A Good Performance of Convolutional Neural Network Based on AlexNet in Domestic Indonesian Car Types Classification

. Classification of car vehicle types has been carried out using CNN. There are weaknesses in the CNN algorithm so that it can be continued in the research we propose. This study aims to improve the previous accuracy by using the Alexnet architecture. To improve the results of the data set used we use threshold and brightness adjustment and data augmentation techniques for Reflection, Rotation, and Translation. Sample images with a resolution of 227x227x3 totaling 840 images used to represent 8 class types of cars, including Avanza, Fortuner, Freed, Inova, Pajero, Terios, Xenia, and Xpander. Alexnet with 10 epochs consisting of a total of 760 iterations, and validation is carried out every 30 iterations, the test results show that the use of the "sgdm" optimization function achieves a training accuracy of 99.74%, while the use of the "adam" optimization function produces an accuracy of 96.85%. This experiment shows the model's ability to classify the types of trainers after a success rate of 100%.


Introduction
With the development of technology, human life has been greatly helped by the presence of machine vision.Recognition of vehicle types has always been a hot topic in the field of object detection, especially cars [1].In 2023, cars will become increasingly popular and have an important role in supporting the development of the automotive and technology industries.Apart from that, it is impossible to find information directly from one person to another to find out the type of car used.This type of car recognition system can help identify and analyze car usage trends [2] and consumer preferences which can help classify car types based on market demand and it is hoped that the results of this research can help car manufacturers plan their marketing and products.strategy development.Data taken from Gaikindo [3] regarding the top 10 retail car sales in 2023, Avanza and its sibling Veloz appeared as Toyota's best-selling vehicles at the 2023 Gaikindo Indonesia International Auto Show (GIIAS) which took place from 10-20 August 2023.These two cars topped sales with a total difference in orders of more than 1,000 units through Vehicle Purchase Orders (VPO) of 1,213 units (20.9%).
Proving that in Indonesia the majority of people choose Toyota for their cars [4].With this data, a car model recognition system is needed to classify the types of car vehicles in Indonesia using digital image processing.The Convolutional Neural Network (CNN) algorithm is an important algorithm that helps many machine vision tasks in image classification, semantic segmentation, and object detection and is often used for car model recognition [5].Previous research has utilized CNNs to recognize distinctive features in car images, thereby enabling more precise classification.Some approaches have used data augmentation techniques, such as rotation, reflection, and brightness adjustment, to increase variation in the training data set, while others have considered various CNN architectures, such as AlexNet.Nonetheless, there is still room for further development in improving the accuracy and efficiency of car type classification.In this context, this paper contributes by introducing a method that combines data augmentation techniques with the AlexNet architecture, which is expected to open up new opportunities to significantly improve the accuracy of car type classification [6].
Data augmentation is a technique commonly used in machine learning to increase variety in a training data set by making small variations on existing images, such as rotating, mirroring, or panning them [7].This helps machine learning models become more robust and able to recognize a wider variety of patterns.This study investigates the use of the AlexNet architecture on the CNN network [8].The main focus of this research is centered on two main problem aspects as follows: a. Found a method of implementing data augmentation on the AlexNet architecture which aims to identify car models.b.Make a comparison between the Alexnet architecture and the implementation of data augmentation techniques such as Reflection, Rotation, and Translation.This research begins with the collection of a dataset that includes various types of cars from various points of view.This car-type dataset is then pre-processed to ensure image quality and consistency.CNN Alexnet is proposed as the main model for performing automatic feature extraction from car images.There are weaknesses in the CNN algorithm [9] that can be continued in the research we propose.This study aims to improve the previous accuracy by using the Alexnet architecture.The research that has been built focuses on the use of data augmentation techniques, such as reflection, rotation, and translation, along with image segmentation using thresholding, brightness adjustment, and CNN Alexnet.This approach can provide greater variation in the training dataset, which in turn can improve classification performance and help increase the accuracy of car type recognition and produce good accuracy.After this training was completed, it resulted in an accuracy of 99.74% and proved to help recognize the type of car.

Data Collection
The dataset used in this training is a private dataset taken in real time on roads around Semarang by taking pictures of passing vehicles.Car images with dimensions of 227x227 pixels total 840 data sets of car types which are categorized into two different data sets, namely training data and test data which includes 8 classes of car type images consisting of 108 Avanza, 100 Fortuner, 100 Freed, 125 Inova, 108 Pajero, 102 Terios, 103 Xenia, 100 Xpander.The training data is used to understand the model for each type of car that needs to be categorized, while the test data is used to evaluate the effectiveness of the proposed method in accurately identifying the desired car type category.

Preprocessing
The initial data processing (preprocessing) of the vehicle image dataset is carried out to obtain a better image so that the results of the process to be carried out next have better results.The following steps are performed for pre-proccessing 1. Cropping the RGB image to adjust the center of the object so that the image on the car dataset can be more focused and uniform.This cutting process is done manually.2. Resize the RGB image to 227x227 pixels.This is necessary so that the image of the car can be processed in practice.The car model dataset must be resized according to the input dimensions specified by the CNN AlexNet network, namely 227x227x3.Figure 3 shows the three RGB color channels namely Red (R), Green (G), and Blue (B).The image resizing process is done automatically using Matlab R2022a.
The preprocessing process of car image segmentation is carried out with 2 techniques, namely Thresholding brightness adjustment to help clarify objects [10] in the image so that the model can focus more on recognizing relevant features and classifying car types better.Divide the collected data into two parts: training data and test data [11].Each vehicle is labeled to distinguish these two data, for each data is divided into 90% training data and 10% test data.In car classification with CNN AlexNet, a rich and comprehensive feature extraction process occurs.This algorithm identifies complex visual patterns and statistically significant features that characterize car types.Feature extraction includes Mean, Variance, Kurtosis, Minimum Value, Maximum Value, Standard Deviation, Entropy, Skewness [12] as in (1) until (8).Calculations in finding the Mean value use the formula (1), where  is the number of elements in the dataset,   is the value of the to- in the dataset.Calculations in finding the standard deviation value use the formula (2), where  is the number of elements in the dataset,   is the value of the to- dalam dataset,  is the average value of the dataset.Calculations in finding the Kurtosis value use the formula (3), where  is the number of elements in the dataset and t,   is the value of the to- dalam dataset,  is the average value of the dataset,  dataset standard deviation.Calculations in finding the Variance value use the formula (4), where  is the number of elements in the dataset,   is the value of the to- dalam dataset,  is the average value of the dataset.Calculations in finding the Skewness value use the formula (5), where  is the amount of data in the sample,   adalah nilai data pada posisi to-,  ̅ is the average of the data,  is is the standard deviation of the data.Calculations in finding the Entropy value use the formula (6), where  is the number of possible distinct values in the data set,   is the intensity value of the pixel value at position to-,   is the probability of occurrence of the value x_i in the dataset.Calculations in finding the Minimum value use the formula (7), where  is the number of elements in the dataset,   is the value of the to- in the dataset.Calculations in finding the Maximum value use the formula (8), where  is the number of elements in the dataset,   is the value of the to- in the dataset. 0230302-04 ℎ  = ∑ (  − ) 4    =1 ×  4   (3) This [-10, 10] range means the image can be shifted randomly between -10 pixels to 10 pixels in the vertical direction.

Classification with CNN Alexnet
The AlexNet architecture approach is used for the classification of car types in the context of visual pattern recognition.In this task, a collection of input images with a resolution of 227 x 227 pixels and 3 color channels (RGB) is needed and pays attention to the structure of the alexnet architecture and the parameters used.AlexNet has layers such as convolution, ReLU, and pooling.Convolution layers take features from the image through filters.ReLU adds non-linearity, while pooling reduces the dimensionality of features.This process is repeated with subsequent layers to recognize increasingly complex features.Finally, a fully connected layer connects these features to classification classes via a softmax layer.In the image above, Alexnet connects Alexnet features to the car-class imagery used for the training and testing process of this study.Accuracy is a measure used to measure the extent to which a model or system can provide correct results or conform to expected data.In the context of pattern recognition or classification, accuracy aims to measure how well the model can correctly identify and classify data based on the labels or categories it should be.Based on the image above, the next step after preprocessing and segmentation is classification using data augmentation techniques and CNN to classify car types.Data augmentation techniques to increase variation in the training process [14] [15].The classification process is carried out with an enhanced AlexNet using several variables that are hyperparameters as a result; the optimation function is based on sgdm and adam; the hardware resource uses a single CPU; MaxEpochs is 10; Validation frequency is 30; and the MiniBatch is 10.In this stage, the classification process succeeded in issuing the appropriate results from the input image class.In running this dataset training using the Matlab2022a tool and for training options.Several features a more concise numerical form that reflects the important characteristics of the image.Table 1 shows accuracy and validation visualize the graph and plot the accuracy of the number of iterations.After undergoing training, the model using the 'sgdm' optimization function obtained 99.74% cursive results in Figure 6 and Figure 7.An experiment using the 'Adam' optimization function yielded an accuracy of 96.85%.3, the classification of car type classes using CNN Alexnet and data augmentation techniques produces excellent accuracy.All samples that passed the test were classified correctly, resulting in a training accuracy of 99.74%.In 18 experimental attempts per class with 2 samples each, the test results consistently indicated very high accuracy, reaching 100%.In this research, we also conducted experiments using the Deep Convulsive Neural Network (DCNN) method with the same number of datasets and training options but did not use the augmentation technique and AlexNet architecture, resulting in an accuracy of 35.71% with 'Adam' and 22.62% with 'sgdm'.With these results, it can be proven that the CNN method processed through data augmentation techniques using the Alexnet architecture can produce better accuracy.

Conclusion
This study uses the AlexNet architecture to improve the accuracy of car vehicle classification by applying data augmentation techniques.The results of the research achieved a training accuracy of 99.74%.The main finding of this research is to provide valuable insights regarding the use of machine vision technology in identifying car types, which is relevant to automotive industry trends and consumer analysis.Gaikindo's data showing the popularity of Toyota cars in Indonesia reinforces the value of this research contribution for car manufacturers, enabling them to understand market trends and preferences in greater depth.In addition, the implementation of a car model recognition system in the technology field can help the automotive industry develop better solutions for identifying and analyzing trends in 0230302-08 car usage.The implications of this research can be used as a basis for further development in optimizing the use of machine vision technology in the automotive industry and related sectors.

) 2 . 3 .Figure 2 .Figure 3 .
Dataset and Data Augmentation Example of Raw Dataset: (a) Terios, (b) Fortuner, (c) Avanza Example of the results of the Data Augmentation process: (a) original image, (b) brightness adjustment, (c) flip horizontal, (d) rotation using 30 0The parameters used in this study are: 1. 'RandXReflection', true: This option indicates whether to perform a random horizontal mirror (create a mirrored version of the image).With the value true, the image will be mirrored horizontally randomly.0230302-052.'RandRotation',[-10, 10]: This option indicates the range of random rotation to be applied to the image.[-10,10]  range means that the image can be randomly rotated between -10 degrees to 10 degrees.3. 'RandXTranslation',[-10, 10]: This option shows a random range of horizontal shifts applied to the image.This[-10, 10]  range means the image can be shifted randomly between -10 pixels to 10 pixels in the horizontal direction.4. 'RandYTranslation',[-10, 10]: This option shows the random vertical shift range applied to the image.

Table 1 .
Extraction Features Figure 6.Training accuracy progress

Table 3 .
Classification Of Car Types