Multi Layer Perception Of Back Propagation Computer Science Essay

Published: Last Edited:

This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.

This paper deals both online and offline hand drawn characters as well as license plate recognition. License plate recognition is an important part of intelligent transportation systems, and image feature extraction and recognition are the key processes. Firstly, multi layer perception is used to detect edges for the segmented characters [MLP] of the plate, and then the features of relative moment are extracted. Secondly, the features are fed into Back Propagation [BP] neural network for classification. The algorithm was tested with different natural vehicle images backgrounds and produce good results. The on-line and off-line recognition of hand-drawn character is an intelligent recognition system. Hand drawn recognition has been one of the active and challenging research areas in the field of image processing and pattern recognition. The Study of on-line identification we have established a neural network model of on-line and off-line hand-drawn graphical symbols identifying and have developed MLP and BP system with a good result of o character recognition. We can recognize and written characters for English alphabets without feature extraction using multilayer Feed Forward neural network. Back Propagation Algorithm mainly focus on compromises Training, Calculating Error, and Modifying Weights.


Neural network, Multi layer perception of Back propagation, Handwritten and license plate Character Recognition, Image processing.


One of the most creative applications of the Neural Network is the Character Recognition [CR] System. For human beings Recognizing characters, letters and symbols is not a big task. Even small child can be done it easily, but doing the same with machine is a difficult task. Machine simulation of human functions has been a very challenging research area since the advent of digital computers. The ultimate goal of designing a character recognition system with an accuracy rate of 100 % is quite illusionary because even human beings are not able to recognize every hand written text without any doubt. Character Recognition System is the base for many applications like businesses, post offices, banks, security systems, and also used in robotics .many of these applications we are using in our daily lives. It Cost effective and less time consuming. Simple example for CR is face or eye scan at the airport entrance, or training a robot to pick up and objector material, you are using the system of Character Recognition.

Simple method CR is Optical Character Recognition (OCR). OCR is widely used today application like airports, airline offices, post offices, banks, and businesses. Advanced version of OCR is used in the field - Handwritten Recognition. Another method Convolution Neural Networks (CNNs) used in handwritten recognition. Multilayer Perceptions (MLPs) is the best method for license plate recognition and Handwritten Recognition because MLP has many more free parameters than a CNN.

Vehicle license plate recognition system is an important research topic of the applications of intelligent transportation by using computer vision, image processing and pattern recognition technology. In license plate recognition system, there are four steps generally. Firstly input pictures, secondly do license plate location, thirdly do character segmentation and lastly do character recognition. We use MLP to obtain edge of vehicle license plate image, and then use relative moments to get some invariant image moment features, and then the feature vectors are put into BP neural network for classification.


Recognition algorithms are generally composed of several processing steps, such as license plate detection of extraction region, character segmentation of license plate and recognition of each character. The section provide a brief explanation for the researchers involved in license plate identification and recognition, regardless of particular application areas .For any character recognition or license plate system there are three major stages, as shown in the figure [1] below:

• License plate detection

• Character segmentation

• Training and Character recognition

FIG: - 1

1. License plate detection:

As far as extraction of the plate region is concerned, techniques based upon combinations of edge and color based. This cannot depend on the license-plate boundary edge; it can detect unclear license-plate boundary of image and also implemented simply and fast. A disadvantage is that edge-based methods alone can hardly be applied to complex images.

Color or gray-scale based processing methods are proposed in the literature for license plate location. Crucial to the success of color (or gray level)-based method is color (gray level) segmentation stage. Since these methods are generally color-based, they fail at detecting various license plates with varying colors. Color processing shows better performance. But these methods are sensitive to the license plate color and brightness and need longer processing time from the conventional color-based methods.

2. Character segmentation:

The license plate candidates determined in the previous stage are examined in the license number identification phase. There are two major tasks involved in the identification phase, character segmentation and character recognition. A number of techniques to segment each character after localizing the plate in the image have also been developed. An algorithm based on the histogram automatically detects fragments and merges these fragments before segmenting the fragmented characters the results are very promising and encouraging indicating that the method could be used for character segmentation in plates with not easily distinguishable characters during off-line operation. But, since the algorithm is computationally complex, it cannot be proposed for real time license plate recognition.

3. Character recognition:

The Character Recognition can be done by MATLAB or VHDL. The matrixes of each alphabet letter can be created along with the network structure. Mainly we need to understand how to pull the Binary Input Code from the matrix, and how to produce the Binary Output Code, which is ultimately produced by computer.

Character Matrixes:

Character matrix contains array of white and black pixels; these are represented by black and white. 0 is white and 1 is black. These fonts are imaginable and created manually by the user; in addition, same alphabet can be used in multiple fonts may even used under separate training sessions. It is just like a process of Image digitization: The process of digitization is important for neural networks. In this process the input image is sampled into binary window which forms the input to the recognition system. The sample of this process is shown in the figure [2] below

0 0 1 1 0 0

0 0 1 1 0 0

0 1 1 1 1 0

0 1 0 0 1 0

0 1 1 1 1 0

1 1 0 0 1 1

1 0 0 0 0 1

1 0 0 0 0 1


Creating a Character Matrix:

The ability to recognize characters with machine, we must first create those characters. Creating a matrix to small and all the letters may not be able to be created with in size, especially if you are trying for two or more different fonts that matrix size is big. Some problems occur in computation like memory cannot able to handle all the neurons in the hidden layer needed to efficient and accurately process the information. We can simply reduce number of neurons, but this will increase the chance for error. A large matrix size of M-N was created. These are done by fast computer performance .different forms of digit as shown in figure [3 4] below:


First Form

Second Form























FIG:-4 Different forms of character


The multilayer perception neural network is built up of simple components. In the beginning, we will describe a single input neuron which will then be extended to multiple inputs. Next, we will stack these neurons together to produce layers .Finally, the layers are cascaded together to form the network.

Single-input neuron:

A single-input neuron is shown in Fig. 5. The scalar input p is multiplied by the scalar weight W to form W p, one of the terms that are sent to the summer. The other input, 1, is multiplied by a bias b and then passed to the summer. The summer output n often referred to as the net input, goes into a transfer function f which produces the scalar neuron output.

Multiple-input neuron:

A neuron has more than one input. A neuron with R inputs is shown in Fig. 6. The inputs are x1, x2.. xn. and W1, 1 W1, 2,W1, R are the weight matrix W.

The neuron has a bias b, which is summed with the weight inputs to form the net input N:

N =W1, 1p1 +W1, 2p2 +... +W1, Np R + b

This expression can be written in matrix form as:

N = W p + b (4)

Where the matrix W for an single neuron has only one row. Now the neuron output can be written as:

A = f (W x+ b)


We are mostly using common neural network architecture called the multilayer perception [MLPs]. Mainly we are focusing on neural network as function approximations. As shown in Fig. 5, we have some unknown function that we wish to approximate. We want to adjust the parameters of the network so that it will produce the same response as the unknown function, if the same input is applied to both systems.


MLP neural networks consist if units arranged in layers .Each layer is composed of nodes and in the fully connected network considered here each node connects to every node in subsequent layers. MLP have minimum of three layers consisting of an input layer, more or one hidden layers and one output layer. The input layer distributes the inputs to subsequent layers. Input nodes have liner activation functions and no thresholds. Each hidden unit node and each output node have thresholds associated with them in addition to the weights. The hidden unit nodes have nonlinear activation functions and the outputs have linear activation functions. Hence, each signal feeding into anode in a subsequent layer has the original input multiplied by a weight with a threshold added and then The Multilayer Perception (MLP) is the most frequently used neural network due to its ability to model non-linear systems and establish non-linear decision boundaries in classification or prediction problems.

Back propagation of multi layer perception:

Neural network model has various types, such as the BP network, Fuzzy neural networks an etc. According to the characteristics of the character recognition we have adopted a widely used BP neural network model.

The BP neural network model topology includes input layer, hidden layer and output layer. Input of images are normally carried from the input layer, hidden layer extract the perdition outputs, and then passes to the output layer, and finally is put out at the output. This is the way a signal is transmitted. In the process of signal transmission the right value of the network is fixed, and the state of neurons of each layer only influence the state of neurons of the layer below. If the output of the output layer has relatively large deviations from the desired output, then the error signals will be put back. The error signal is the difference between the actual output of network and desired output. The backward transmission of error signal is transmission of the signal from the output to the input. In this process, the network value is adjusted by the error feedback.

Back Propagation (BP) neural network is trained by error back propagation algorithm of multilayer feed-forward networks. It is currently the most widely used neural network model. BP neural network can learn and store a lot of input-output model mappings, without revealing in advance the mathematical equations describing the mappings. Its learning rule is to use the gradient descent method, and by back-propagation network to continuously adjust the weights and thresholds. As a result, the minimum of the squared error of the network is gained. We design two BP neural networks: one is for Character classifying, and the other is for numbers and letters identifying. The input layer accepts the all image character, so there are many input nodes. We use binary code as the output of the network. And the number of output nodes is M-N. For characters classifying we used number plates including some special license plates like military.

The parameters of BP neural networks are Input values [0, 1] used for activation function of log of Sigmoid, three layers, the hidden layer uses 36 neurons based on experience. These 36 neurons are 26 characters [A, B…Z] and 10 numbers [0, 1…9]. We collected some license plate images from the internet for experiments. The image database is divided into two parts: one is 100 images as training data, from which selecting 80 numbers (each number has 10 samples), 100 letters, 100 characters; the other is 20 images as test data. Firstly, the feature vectors of different types of training samples are put into the corresponding neural network. Secondly, after the networks are trained, the feature vectors of test samples are put into the trained networks. Figure 2 shows one of recognition test results. And the final recognition rate is about 95%

BP neural network model overview:

The constant adjustment of the network value makes it possible for output of the network to reach the expected value. Fig.1 is a three-layer BP neural network model of the structure diagram. Fig. of representative training samples and a highly efficient and stable, fast convergence of learning. The process of BP neural network algorithm learning and training is shown in Fig.2.

Fig.3 Flow chart of BP network identification

BP neural network model contain three layers to determine the number of neurons, which is the key to BP neural network design. The first step of BP network applications is to use the known training samples to train the BP network. Here, the number of BP network input layer nodes is the number of characteristics of the output dimension of image after pretreatment. For feature extraction we use the pixel-by-feature extraction method that is to use the pixel value of each point as the feature. So for every one input sample image contain huge amount of characteristics can be compressed to 8- 6 characteristics by line and column projection. No rules are taken for hidden layer nodes. Generally, if the number of hidden layer neurons is larger, then the BP network is more precise and the training time is longer. If the number of hidden layer neurons is changed then the training time will increase the recognition rate is not greatly improved.



Result are taken by using different images. Two types of images have been used in is black and white image and other one is colured image. Those two images are processed nicely and producing good result. These resuults are very accurate for vehicle plate recognization. Fig 4 shows the output result of an both types of images.


Multi layer perception and back propagation are applied to license plate recognition. First multi layer perception is used to detect edges of the pre-processed license plate images, then many invariant moments of shape features are extracted, and then the feature vectors are put into BP neural network for classification. It achieves fast and efficient license plate character recognition. Experiments show that the method is simple and has high accuracy. This method can be used in field of vehicle control, road pricing and parking management. Mainly this is used for tool plaza.