From a single glance a human can understandthe world which is assumed to be a great accomplishment. All it takes for secondsto categorize certain environment or an object and its importance in the fieldof scene recognition.
One of the human capability is to learn and memorizeplaces by analyzing the world in some seconds. Our neural design constantlysaves different input even for a short span of time. It is common that peoplecan recognize scene they see such as a mosque, tomb or any fort.
The recent researchesshow that in 36 milliseconds processing time with 80% accuracy viewer canrecognize the scene. Now question raises how much an artificial machine willlearn before stretching out possession of human being and how we can recognizescene that fast and what information do we need?It is important to understand scene perception because in researchesit if found that a scene uses our certain knowledge that is connected to somescene category (eg, mosque, tomb, fort). Such knowledge forces that we must payattention and it may be helping in recognizing a particular object and determineswhat kind of information do we need to memorize from the scene.
Researchers onscene recognition travel a point between two system i.e cognition andperception and it is a problem that it considered to be a challenging task fora person working in artificial intelligence. This kind of researches can beused to design a system in artificial intelligence that have the ability torecognize the scene and its categoryA CNN is a class of deep learning and feedforward AI that has been successfully applied in scene recognition and visualimagery. Convolutional neural network directly learnfrom image data set and removing the need to manually feature extracting toproduce better recognition result. CNN are much like ordinary neural and CNNconsist of multilayers that is used to minimal preprocessing. CNN is also knownas shift variant. Conventional neural network appears to be successful in manyreal life studies and application such as image classification, facerecognition and much more. We will have to go back 2012 to understand itssuccess when Alex Krizhevskyused CNN to win 2012 imageNet event, resulting in error reduced from 26% to 15% A convolutional neural network may even have hundredlayers to discover different features in an image.
CNNarchitecture is design to take 2D structure advantage. It is achieved by localconnection and tied weight followed with certain pooling that result theinvariant translation feature. Additional advantage of conventional neuralnetwork is that it is trained easily and require few parameters that canconnect fully operational network with identical no. of hidden units.
CNN usesa variation of multilayerfollowed by full connected layer designedto require minimal preprocessing. CNNnetworks is designed similar to biological processes inwhich the connectivity pattern between neurons isinspired by the organization of the animal visual cortex. InCNN every neuron will be receiving input and carry out a dot product followedby non-linearity. Thereare three kinds of layer in CNN input, output, hidden.
The hidden layer containspooling, convolutional, fully connected and normalized layer