ABSTRACT

Crack detection is one of the important contents of the nuclear fuel pellet quality inspection. Aiming at the problem of high crack false detection rate caused by low image contrast of nuclear fuel pellets, complex image background and fine cracks on the pellet surface, a Weighted Object Variance (WOV) threshold crack detection method guided by SE-CrackNet convolutional neural network is proposed. The method first uses the sliding window scanning technology and SE-CrackNet network to locate the crack regions in the pellet image, and then uses the WOV threshold method to extract the cracks to achieve accurate identification of the cracks on the surface of the nuclear fuel pellet. The pixel-level F1-measure of the method is about 92%, which can accurately identify cracks on the surface of nuclear fuel pellets, greatly reduce the crack false detection rate, meet the real-time quality inspection requirements of nuclear fuel pellet production lines, and vastly improve the performance of traditional machine vision inspection systems. At the same time, the method can be extended to the quality inspection of other industrial products.

摘要

裂纹检测是核燃料芯块质量检验的重要内容之一。针对核燃料芯块图像对比度低、图像背景复杂、芯块表面细小裂纹导致裂纹误检率高的问题，一种基于SE-CrackNet卷积神经网络的加权对象方差（WOV）阈值裂纹检测方法被提议。该方法首先利用滑动窗口扫描技术和SE-CrackNet网络定位芯块图像中的裂纹区域，然后利用WOV阈值法提取裂纹，实现对核燃料芯块表面裂纹的准确识别。 . 该方法像素级F1-measure约为92%，可以准确识别核燃料芯块表面裂纹，大幅降低裂纹误检率，满足核燃料芯块生产线实时质量检测要求，大幅提升传统机器视觉检测系统性能。同时，该方法可以推广到其他工业产品的质量检测。