Marble surface anomaly detection using autoencoder architecture

Marble is a type of metamorphic rock composed of limestone that has undergone structural and chemical transformations due to high pressure and heat in the earth's crust. Marble has a smooth texture and often contains beautiful patterns or veins. The stone usually comes in a variety of colors, such as white, gray, pink and green, depending on the minerals contained in the rock it comes from. Marble has a unique aesthetic and elegance that makes it highly valued in art, architecture and decoration. It is widely used for a variety of applications, including floors, walls, sculptures, tables and countertops. Its relatively high hardness makes it wear resistant and durable. As its use in various applications and industries advances each year, the number of marble users will continue to fluctuate depending on trends and economic conditions in various regions and demand for special projects. As many consumers buy marble for its beautiful patterns and long-lasting durability, monitoring surface defects in marble factories and maintaining the quality of the marble is crucial to increase consumer confidence. In addition, early prevention of problems caused by faulty production prevents financial losses that might otherwise occur. Therefore, it is very important to detect surface defects in the manufacturing industry. Traditional surface defect inspection usually involves visual inspection by humans or the use of simple equipment such as microscopes. In a manufacturing context, manual inspection by operators to detect defects such as scratches, cracks or surface unevenness is still widely used. While these methods can be effective, they tend to be time-consuming, depend on operator expertise and may not provide a consistent level of accuracy. Automated inspection systems are becoming a progressive solution to improve the efficiency and accuracy of surface defect inspection. By integrating technologies such as image recognition, artificial intelligence and machine vision, automated inspection systems can automatically detect and classify surface defects with high accuracy. This reduces reliance on the human factor and increases efficiency in production environments. An autoencoder is a type of neural network that automatically learns the data representation and in this context is used to reconstruct normal data with high accuracy. When the model sees a large number of normal data samples, it can produce a good reconstruction. The autoencoder model also uses a loss function to measure how well the model reconstructs the data. When the model encounters abnormal data, the loss function tends to increase due to the difficulty in reconstructing unusual data. A threshold value can then be set using the value of the loss function. Data with a loss value that exceeds the predefined threshold value can be considered anomalous. This allows the autoencoder to automatically detect patterns that do not correspond to normal data. To get good results in this study we conducted several trials using the same dataset but with different autoencoder architecture models and taking different threshold values. in the first model we created an autoencoder architecture with four convolution layers and four pooling layers using MaxPooling2D in the encoder section and also four convolution layers and four pooling layers using UpSampling2D in the decoder section. Then we provide 300 image data as training data and 50 image data as validation data to train the model that has been created. In addition, hyper parameters are also used in the form of Adam as an optimizer, 50 number of batch sizes and the loss function 'mean_squared_error'. then with the model that has been trained, a function is made to find the average value of the reconstruction error and the standard deviation value of the input data and the two values are summed up and then used as a threshold as a parameter in detecting anomalies. Then for the second model we created an autoencoder architecture with three convolution layers and three pooling layers using the MaxPooling2D function in the encoder section and also in the decoder section we created three convolution layers and three pooling layers using the UpSampling2D function. After that, with the model that has been trained, we look for the threshold value using the previously created function as a parameter to detect anomalies. Then in the third experiment we used the same model in the second experiment, namely three convolution layers and three pooling layers in the encoder section as well as for the decoder layer made three convolution layers and three pooling layers. But in this experiment we took the threshold value directly from the average reconstruction error value of the trained model and used it as a parameter in detecting anomalies. After the parameters are obtained, a function is created to detect anomalies on the marble surface using the "marble surface anomaly detection" dataset obtained from one of the dataset provider websites. Fifty images of normal marble surfaces and fifty images of marble surfaces with defects (points, joints, cracks) were used to test the anomaly detection function. For the first experiment, the accuracy result is 74%, for the second experiment, the accuracy result is 84% and finally in the third experiment, the result is 88%. The conclusions of this study include the effectiveness of this method in detecting anomalies, the advantages of the autoencoder architecture compared to other methods, and the potential practical applications of these findings in various fields. By utilizing the ability of the autoencoder to reconstruct data, anomaly detection can be performed by comparing the reconstructed results with the original data. The main advantage of this approach lies in its ability to address the problem of anomaly detection without the need for class labels.