The vehicle passing through the road can be monitored continuously using the ALPR systems mounted on the roadside. The same system can be used in airport parking, mall parking, and hospital parking, etc. ALPR is implementedinmassive traffic zones for traffic rule enforcement. It can also be used for automatic toll collection on highways. Figure 1 shows a typical ALPR system. In such systems capturing of the image is very important as the image must include the licence plate of the vehicle for its identification.
Now the captured image is pre-processed to find the region containing the licence plate of the vehicle only. The numeral and alphabets of the licence plates are segmented, and their features are extracted. Finally, the features are given to classifier for the recognition of the characters[7, 8].
Overview of Literature
Various Methods are used for implementation of each stage of ALPR system has been proposed in available literature. A number of methods are proposed for licence plate extraction, and the main categories of existing methods are based on colour, gray-level, binary, and classifiers [9, 10]. Colour-based licence plate region extraction use colour criteria[11]. Colour based method for object recognition is very dominant, but the colour features of the vehicle image may vary because of environmental conditions, illumination condition, and the superiority of image capturing system. Such variations badly affect system performance. Colour combinations of Chinese licence plates are used to extract the licence plate region[11]. Fuzzy logic based extraction is proposed and fuzzy rules are based on a human’s colour sensitivity [12].The licence plate extraction based on gray-level is preferred and mostly used as compare to colour-based. Gray-level extraction methods perform superior but, they take longer processing time. It has been analysed from literature survey that feature extraction is easy even though; if high resolution images has been provided as an input for feature extraction. Although; a large number of techniques in consideration of quantization of vector transforms in line with features are used like Hough and Gabor transform, wavelet analysis has been used by various researchers in past. It has been seen that characters are being processed with hard pixels on plain background with low intensity. This kind of practices leads to blur images and somehow made it impossible to extract information with analytical techniques. However; a technique using adaptive algorithm has been used to extract information with different intensity patches with different refraction indexes [13]. Various Transforms are also used in conjunction with different techniques and Hough transform along with filter techniques and priority quantization like Gabor filters and vector quantization is also used to extract the licence plate region [14, 15]. Wavelet analysis of the image is used to find the image contrast and hence the licence plate region [16].
In classifier-based techniques, firstly the system is trained using the training sets of rectangular shape licence plate samples and then the trained system is used to find the coordinates of a new testing licence plate image. Such classifier based ALPR systems are slower if licence plate region is large. Classifiers such as genetic algorithm based [17, 18], genetic algorithms and ANN based [19], and fuzzy logic and ANNs based [20]techniques are used to anneal the various processes to segment features so that these can be easily extracted from a plate. The binary image processing methods which are simpler and faster are used for licence plate extraction. These methods are independent of shape size and colour of the licence plate region. These methods performs good and are based on edge statistics and mathematical morphology[9, 21]. The licence plate region in the input image contains the more number of edges as well as texture information. The benefit of using edge statistics and morphological operations is proposed to extract the licence plate region[22].The techniques based on contour tracking, horizontal and vertical projections and mathematical morphology are reported for the segmentation of the characters from the extracted licence plate region. Vertical projection based technique[11], adaptive morphological technique based on histogram equalization [23] and contour model based[24] techniques are proposed for separating the licence plate characters from each other.
The segmented characters are recognized at the last stage of ALPR system. The techniques proposed for the character recognition in the literature are based on ANN[25], Markov model [14], support vector machine[26] used, and pattern recognition[27, 28]. Proposed techniques by various researchers are good and inherent characterization but there is a scope of improvement with considering one of the properties of technique and merged with another to have a proposed structure in order to have better accuracy. In literature various techniques have been proposed and discussed but deep learning and random forest are performing very well in various applications. In this work, a novel technique has been proposed with conjunction of heuristic approach and Meta heuristic approach with characteristics of deep learning and random forest for ALPR system.
Overview of The Techniques Used
In the proposed technique, the input image of the vehicle is given as input to the ALPR system. As discussed earlier, input image undergoes various pre-processing steps includes conversion of one system to another comprises of binary structure, filtering, dilation, labeling, the histogram with an objective to differentiate a region from background region so that features should not be aligned with background features and resolution must be intact.[29, 30]. The region is taken for analyzed and separated out from a background and processed for its parametric evaluation in order to acess characters of plates for proper recognition. This process filters out unnecessary portion and disturbances in an image due to background overlapping. Next step in the proposed technique is to find the features of segmented characters and classify them. Soft computing techniques[31], are becoming more and more popular in designing real life applications such as home appliances, electronic gadgets real, consumer electronics. The reason of this revolution is to achieve reduce cost, automation and robustness. Genetic algorithm, fuzzy logic, ANN, machine learning, probabilistic reasoning are the most commonly used soft computing techniques. These soft computing based products have extraordinary ability to reason, learn from experience and make intelligent decisions. This all motive to implement of the ALPR system using soft computing techniques. The soft computing techniques used are explained as below.
Deep Learning
This technique has been applicable for various applications like text detection and recognition[32], tracking of object[33], pose estimation[34], classification of images, visual saliency detection[35], action recognition[36], scene labelling[37] and many more. There are number of deep learning models in the literature such as deep belief networks, sparse coding, restricted Boltzmann machine, auto encoder, and CNN. Out of the all CNN is not only widely used[38] but it has shown high performance. CNN was inspired by the visual cortex of animals[39]. In this paper, a CNN model has been proposed built for image classification [40].
A typical CNN consists of number of layers, first layer of the CNN is the input layer, and second layer of the CNN architecture is a convolutional layer. In CNN a feature map performs discrete convolution filtering operation with the input image [41]. After each convolution layer, there is a third layer rectified linear unit, is used to generate the nonlinearity in the system, and it is also known as ReLU layer. It is conventional to have ReLU layer after convolutional layer[42]. The main function of this layer to keep all attributes of function to be at origin so that overlapping can be avoided and input element is negative[43], i.e.,
$$f\left(x\right)=\left\{\begin{array}{c}x, x\ge 0\\ 0, x<0\end{array}\right.$$
1
Next layer in the architecture is pooling layer which is used to do downsampling. The fifth and sixth layer corresponds to connected mechanism with aligned structure and network configuration. In multi-class classification problem, output unit comprises of softmax activation function[44]:
$${\text{P(c}}_{\text{r}}\text{/x, θ) = }\frac{{\text{P(x, θ/c}}_{\text{r}}\text{)}{\text{P(c}}_{\text{r}}\text{)}}{\sum _{\text{j=1}}^{\text{k}}{\text{P(x, θ/c}}_{\text{j}}\text{)}{\text{P(c}}_{\text{j}}\text{)}}$$
$$=\frac{\text{e}\text{x}\text{p}\left({a}_{r}\right(x,\theta \left)\right)}{\sum _{\text{j}=1}^{\text{k}}\text{e}\text{x}\text{p}\left({a}_{j}\right(x,\theta \left)\right)}$$
2
Where\(0\le {\text{P}(\text{c}}_{\text{r}}/\text{x}, {\theta })\le 1\)and\(\sum _{\text{j}=1}^{\text{k}}{\text{P}(\text{c}}_{\text{j}}/\text{x}, {\theta })= 1\)
Moreover,\({\text{a}}_{\text{r}}=\text{l}\text{n}\left({\text{P}(\text{x}, {\theta }/\text{c}}_{\text{r}}\right)\text{P}\left({\text{c}}_{\text{r}}\right)\)
\({\text{P}(\text{x}, {\theta }/\text{c}}_{\text{r}})\) Signify conditional probability of the sample given class r, with P (cr) is the class prior probability. The seventh layer represents classification layer[26, 45].
Random Forest
It is a classifier with supervised mechanism with two different kinds of stages comprises of creation and prediction. The main structure element is combination of branches in an open ended modes where nodes are grouped together to form a network and termed as tree. Various Algorithms are being used to process a group of trees which are being assembled to form a forest with random attributes [46]. Algorithms has been categorized in two stages of codes and termed as creation and prediction Random Forest (RF) pseudo code
RF Creation Pseudo Code
RF creation pseudo code involves the following steps:
Step 1 Root has been constructed with best fit attributed of selected dataset
Step 2Data Set has been classified into different subset in consideration of same feature extraction. Data Set should be divided into sets with a fixed division in a pair of two.
Step 3 Processing has been carried out in a loop comprises of Step 1 and Step 2 till every subset only left with single element.
Step 4 Create a forest with repeating previous steps number of times with a value “n” in order to create a “n” number of trees to build a forest with random values