A CNN Model Based on Spatial Attention Modules for Casting Type Classification on Pseudo-color Digital Radiography Images
Chuanfei Hu, Yongxiong Wang, Kai Chen, Yulong Qin, Hang Shao, Jingkun Wang · 2019
In the casting manufacturing, the dependable automation of classifying casting types on digital radiography (DR) images is a crucial technology to automate the downstream tasks, such as defects detection. Generally, DR images are constructed by single gray-scale information, which constricts feature representations of castings on the DR images. Meanwhile, such the complicated background of DR image acquisition is an undesirable issue for the classification performance. Recently, the neural network, especially convolutional neural network (CNN), has great successes in the conventional tasks. However, CNNs are unable to be applied to the industrial tasks without any additional adjustments. In this paper, an improved pseudo-color processing is first proposed to enhance the original DR images. Then, the query DR images are colorized using pseudo-color processing. Ultimately, we design a novel CNN model based on spatial attention modules which can increase the model capability of focusing on the valid regions of castings. The experiments demonstrate the proposed model is capable of recognizing the casting types precisely on pseudo-color DR images. The superior performance also shows the practical value of these methods in casting type classification.