| Summary: | 碩士 === 國立高雄科技大學 === 電腦與通訊工程系 === 107 === Recently, the use of Convolutional Neural Network (CNN) automatically extracts effective features from raw data. CNN is different from the hand-crafted features which is designed by human being for solving the specific image processing problems. In order to investigate the complement relation between CNN features (or called deep feature) and hand-crafted features, this work integrates two kinds of features for thermal obstacle detection. Deep feature is based on the YOLO detector; hand-crafted features used are HLID and LBP which are the two widely used features in thermal images. The way for investigating the integration effect is by respectively impose the hand-crafted features to different CNN layers. In order to achieve this, we change the concatenation way of the above-mentioned hand-crafted features and integrate it into the convolutional neural network to observe the effect of hand-crafted features on detection performance. In the experimental, we use the self-labeled thermal image dataset for verification, and integrate the hand-crafted features into some layers of the neural network. The metric used for evaluation is mAP.
|
|---|
