Abstract:
Automated shell abrasion detection is essential for secondhand cell phone quality assessment and pricing in industry, which remains a very challenging problem due to the ...Show MoreMetadata
Abstract:
Automated shell abrasion detection is essential for secondhand cell phone quality assessment and pricing in industry, which remains a very challenging problem due to the abrasions usually appearing with irregular distributions, various scales and shapes, and blurred boundaries. In this article, we propose an automated shell abrasion detection network for cell phone based on GraphFPN with learnable sparse priors. To fully use rich multiscale abrasion features, we first apply a GraphFPN to the input image, which can adapt its topology to different abrasion structures and support synchronous feature interaction across all scales. To achieve a more accurate multiscale abrasion feature fusion, we further apply a spatial squeeze and excitation to recalibrate multiscale abrasion feature maps extracted from GraphFPN. Finally, we can efficiently obtain the abrasion detection results based on an end-to-end sparse R-convolutional neural network (CNN) with learnable sparse priors. Our method is extensively evaluated on our collected cell phone shell (CPS) dataset, which includes tens of thousands of CPS photographs for different brands (including Apple, Huawei, Oppo, Vivo, and etc.) and with various sizes of abrasions. The experimental results of ablation studies demonstrate the effectiveness of the proposed method in handling challenging abrasions with various sizes, irregular shapes, and low-contrast boundaries. Ablation studies and comparisons with state-of-the-art methods are also conducted to verify the advantages of the proposed method in shell abrasion detection with a higher accuracy.
Published in: IEEE Transactions on Instrumentation and Measurement ( Volume: 71)