• Journal of Multimedia Information System
• /
• v.9 no.3
• /
• pp.177-182
• /
• 2022

A deep neural network (DNN) includes variables whose values keep on changing with the training process until it reaches the final point of convergence. These variables are the co-efficient of a polynomial expression to relate to the feature extraction process. In general, DNNs work in multiple 'dimensions' depending upon the number of channels and batches accounted for training. However, after the execution of feature extraction and before entering the SoftMax or other classifier, there is a conversion of features from multiple N-dimensions to a single vector form, where 'N' represents the number of activation channels. This usually happens in a Fully connected layer (FCL) or a dense layer. This reduced 2D feature is the subject of study for our analysis. For this, we have used the FCL, so the trained weights of this FCL will be used for the weight-class correlation analysis. The popular DNN models selected for our study are ResNet-101, VGG-19, and GoogleNet. These models' weights are directly used for fine-tuning (with all trained weights initially transferred) and scratch trained (with no weights transferred). Then the comparison is done by plotting the graph of feature distribution and the final FCL weights.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 2009.01a
• /
• pp.49-54
• /
• 2009

A fuzzy weighted mean method on a 2D shape recognition system is introduced in this paper. The bispectrum based on third order cumulant is applied to the contour sequence of each image for the extraction of a feature vector. This bispectral feature vector, which is invariant to shape translation, rotation and scale, represents a 2D planar image. However, to obtain the best performance, it should be considered certain criterion on the calculation of weights for the fuzzy weighted mean method. Therefore, a new method to calculate weights using means by differences of feature values and their variances with the maximum distance from differences of feature values. is developed. In the experiments, the recognition results with fifteen dimensional bispectral feature vectors, which are extracted from 11.808 aircraft images based on eight different styles of reference images, are compared and analyzed.

• The Journal of the Acoustical Society of Korea
• /
• v.31 no.2
• /
• pp.87-92
• /
• 2012

In this paper, a robust feature compensation method to deal with the environmental mismatch is proposed. The proposed method applies energy based weights according to the degree of speech presence to the Mean subtraction, Variance normalization, and ARMA filtering (MVA) processing. The weights are further smoothed by the moving average and maximum filters. The proposed feature compensation algorithm is evaluated on AURORA 2 task and distant talking experiment using the robot platform, and we obtain error rate reduction of 14.4 % and 44.9 % by using the proposed algorithm comparing with MVA processing on AURORA 2 task and distant talking experiment, respectively.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2022.11a
• /
• pp.594-596
• /
• 2022

Over the last decades, deep neural networks have demonstrated significant success in various tasks. To address the special vision task, choosing a hot network as backbone to extract feature is a common way in both research and industry project. However, the choice of backbone usually requires the expert experience and affects the performance of the classification task. In this work, we propose a novel idea to support backbone decision-making by exploring the feature attribution and weights distribution of hidden layers from various backbones. We first analyze the visualization of feature maps on different size object and different depth layers to observe learning ability. Then, we compared the variance of weights and feature in last three layers. Based on analysis of the feature and wights, we summarize the traits and commonalities of existing networks.

• Proceedings of the KIEE Conference
• /
• 2004.11c
• /
• pp.162-164
• /
• 2004

This paper proposes a new classification system combining the adaptive feature weighting algorithm using the genetic algorithm and the modified KNN rule. GA is employed to choose the middle value of weights and weights of features for high performance of the system. The modified KNN rule is proposed to estimate the class of test pattern using adaptive feature space. Experiments with the unconstrained handwritten digit database of Concordia University in Canada are conducted to show the performance of the proposed method.

• Journal of KIISE:Computing Practices and Letters
• /
• v.9 no.2
• /
• pp.158-165
• /
• 2003

Image similarity is often decided by computing the distance between two feature vectors. Unfortunately, the feature vector cannot always reflect the notion of similarity in human perception. Therefore, most current image retrieval systems use weights measuring the importance of each feature. In this paper new initial weight selection and update rules are proposed for image retrieval purpose. In order to obtain the purpose, database images are first divided into groups based on human perception and, inner and outer query are performed, and, then, optimal feature weights for each database images are computed through searching the group where the result images among retrieved images are belong. Experimental results on 2000 images show the performance of proposed algorithm.

• Journal of the Institute of Electronics and Information Engineers
• /
• v.51 no.11
• /
• pp.66-72
• /
• 2014

Histogram of oriented gradient (HOG) feature is widely used in vision-based pedestrian detection. The interpolation is the most important technique in HOG feature calculation to provide high detection rate. In interpolation technique of HOG feature calculation, two nearest orientation bins to gradient orientation for each pixel and the corresponding weights are required. In this paper, therefore, an efficient gradient orientation bin and weight calculation circuit for HOG feature is proposed. In the proposed circuit, pre-calculated values are defined in tables to avoid the operations of tangent function and division, and the size of tables is minimized by utilizing the characteristics of tangent function and weights for each gradient orientation. Pipeline architecture is adopted to the proposed circuit to accelerate the processing speed, and orientation bins and the corresponding weights for each pixel are calculated in two clock cycles by applying efficient coarse and fine search schemes. Since the proposed circuit calculates gradient orientation for each pixel with the interval of $1^{\circ}$ and determines both orientation bins and weights required in interpolation technique, it can be utilized in HOG feature calculation to support interpolation technique to provide high detection rate.

• International Journal of Quality Innovation
• /
• v.3 no.2
• /
• pp.113-131
• /
• 2002

The relationships among multi-dimensional data (such as medical examination data) with ambiguity and variation are difficult to explore. The traditional approach to building a data classification system requires the formulation of rules by which the input data can be analyzed. The formulation of such rules is very difficult with large sets of input data. This paper first describes two classification approaches using back-propagation (BP) neural network and Mahalanobis distance (MD) classifier, and then proposes two classification approaches for multi-dimensional feature selection. The first one proposed is a feature selection procedure from the trained back-propagation (BP) neural network. The basic idea of this procedure is to compare the multiplication weights between input and hidden layer and hidden and output layer. In order to simplify the structure, only the multiplication weights of large absolute values are used. The second approach is Mahalanobis-Taguchi system (MTS) originally suggested by Dr. Taguchi. The MTS performs Taguchi's fractional factorial design based on the Mahalanobis distance as a performance metric. We combine the automatic thresholding with MD: it can deal with a reduced model, which is the focus of this paper In this work, two case studies will be used as examples to compare and discuss the complete and reduced models employing BP neural network and MD classifier. The implementation results show that proposed approaches are effective and powerful for the classification.

• Journal of the Korea Society of Computer and Information
• /
• v.14 no.12
• /
• pp.55-62
• /
• 2009

This paper proposes a decision method of fuzzy weights by utilizing degrees of contribution in order to classify insect footprint patterns having difficulties to classify species clearly. Insect footprints revealed delicately in the form of scattered spots since they are very small. Therefore it is not easy to define shape of footprints unlike other species, and there are lots of noises in the footprint patterns so that it is difficult to distinguish those from correct data. For these reasons, the extracted feature set has obvious feature values with some uncertain feature values, so we estimate weights according to degrees of contribution. If the one of feature values has distinct difference enough to decide a class among other classes, high weight is assigned to make classification. A calculated weight determines the membership values by fuzzy functions and objects are classified into the class having a superior value.atu present experimental resultseighrontribution. Iinsect footprints with noises by the proposed method.

• Journal of Institute of Control, Robotics and Systems
• /
• v.5 no.7
• /
• pp.821-826
• /
• 1999

This paper proposes a face recognition method using feature spaces and image vectors in the image plane. We obtain the 2-D feature space using the self-organizing map which has two inputs from the axis of the given image. The image vector consists of its weights and the average gray levels in the feature space. Also, we can reconstruct an normalized face by using the image vector having no connection with the size of the given face image. In the proposed method, each face is recognized with the best match of the feature spaces and the maximum match of the normally retrieval face images, respectively. For enhancing recognition rates, our method combines the two recognition methods by the feature spaces and the retrieval images. Simulations are conducted on the ORL(Olivetti Research laboratory) images of 40 persons, in which each person has 10 facial images, and the result shows 100% recognition and 14.5% rejection rates for the 20$\times$20 feature sizes and the 24$\times$28 retrieval image size.