Multiscale Attention-Based Instance Segmentation for Measuring Crystals With Large Size Variation

Quantitative measurement of crystals in high-resolution images allows for\nimportant insights into underlying material characteristics. Deep learning has\nshown great progress in vision-based automatic crystal size measurement, but\ncurrent instance segmentation methods reach their limits with images that have\nlarge variation in crystal size or hard to detect crystal boundaries. Even\nsmall image segmentation errors, such as incorrectly fused or separated\nsegments, can significantly lower the accuracy of the measured results. Instead\nof improving the existing pixel-wise boundary segmentation methods, we propose\nto use an instance-based segmentation method, which gives more robust\nsegmentation results to improve measurement accuracy. Our novel method enhances\nflow maps with a size-aware multi-scale attention module. The attention module\nadaptively fuses information from multiple scales and focuses on the most\nrelevant scale for each segmented image area. We demonstrate that our proposed\nattention fusion strategy outperforms state-of-the-art instance and boundary\nsegmentation methods, as well as simple average fusion of multi-scale\npredictions. We evaluate our method on a refractory raw material dataset of\nhigh-resolution images with large variation in crystal size and show that our\nmodel can be used to calculate the crystal size more accurately than existing\nmethods.\n