A neural network based retrainable framework for robust object recognition with application to mobile robotics

Semantic Segmentation Leveraging Simultaneous Depth Estimation

Semantic segmentation is one of the most widely studied problems in computer vision communities, which makes a great contribution to a variety of applications. A lot of learning-based approaches, such as Convolutional Neural Network (CNN), have made a vast contribution to this problem. While rich context information of the input images can be learned from multi-scale receptive fields by convolutions with deep layers, traditional CNNs have great difficulty in learning the geometrical relationship and distribution of objects in the RGB image due to the lack of depth information, which may lead to an inferior segmentation quality. To solve this problem, we propose a method that improves segmentation quality with depth estimation on RGB images. Specifically, we estimate depth information on RGB images via a depth estimation network, and then feed the depth map into the CNN which is able to guide the semantic segmentation. Furthermore, in order to parse the depth map and RGB images simultaneously, we construct a multi-branch encoder-decoder network and fuse the RGB and depth features step by step. Extensive experimental evaluation on four baseline networks demonstrates that our proposed method can enhance the segmentation quality considerably and obtain better performance compared to other segmentation networks.

Image Inpainting Using Two-Stage Loss Function and Global and Local Markovian Discriminators

Image inpainting networks can produce visually reasonable results in the damaged regions. However, existing inpainting networks may fail to reconstruct the proper structures or tend to generate the results with color discrepancy. To solve this issue, this paper proposes an image inpainting approach using the proposed two-stage loss function. The loss function consists of different Gaussian kernels, which are utilized in different stages of network. The use of our two-stage loss function in coarse network helps to focus on the image structure, while the use of it in refinement network is helpful to restore the image details. Moreover, we proposed a global and local PatchGANs (GAN means generative adversarial network), named GL-PatchGANs, in which the global and local markovian discriminators were used to control the final results. This is beneficial to focus on the regions of interest (ROI) on different scales and tends to produce more realistic structural and textural details. We trained our network on three popular datasets on image inpainting separately, both Peak Signal to Noise ratio (PSNR) and Structural Similarity (SSIM) between our results, and ground truths on test images show that our network can achieve better performance compared with the recent works in most cases. Besides, the visual results on three datasets also show that our network can produce visual plausible results compared with the recent works.

SDAE-BP Based Octane Number Soft Sensor Using Near-infrared Spectroscopy in Gasoline Blending Process

As the most important properties in the gasoline blending process, octane number is difficult to be measured in real time. To address this problem, a novel deep learning based soft sensor strategy, by using the near-infrared (NIR) spectroscopy obtained in the gasoline blending process, is proposed. First, as a network structure with hidden layer as symmetry axis, input layer and output layer as symmetric, the denosing auto-encoder (DAE) realizes the advanced expression of input. Additionally, the stacked DAE (SDAE) is trained based on unlabeled NIR and the weights in each DAE is recorded. Then, the recorded weights are used as the initial parameters of back propagation (BP) with the reason that the SDAE trained initial weights can avoid local minimums and realizes accelerate convergence, and the soft sensor model is achieved with labeled NIR data. Finally, the achieved soft sensor model is used to estimate the real time octane number. The performance of the method is demonstrated through the NIR dataset of gasoline, which was collected from a real gasoline blending process. Compared with PCA-BP (the dimension of datasets of BP reduced by principal component analysis) soft sensor model, the prediction accuracy was improved from 86.4% of PCA-BP to 94.8%, and the training time decreased from 20.1 s to 16.9 s. Therefore, SDAE-BP is proposed as a novel method for rapid and efficient determination of octane number in the gasoline blending process.

Improving Neural-Network Classifiers Using Nearest Neighbor Partitioning

This paper presents a nearest neighbor partitioning method designed to improve the performance of a neural-network classifier. For neural-network classifiers, usually the number, positions, and labels of centroids are fixed in partition space before training. However, that approach limits the search for potential neural networks during optimization; the quality of a neural network classifier is based on how clear the decision boundaries are between classes. Although attempts have been made to generate floating centroids automatically, these methods still tend to generate sphere-like partitions and cannot produce flexible decision boundaries. We propose the use of nearest neighbor classification in conjunction with a neural-network classifier. Instead of being bound by sphere-like boundaries (such as the case with centroid-based methods), the flexibility of nearest neighbors increases the chance of finding potential neural networks that have arbitrarily shaped boundaries in partition space. Experimental results demonstrate that the proposed method exhibits superior performance on accuracy and average f-measure.