Invited column-“Deep learning and its application”
2018, 47(7): 703001.
doi: 10.3788/IRLA201847.0703001
The traditional classification methods of PolSAR image generally required the feature extraction in the early stage, involving more human participation, and the classification accuracy needed further improvement. In addition, when using supervised classification method, there were sometimes small sample problems. In view of these problems and combining the requirement of PolSAR image fine classification, a PolSAR image classification method based on 3D convolution neural network was proposedr. The traditional convolution neural network was extended to three dimensions and applied to PolSAR image classification, and the classification method was described in detail. Thus, the characteristics of the multichannel PolSAR image could be fully excavated and improve the classification performance. Moreover, the method of virtual sample expansion was used to improve the small sample situation of certain category and get better classification results. Experimental results showed that 3D convolution neural network could get better performance than 2D convolution neural network in PolSAR image classification and the virtual sample expansion method could effectively improve the small sample classification problem.
2018, 47(7): 703002.
doi: 10.3788/IRLA201847.0703002
In recent years, deep learning techniques have achieved great breakthrough for its powerful non-linear computations in the tasks of target recognition and detection. Existing deep networks were almost designed based on the precondition that the visual data reside on the Euclidean space. However, many data in computer vision have rigorous geometry of manifolds, i.e., image sets can be represented as Grassmann manifolds. The deep network was devised based on the non-Euclidean structure of the manifold-valued data, which combined the differential geometry and deep learning methods theoretically. Furthermore, a deep network for image-set recognition based on the Grassmann manifold was proposed. In the training process, the model was updated by the use of the backpropagation algorithm derived from the matrix chain rule. Learning of the weights can be transformed as the Riemannian optimization problem on the Grassmannian. The experimental results show that this method not only improves the accuracy of recognition, but also accelerates the training and test process in one magnitude.
2018, 47(7): 703003.
doi: 10.3788/IRLA201847.0703003
In order to improve the accuracy of infrared fault image recognition for electrical equipment, an infrared fault image recognition method for electrical equipment based on double supervised signal deep learning was proposed. Firstly, a Slic super pixel segmentation algorithm was adopted to merge the similar pixel regions into blocks. According to the luminance information provided by the improved HSV space transformation, the temperature abnormal regions were determined. Secondly, the connected areas and the corresponding device of this region were separated. Finally, based on the GoogLeNet convolution neural network model, fault features of infrared images for electrical equipments were extracted, then trained and supervised by two kinds of signals, i.e., the softmax loss and the center loss signal. Among an established 700 infrared fault of electrical equipment images dataset, 500 of which are for training, and 200 for testing. Experiments results show that the test accuracy rate can reach to 98.6% which enhanced 1% when being compared with the classic method simply using the single softmax loss. The algorithm can accurately locate five kinds of electrical equipments which include the transformer bushing, current transformer, surge arrester, isolating switch, insulators, as well as identify the corresponding faults.
2018, 47(7): 703004.
doi: 10.3788/IRLA201847.0703004
Since the current image quality assessment methods are generally based on hand-crafted features, it is difficult to automatically and effectively extract image features that conform to the human visual system. Inspired by human visual characteristics, a new method of full-reference image quality assessment was proposed by this paper which was based on convolutional neural network (DeepFR). According to this method, the DeepFR model of convolutional neural network was designed which was based on the understanding of the dataset by itself using the human visual system to weight the sensitivity of the gradient, and the visual gradient perception map was extracted that was consistent with human visual characteristics. The experimental results show that the DeepFR model is superior to the current full-reference image quality assessment methods, its prediction score and subjective quality evaluation have good accuracy and consistency.
2018, 47(7): 703005.
doi: 10.3788/IRLA201847.0703005
In order to improve the small object detection ability of SSD object detection algorithm, the transposed convolution structure in SSD algorithm was proposed, the low resolution high semantic information feature map was integrated with high resolution low semantic information feature map using transposed convolution, which increased the ability of low level feature extraction and improved the average accuracy of SSD algorithm. At the same time for the problem that SSD algorithm model being large, running memory consumption high, without running on the embedded equipment ARM, a lightweight feature extraction minimum unit was proposed based on DenseNet, combining depthwise separable convolutions, pointwise group convolution and channel shuffle, running on the embedded equipment ARM cloud be realized. The comparative experiments on PASCAL VOC data set and KITTI autopilot data set show that the mean average is significantly improved by improved network structure, and the number of model parameters is effectively reduced.
