods But these methods were not suitable for most of the complex and crowded images Therefore using higher level featureswas a necessity These features were originally extracted manually Afterward with the development and expansion of deeplearning algorithms these methods replaced the former ones One of the most popular structures used saliency detectionis Convolutional Neural Networks CNN CNN s ability to extract features leads to great progress in saliency detection In this research the theoretical and basic foundations for salient object detection is studied Then the relevant databasesand evaluation criteria of them have been investigated In this research we chose UNet as the basic architecture for salientobject detection It was shown that this architecture has some shortcomings in salient object detection Thus we developedfour new architectures In the first network we increased the convergence and trainability of the network by adding residualblocks and batch norm layers In the second network in order to improve the up sampling procedure and increase theaccuracy of segmentation we used new blocks composed of transposed convolution and cubic spline up sampling Onthe third architecture we used multiscale feature by adding Inception ResNet blocks with pre trained weights in each stepof encoder path In the final proposed architecture the atreus spatial pyramid pooling was added in the last layer of theencoder to use global features as well as local features The final model showed its superiority to other models in most ofthe evaluation criteria Key Words Saliency Salient Object Detection Deep Learning Convlutional Neural Network

ضيايي، نويد

تشخيص جسم برجسته در تصاوير با روش‌هاي يادگيري عميق

تشخيص برجستگي , جسم برجسته , يادگيري عميق , شبكه هاي عصبي پيچشي