Improvement of Image Quality Using Convolutional Neural Networks Method

. Purpose: This desire for high resolution stems from two main application areas, namely improving pictorial information for human interpretation and assisting automatic machine perception in representing images or videos. Image resolution describes the detail contained in an image, the higher the resolution, the more detail there is. The resolution of a digital image can be classified into various types, namely pixel resolution, spatial resolution, temporal resolution, and radiometric resolution. In this context, we are interested in spatial resolution. Methods: Elements of a digital image consist of a collection of small images called pixels. Spatial resolution refers to the pixel density of an image and is measured in pixels per unit area. A quality digital image is determined by the size of the resolution it has. A low resolution or low-resolution is a drawback of a digital image because the information contained in the image means little compared to a high-resolution image. Result: Therefore, in this study, a digital image processing program was created in the form of Image Super-Resolution with the Convolutional Neural Network method to utilize low-resolution images to produce high-resolution images. With a fairly short training process, namely 6050 datasets with 100 CNN epochs, the average PSNR image is 5% higher. Novelty: Image quality can be improved by changing the parameters in the CNN method so that image quality can be improved.


INTRODUCTION
In an application of Digital Imaging, image (pictures or video) with a resolution higher is desirable for processing or analysis. The desire to get a resolution high is originated from two areas of applications principal namely an increase in the information display for the interpretation of man and help the perception of the machine automatically in representing images or videos. Image Resolution (resolution image) describes the detail contained in an image, the higher the resolution, the more a lot of detail there [1]. Digital images that are often encountered have low resolution, which means they store little information and make it difficult for image users, namely humans who see and computer machines that use these digital images for certain purposes. Image Super-Resolution (SR) is a technique to reconstruct the image of a resolution higher than the image resolution of the low was observed [2]. Because it required an application Super Resolution that can transform an image digitally -resolution low into an image with a resolution higher [3]. An image is a combination of points, lines, planes, and colors to create an imitation of an objectusually a physical object or a person. Imagery can be a tangible image (picture) are two -dimensional, like a painting, photograph, and tangible three-dimensional, like a statue [4], [5]. While digital images are images that are displayed on a computer as a set of digital values (resolution) in the form of pixels. The resolution can be low or high and it determines the quality of a digital image [6], [7]. In an application of Digital Imaging, an image (pictures or video) with a resolution higher is desirable for processing or analysis. The desire to get a resolution high originated from two areas of applications principal namely an increase in the information display for interpretation humans and helps automatic machine perception in representing images or videos. Image resolution describes the detail contained in an image, the higher the resolution, the more a lot of detail there [3].
The resolution of a digital image can be classified into various types, namely pixel resolution, spatial resolution, temporal resolution, and radiometric resolution. In this context, we are interested in spatial resolution [8]. Elements image digitally consists of a collection of images of small are called pixels. Spatial resolution refers to the pixel density in an image and in measuring the pixels per unit area [5].
In getting the image digitally, a spatial resolution that is higher obtained from how good the sensor that is used to capture the image of the good lens optics are used. Super-Resolution is a technique that is used to build a high image resolution (High-Resolution) or image resolution Low (Low-Resolution) which has been observed [1].

LITERATURE REVIEW Convolutional Neural Networks
Convolutional Neural Networks, or hereinafter referred to as CNNs ( Figure 1, Figure 2) are artificial neural networks whose neurons are represented as a matrix (n x n for an example of a 2-dimensional matrix), unlike neural networks in general which are vectors (n x 1) [3], [9], [10].

CNNs Architecture
The following is the architecture of a network of CNNs [11].

Input layers
Layer that contains a pair of images (x, y) where x is the original high-resolution image and y is the image x that has been processed into a low resolution. CNN's have a special layer called the convolutional layer. This layer functions as a filter when preprocessing the input neurons. The convolutional layer is represented in a matrix that has the same dimensions as the input layer. For example, if the input is 2D, then the Conv layer is also 2D [12]. The layer size is smaller than the input, which is commonly used as 3x3 for 2D. This layer aims to find out the features contained in an image such as vertical, horizontal edges, gradients, and others. To find out the various features, different filters are needed. In getting the output from this layer, the dot product is done between the convolutional layer and the input layer as shown in Figure 3.

ReLU Activation Function
The output layer of the convolutional layer will be subjected to the activation function ( Figure 3). The activation function used is the Rectified Linear Unit (ReLU) [1].
Where x if x > 0, and 0 otherwise.
If the element is negative then the value is set to 0, with no exponential, multiplication, or division operations. With such characteristics, the advantages of ReLU will appear when dealing with networks that have a lot of neurons so that it can reduce training and testing time significantly.

Pool Layers
Layer that is inserted after Conv layer + ReLU (figure 5) serves to reduce the size of the representation (matrix) to reduce the number of parameters (each neuron is a parameter) and computations in the neural network. There are two types of pooling, namely Max-pooling to choose the highest value or averagepooling to find the average [1]. Figure 5. Pooling layer process with max-pooling filter 2x2

Fully Connected Layer
The neurons in the screen FC are fully connected to all activations in the previous layer as shown in Figure 6.  To get an optimal network, the error rate or error must be as small as possible. Therefore we need a function called Loss Function [9]. The loss function that will be used is the mean squared error, or MSE, an error estimation method commonly used in signal processing [13]. MSE is a pixel-wise comparison with the formula 2 where N is the number of pixel channels in the image [6]. For example, for an image with a pixel size of RGB color. From MSE we can also calculate the peak signal-to-noise ratio, or PSNR, which is used to evaluate the existing neural network [1], [13].
where L is the dynamic range of each image [6]. Examples for color 8-bit integer, L = 255. Unit PSNR measured in decibels, the higher the value the network is getting good. The image that is identical having i value of MSE = 0 and PSNR = ∞ dB [14].

RESULT AND DISCUSSION Data Acquisition
In developing this Super Resolution application, there are 2 types of datasets used on the CNN network. The first dataset is used as training data and the second dataset is used as test data for 6050 HR images from the "Kou photo collection" which is free to download by visitors. The dataset consists of various types of images such as buildings, furniture, and food [13].

Figure 8. Sample Dataset Kou. Photo Collection
Set14 is a dataset that is often used to test a network [15]. This dataset consists of 14 images, but only 9 images will be used in the test.

Results
The model training was carried out with a configuration of x2 scaling and epoch 100. The following are the results of testing using the Set14 dataset which consists of 9 images, namely by comparing the MSE and PSNR values of LR images that have been processed with Lanczos filtering and CNN [16]. The results of calculations with upscaling 2x obtained the values presented in Table 1 Table 3 training changes using upscaling 4x, this process is used to test the changes in the value of the training image.  Table 4 training changes using 4x upscaling, this process is used to test the value changes in the training image using CNN. According to Table 2, Table 3, and Table 4, it can be seen that the CNN, MSE value is smaller than Lanczos in all tested images, while the PSNR value is higher. The comparative performance of upscaling with the CNN method can produce better image improvements than with the usual method. Learning by using lowresolution images, image quality can be improved by changing the parameters in the CNN method so that image quality can be improved. the result of the comparison is presented in Figure 9.