Due to the complexity and similarity of plant leaves, it is very important to study an effective leaf-feature extraction method to improve the recognition rate of plant leaves. We study five multi-scale triangle representations: the triangle unsigned area representation (TUA), the triangle vertex angle representation (TVA) and three new representations, which we define as the gray average (TGA), the gray standard deviation (TGSD) and the side length integral (TSLI) on the triangle. In this method the curvature features of the contour, the texture features and the shape area feature are extracted to provide a multiscale leaf-feature description, and a new adaptive KNN for optimization method is proposed to improve the retrieval rate of leaf datasets. Experiments show that compared with the state-of-the-art methods, our method has higher accuracy on the Swedish and Flavia plant leaf datasets, which are respectively 99.35% and 99.43% with 84.76% Mean Average Precision (MAP) value and has comparable results on MPEG-7, kimia99 and kimia216 datasets. When our method is combined with KNN for optimization, the retrieval rate of the above datasets has been significantly improved, especially MAP on the Flavia dataset increases to 94.48%.

With the development of E-commerce, online advertising began to thrive and has gradually developed into a new mode of business, of which Click-Through Rates (CTR) prediction is the essential driving technology. Given a user, commodities and scenarios, the CTR model can predict the user’s click probability of an online advertisement. Recently, great progress has been made with the introduction of Deep Neural Networks (DNN) into CTR. In order to further advance the DNN-based CTR prediction models, this paper introduces a new model of FO-FTRL-DCN, based on the prestigious model of Deep&Cross Network (DCN) augmented with the latest optimization technique of Follow The Regularized Leader (FTRL) for DNN. The extensive comparative experiments on the iPinYou datasets show that the proposed model has outperformed other state-of-the-art baselines, with better generalization across different datasets in the benchmark.

Semantic segmentation is one of the most fundamental techniques for visual intelligence, which plays a vital role for indoor service robotic tasks such as scene understanding, autonomous navigation and dexterous manipulation. However, semantic segmentation of indoor environments poses great challenges for existing segmentation techniques due to the complex overlaps, heavy occlusions and cluttered scenes with objects of different shapes and scales, which may lead to the loss of edge information and insufficient segmentation accuracy. And most of the semantic segmentation networks are very complex and cannot be applied to mobile robot platforms. Thus, it is of significant importance for ensuring as few network parameters as possible while improving the detection of meaningful edges in indoor scenes. In this paper, we present a novel indoor scene semantic segmentation method that can refine the segmentation edges and achieve a balance between accuracy and model complexity for indoor service robots. Our approach systematically incorporates dilated convolution and rich convolutional features from the intermediate layers of Convolutional Neural Networks (CNN), which is based on two motivations: (1) The middle hidden layer of CNN contains a lot of potentially useful information for better edge detection which is, however, no longer present in latter layers in traditional structures. (2) The dilated convolution can change the size of receptive field and obtain multi-scale feature information without losing the resolution and introducing any additional parameters. Thus we propose a new end-to-end Multi-Scale Convolutional Features (MSCF) network to integrate the dilated convolution and rich convolutional features extracted from the intermediate layers of traditional CNN. Finally, the resulting approach is extensively evaluated on the prestigious indoor image datasets of SUN RGB-D and NYUDv2, and shows promising improvements over state-of-the-art baselines, both qualitatively and quantitatively.

Massively pre-trained transformer models such as BERT have gained great success in many downstream NLP tasks. However, they are computationally expensive to fine-tune, slow for inference, and have large storage requirements. So, transfer learning with adapter modules has been introduced and has become a remarkable solution for those problems. Nevertheless, recent studies reveal that the parameters in adapters are actually still quite redundant, which could slow down inference speed when fusing multiple adapters for a specific downstream task, and thus, they can be further reduced. To address this issue, we propose three novel ways to prune the adapter modules iteratively based on the prestigious Lottery Ticket Hypothesis. Extensive experiments on the GLUE datasets show that the pruned adapters can achieve state-of-the-art results, with sizes reduced significantly while performance remains unchanged, and some pruned adapters even outperform the ones with the same size that are fine-tuned alone without pruning.