Abstract:
The objective of the proposed project is to research, explore and develop image
classification technique known as Convolution Neural Networks to classify and
detect road irregularity. This report explores detailed key ingredients used for
classification of road irregularity. Several building blocks of this technique; namely
convolution, dynamic feature mapping, non-linearity algorithm, back propagation,
inter and inter class classification along with several other layers will be researched
and discussed. Eventually a system will be developed using python programming
language and android development studio to fully utilize this technique.
In our proposed project uses the Convolution Neural Network technique to develop a
framework. The main merit of using this technique is its ruggedness to shifts and
distortion in the image, fewer memory requirements, better training and high
accuracy to name few. Different aspects of this architecture are discussed and
dynamic adjustment of weights and feature maps are used to attain better accuracy.
After several experiments and adjustments, a proper network will be defined. For
beginning a network structure consisting of 3 convolution layer and 2 fully
connected feed-forward neural network is used along with numerous dynamically
adjusting neurons, weights and feature maps.
The system initiates with the pre-processing of input video and will end with
classified labels of frames in video. A road map for future development and
conclusions are also included in the report.
