Semantic segmentation in flaw detection

Abstract:

The paper presents a review of study on detection and classification of defects using semantic image segmentation based on convolutional neural networks. Taking into account the revealed general features of flaw detection tasks of various industries related to the lack of a large marked data set and the need to detect defects of small sizes. The convolutional neural network of the u-net architecture was chosen as the basis for the decision support system. Testing of this architecture on several datasets yielded positive results regardless of the area of use.

Introduction

Flaw detection requires a wide range of appropriate description methods and means of control of materials and products. Despite the diversity of physical principles used in the basis, many technical solutions consist of two stages: image acquisition and its analysis in order to detect any deviation from the specific characteristic. In some cases, it is not enough only to detect deviations; it is important to detect and identify structural defects. This study [1] classifies defects found in the structures of saturated metallic composite castings. The proposed procedure for the detection and identification of structural defects of saturated metallic composite castings gravimetry, ultrasonic and X-ray, tomography, macroscopic tests, microscopic examination using light or scanning electron microscope, then its classification is carried out using the obtained image. The implementation of lightweight constructions based on composite materials requires the determination of the minimum damage size to still ensure safe conditions have to be identified and established in production as well as during the application, a review is presented in [2]. To assess the quality of welded joints, where one of the prospective flaw detections has a traditionally significant role, the magneto-optical eddy current (MOEC) method, in which surface, subsurface and fatigue defects can be recorded in products from both magnetic and nonmagnetic metals, as well as flaw detection of welds are considered in [3]. The measured impact duration can be used to obtain a “scan image” in various materials (especially honeycomb sandwich composites) [4]. The absence of visible defects along the route is an important condition for the safe movement of all modes of transport. Timely identification of defects and understanding of the operating conditions of materials and structures allow us to assess the time of their fault tolerance.

Читать статью