Machine vision based detection of object surface defects on industrial applications

Although there are many disadvantages compared to the controls made with machines, controls made with the human eye continue to be used frequently in the industry. Machines can perform quality control with high performance rates on object surfaces that are easily perceptible, whose surface texture is homogeneous and uniform, where the flawless surface is clearly separated from the defect, which does not reflect irregularly in shape, and where sophisticated algorithms and high calculations are not required for the detection of defects. However, in the quality control of objects with mixed and irregular surface texture, reflecting light due to their material properties, and irregular reflection due to their shape, computer vision-based systems are insufficient for detecting defects compared to human abilities. This is especially true in images where imperfections are difficult to distinguish from the surface texture. Another problem is that the defect detection and separation cannot be done frequently within the requested time limits of the developed methods. In this study, a system that performs end-to-end quality control of military cartridge cases is proposed. First of all, a suitable design for the mechanical, electronic, communication and lighting of the system was made. Images were taken from the designed system and a data set was created for the hive surfaces. In order to create the hive data set, the raw images taken from the cameras were processed with the new methods developed and the hive surface was separated at the desired position for further processing. Military cartridge cases are metallic, not flat in shape, have an irregular surface texture and are highly reflective. These features make surface defect detection of hives a challenging problem. Various methods have been proposed that can operate at high accuracy rates and low speeds to detect manufacturing defects based on computer vision and distinguish them from the unique texture of the hive surface. High-performance methods have been proposed for deep neural networks developed both by image processing techniques and by being influenced by human ability, and new contributions have been made to the literature. Experimental results were carried out using real images taken from the prototype and successful results were obtained.