Convolutional Neural Networks for Image Processing with Applications in Mobile Robotics By, Sruthi...

Post on 25-Dec-2015

217 views 1 download

Transcript of Convolutional Neural Networks for Image Processing with Applications in Mobile Robotics By, Sruthi...

Convolutional Neural Networks for Image Processing with Applications in Mobile Robotics

By, Sruthi Moola

Convolution

•Convolution is a common image processing technique that changes the intensities of a pixel to reflect the intensities of the surrounding pixels. A common use of convolution is to create image filters

Convolutional neural network

• Type of feed-forward MLP. • Conv. Nets are inspired by biological

processes in visual cortex. • So it is used in image recognition and

handwritten recognition. • high performance in MNIST database. • Designed by Yann Lecun.

Convolutional neural network

•Architecture of applying neural networks

to 2-D arrays (usually images), based on spatially localized neural input.

•Technique of sharing weights or receptive fields.

Types of layers

• Convolutional layers ▫ Feature Map or filter▫ Shared weights

• Subsampling or Max pooling • Full connected layer (classification)

Convolutional layer

•Rectangular grid of neurons•Input from a rectangular section of

previous layer•Weights are same for each neuron•Image convolution of previous layer•Weights specify convolutional filters•Several grids in each layer, each grid

takes inputs from all layers using different filters

Max pooling layer

•Takes smaller blocks from convolutional layer

•Subsamples to produce single output from that block

•Several ways- average or maximum or learned linear combination of neurons

•Max pooling layers take maximum out of that block

Full-connected layer

•High level reasoning in NN•Takes all neurons from previous layer and

connects it to every single neuron it has•These are not spatially located (visualize

as one-dimensional)•Therefore, no convolutional layers after

fully connected layer

Convolutional neural network

•Network structure designed extracts relevant features, restricting neural weights of one layer to a local perceptive field in previous layer. Thus, feature map obtained in second layer

•The degree of shift and distortion variance is achieved by reducing the spatial resolution of the feature map

A six-layer convolutional neural network

Training

•Back propagation•In feature map, all neurons share the

same weight and bias, the number of parameters is smaller than in fully connected multilayer perceptron, leading to a reduction in gap

•Subsampling/pooling layers have one trainable weight and one trainable bias, so number of free parameters is even lower when compared

•Because of low number of free parameters, training of CNN requires far less computational effort than training multilayer perceptron

Applying CNN to real-world problems

•Image processing system on mobile robot•Task – detect and characterize cracks and

damage in sewer pipe walls.•Uses monochrome CCD camera•Task of CNN-

▫filter raw data▫Identify spatial location of cracks▫Enable characterization of length, width of

damage

Example input and target images for large cracks on a concrete pipe

•Horizontal feature – pipe joint•Significant challenges for filtering system

▫Differentiating between pipes and joints▫Accounting for shadows and lighting effects

•Training was conducted using standard weight updates rules and approximately 93% of pixels in the validation set were correctly classified

•Not every pixel was used for training▫Computational expense▫Low proportion of ‘crack’ to ‘clean’ training

samples tended to bias the network towards classifying all samples as ‘clean’

Example input and CNN output frames

Three frames including Crack present, no crack and pipe joint, and crack and joint together represent the data set.

Using a subsequent crack detection algorithm, the network ignores the presence of joints and attenuated lighting effects while enhancing the cracks.

Conclusion• Issues- Training uses bitmap type images that

results in over abundance of training sample points.

•Key characteristics like sharing weights are appropriate when input data is spatially distributed.

•Concept of CNNs and weight sharing not only reduces the need for computation, but also offers a satisfying degree of noise resistance and invariance to various forms of distortion.

•In the present system, CNNs are expected to achieve better results than standard feed forward tasks.

Thank you