| Summary: | Ambiguous appearance discrimination plays an important role in the impurity detection task. Among the majority of deep learning models, images from every sequence are processed separately instead of being considered collectively. Therefore, the outputs of these models given a single region proposal might not be accurate. In this paper, a gallery-guided graph architecture is proposed and integrated to overcome such limitations. Specifically, region proposals are firstly generated using a two-stream fusion network; then their feature embeddings are extracted from a convolutional neural network by reducing intra-class variations while increasing inter-class ones. Secondly, a graph representing clusters among different training sequences updates relationships between region proposals in the test sequence. Finally, the features of the graph are classified by a graph convolutional neural network. Different from those learned weights in conventional common object detectors, region features from all the training sequences are explicitly integrated into a gallery-guided graph architecture. Extensive experiments on IML-DET dataset demonstrate that our proposed method can obtain competitive performances compared with previous state-of-the-art object detection approaches transferred into this task.
|
|---|