| Summary: | 碩士 === 國立清華大學 === 資訊系統與應用研究所 === 104 === Mobile product recognition aims to recognize the product image by retrieving the similar images from the dataset. Recognition accuracy is largely affected by two challenging issues: inter-product similarity and intra-product variations. In this work, we first introduce a multi-stage method to tackle the two issues through multiple convolutional neural networks. After validating the effectiveness, we further propose to simplify the repetitive convolutional operations involved in different stages. In the second proposed method (the two-stage method), we aim to design a deep neural network that can jointly solve the two above-mentioned issues. We first adopt Faster RCNN to simultaneously locate the product and roughly categorize the query image. We next re-use the feature representation learned in earlier layers and involve the part configuration and fine-grained appearance into a deep neural network in the final product recognition stage. In the final stage, we design two kinds of deep neural networks that are rotation-invariant and measure the similarity accurately. The second proposed two-stage method tackles the product recognition task in a more elegant and efficient way without compromising the performance. In addition, we collect a new publicly available dataset PRODUCT-100, which contains 100 products taken under real-world scenarios. Our experiments demonstrate that our method achieve promising results and outperforms existing methods on both PRODUCT-100 and SHORT dataset.
|
|---|
