As well known, fully convolutional network (FCN) becomes the state of the art for semantic segmentation in deep learning. Currently, new hardware designs for deep learning have focused on improving the speed and parallelism of processing units. This motivates memristive solutions, in which the memory units (i.e., memristors) have computing capabilities. However, designing a memristive deep learning network is challenging, since memristors work very differently from the traditional CMOS hardware. This paper proposes a complete solution to implement memristive FCN (MFCN). Voltage selectors are firstly utilized to realize max-pooling layers with the detailed MFCN deconvolution hardware circuit by the massively parallel structure, which is effective since the deconvolution kernel and the input feature are similar in size. Then, deconvolution calculation is realized by converting the image into a column matrix and converting the deconvolution kernel into a sparse matrix. Meanwhile, the convolution realization in MFCN is also studied with the traditional sliding window method rather than the large matrix theory to overcome the shortcoming of low efficiency. Moreover, the conductance values of memristors are predetermined in Tensorflow with ex-situ training method. In other words, we train MFCN in software, then download the trained parameters to the simulink system by writing memristor. The effectiveness of the designed MFCN scheme is verified with improved accuracy over some existing machine learning methods. The proposed scheme is also adapt to LFW dataset with three-classification tasks. However, the MFCN training is time consuming as the computational burden is heavy with thousands of weight parameters with just six layers. In future, it is necessary to sparsify the weight parameters and layers of the MFCN network to speed up computing.

Memristive Fully Convolutional Network: An Accurate Hardware Image-Segmentor in Deep Learning

Denoising and recognition using hidden Markov models with observation distributions modeled by hidden Markov trees

Global optimization of wavelet-domain hidden Markov tree for image segmentation

In this paper, we propose a new approach to intelligently segment jacquard warp-knitted fabric images by combining wavelet texture decomposition, multiresolution Markov random field (MRF) modeling ...

Intelligent segmentation of jacquard warp-knitted fabric using a multiresolution Markov random field with adaptive weighting in the wavelet domain

Multiscale fusion of wavelet-domain hidden Markov tree through graph cut

The invention relates to a sight-testing and on-line adjusting method for printing pressure machine. Printing matters are printed and then accurately located though a locating device and then an executive command on the sight test of the printing pressure is sent out; the testing region is identified though a sight testing unit and the printing pressure of each color class is calculated through feature mapping; the current pressure and the printing pressure of original draft are compared with each other and if the difference exceeds the error set value, the pressure is adjusted according to the error; the printing pressure of each color class is adjusted on line through the driving and pneumatic servo device of a main control unit and the position of a cylinder piston is fed back to the main control unit. Through constant detection and adjustment, basically, the printing pressure can be accurate. Besides, the method changes the known detecting method for printing pressure, thus realizing continuous sight testing of the print pressure and solving the known problem of no on-line adjustment of the printing pressure in the printing process, thus guaranteeing the printing process.

Printing pressing machine vision checking and on-line adjusting method

We present a design methodology for automatic machine vision application aiming at detecting the size ratio of tobacco leaves which will be feedback to adjust running parameters of manufacture system. Firstly, the image is represented by Markov Random Field(MRF) model which consists of a label field and an observation field. Secondly, according to Bayes theorem, the segmentation problem is translated into Maximum a Posteriori(MAP) estimation of the label field and the estimation problem is solved by Iterated Conditional Model(ICM) algorithm. Finally we give the setup of the inspection system and experimented with a real-time image acquired from it, the experiment shows better detection results than Otsu's segmentation method especially in the larger leaf regions.

Automatic inspection of tobacco leaves based on MRF image model

The wavelet domain hidden Markov tree (WHMT) model can decompose the original image into a multiscale and multiband representation. In traditional methods, each WHMT model has to be trained with a single texture image, i.e., each texture is represented by a corresponding WHMT model. This method is memory consuming and do not work for unknown textures. More importantly, the model training of the wavelet domain hidden Markov tree does not take into consideration of the classification likelihood of the foreground and background observations in an optimum sense. In this paper, we develop a probabilistic approach to learn the a priori distribution of foreground objects and backgrounds of WHMT based on the Bayesian optimum statistical classifiers. Instead of computing the class labels of each pixel in the image, we only compute the likelihood of each pixel that belongs to foreground and background, which is then assigned to the classification likelihood of WHMT model. The robustness and accuracy of the proposed algorithm is demonstrate by using four real world horse image come from the benchmark of Weizmann Horse database.

Multiscale Image Segmentation Using Bayesian Optimum Statistical Estimation

Conventional uncooled infrared temperature detectors can only detect targets in two-dimensional temperature field. We propose to achieve three-dimensional temperature field in microwave reaction chamber through binocular visual reconstruction. The infrared and visible multispectral images at two-views are acquisitioned via an infrared detector. To improve consistency of feature point detection results, we perform stereo matching at visible band with high spatial resolution. Disparity mapping with epipolar geometry constraints between two multiscale images are estimated, with which three-dimensional temperature field of depth information is reconstructed by triangulation back projection. This method does not require calibration of infrared detectors and possess certain research and application significance for carrying out the full range observation of heated objects as well as timely regulation of microwave intensity.

Zifen He

Papers

Multiscale fusion of wavelet-domain hidden Markov tree through graph cut

Printing pressing machine vision checking and on-line adjusting method

Automatic inspection of tobacco leaves based on MRF image model

Multiscale Image Segmentation Using Bayesian Optimum Statistical Estimation

Stereo Vision Reconstruction of Infrared Temperature in Microwave Reaction Chamber