基于单步目标识别架构的轻量级裂纹图像自动识别算法

Crack is one of the important factor affecting the safety of structures such as engines.Widely used manual visual test is inefficient and unreliable.Reliable automatic image recognition method for cracks are needed to promote the detecting efficiency.In this paper, based on convolutional neural network (CNN) and SSD model of monocular recognition.The network topology structure for crack recognition was improved, and a lightweight crack image recognition algorithm was developed.Considering the composite materials used in rocket structure, it was difficult to obtain the crack of propellant grain because of the migration learning characteristic of deep neural network. In this paper, CNN training was carried out based on the standard test pieces of metal materials was used in rocket as an application case.The algorithm expands the spatial resolution used in model training and reduces the neural network topology-the compression of the recognition task image, so that it can effectively capture small-scale graphic details.Covering overlap algorithm which is more suitable for special multi-scale separable graphics of crack, is proposed in training sample judgement.Through the crack image recognition experiment, the effectiveness of the proposed algorithm and above strategy has been verified.The proposed algorithm can increase the efficiency of image recognition by 51.13% compare with the typical multi-step target recognition method faster R-CNN.In addition, the hardware requirements are significantly reduced for model deployment while maintaining similar accuracy to faster R-CNN recognition.