Thresholding-accelerated convolutional neural network for aeroengine turbine blade segmentation

Abstract

Turbine blades can only be detected nondestructively and precisely using industrial computed tomography (CT). The accuracy of CT image segmentation, which is a key step in CT detection, significantly affects CT detection precision. However, CT images of turbine blades cannot be segmented precisely because they contain inhomogeneity, noise, artifacts, and low-contrast areas. In addition, it is crucial to consider both efficiency and precision in industrial applications. Unfortunately, large-size networks often tend to be inefficient, while lightweight networks lack the necessary accuracy. To address these issues, we propose an innovative approach utilizing an ultra-lightweight neural network that fuses a threshold-based segmentation algorithm with a convolutional neural network for acceleration. The integrated thresholding algorithm was developed through the convolution of the Phansakar method. The convolutional neural network was designed with a simplified multichannel convolution operation to minimize computational cost. Notably, the proposed neural network utilizes only 0.11 M learnable parameters, which is significantly less than the number utilized by most classical deep neural networks. Experimental results for CFM56-7BE and Pratt & Whitney F100 aeroengine hollow turbine blade CT images demonstrate that compared to state-of-the-art algorithms, our proposed network can achieve 0.44 % mean intersection over union and 0.23 % mean Dive similarity coefficient improvements on testing datasets with significantly reduced computational cost. Therefore, the proposed method has high practical application significance and can achieve real-time detection.