• Smart Media Journal
• /
• v.12 no.9
• /
• pp.45-59
• /
• 2023

Weakly supervised object localization (WSOL) is a task of localizing an object in an image using only image-level labels. Previous studies have followed the conventional class activation mapping (CAM) pipeline. However, we reveal the current CAM approach suffers from problems which cause original CAM could not capture the complete defects features. This work utilizes a convolutional neural network (CNN) pretrained on image-level labels to generate class activation maps in a multi-scale manner to highlight discriminative regions. Additionally, a vision transformer (ViT) pretrained was treated to produce multi-head attention maps as an auxiliary detector. By integrating the CNN-based CAMs and attention maps, our approach localizes defective regions without requiring bounding box or pixel-level supervision during training. We evaluate our approach on a dataset of apple images with only image-level labels of defect categories. Experiments demonstrate our proposed method aligns with several Object Detection models performance, hold a promise for improving localization.

• Journal of the Institute of Electronics and Information Engineers
• /
• v.50 no.9
• /
• pp.194-200
• /
• 2013

In this paper, we propose a multi-scale crack detection method using scaling. It is based on morphology algorithm, crack features, and scaling. We use a morphology operator which extracts patterns of crack. It segments cracks and background using opening and closing operations. Morphology based segmentation is better than existing integration methods using subtraction in detecting a crack it has small width. However, morphology methods using only one structure element could detect only fixed width crack. Thus, we use a scaling method. We use bilinear interpolation for scaling. Our method calculates values of properties such as the number of pixels and the maximum length of the segmented region. We decide whether the segmented region belongs to cracks according to those data. Experimental results show that our proposed multi-scale crack detection method has better results than those of existing detection methods.

• Korean Journal of Remote Sensing
• /
• v.34 no.6_3
• /
• pp.1351-1367
• /
• 2018

Image registration is an essential process for image fusion, change detection and time series analysis using multi-sensor images. For this purpose, we need to detect accurately the difference of scale and rotation between the multi-sensor images with difference spatial resolution. In this paper, we propose a new feature matching method using variable circular template for image registration between multi-resolution images. The proposed method creates a circular template at the center of a feature point in a coarse scale image and also a variable circular template in a fine scale image, respectively. After changing the scale of the variable circular template, we rotate the variable circular template by each predefined angle and compute the mutual information between the two circular templates and then find the scale, the angle of rotation and the center location of the variable circular template, respectively, in fine scale image when the mutual information between the two circular templates is maximum. The proposed method was tested using Kompsat-2, Kompsat-3 and Kompsat-3A images with different spatial resolution. The experimental results showed that the error of scale factor, the error of rotation angle and the localization error of the control point were less than 0.004, $0.3^{\circ}$ and one pixel, respectively.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.5 no.12
• /
• pp.2374-2391
• /
• 2011

A variety of reconstruction methods has been developed to convert a set of scattered points generated from real models into explicit forms, such as polygonal meshes, parametric or implicit surfaces. In this paper, we present a method to construct multi-scale implicit surfaces from scattered points using multiscale kernels based on kernel and multi-resolution analysis theories. Our approach differs from other methods in that multi-scale reconstruction can be done without additional manipulation on input data, calculated functions support level of detail representation, and it can be naturally expanded for n-dimensional data. The method also works well with point-sets that are noisy or not uniformly distributed. We show features and performances of the proposed method via experimental results for various data sets.

• International Journal of Control, Automation, and Systems
• /
• v.4 no.5
• /
• pp.615-623
• /
• 2006

The gyroscope free strap-down INS is composed only of accelerometers. Any gyroscope free INS navigation error is deeply affected by the accuracy of the sensor bias, scale factor, orientation and location error. However these parameters can be found by calibration. There is an important research issue about a multi-position calibration method in this paper. It provides a novel method to find the error parameters for the six-accelerometer INS. A superior simulation is shown that the multi-position calibration can find the specifications of a six-accelerometer INS in laboratory. From these parameters the six-accelerometer INS could apply in realistic navigation.

• Journal of the Korea Society of Computer and Information
• /
• v.10 no.4 s.36
• /
• pp.239-248
• /
• 2005

The reliability and efficiency of network must be considered in the large scale wireless sensor networks. Broadcast method must be used rather than unicast method to enhance the reliability of networks. In recently proposed GRAB (GRAdient Broadcast) can certainly enhance reliability of networks fy using broadcast but its efficiency regarding using energy of network is low due to using only one sink. Hence, the lifetime of networks is reduced. In the paper we propose the scheme of SMSGB (Selective Multi Sink Gradient Broadcast) which uses single sink of multi-sink networks. The broadcast based SMSGB can secure reliability of large scale wireless sensor networks. The SMSGB can also use the network's energy evenly via multi sink distribution. Our experiments show that using SMSGB was reliable as GRAB and it increased the network's lifetime by 18% than using GRAB.

• Proceedings of the KSRS Conference
• /
• 2003.11a
• /
• pp.1209-1211
• /
• 2003

The wavelet and semivariogram analysis method are used to identify the city landscape and farmland landscape pattern on the 1m resolution IKONOS images. The results prove that wavelet method is a potential way for landscape pattern analysis. Compared to semivariogram analysis, Wavelet analysis can not only detect the overall spatial pattern, but also find multi-scale and direction structures. In this experiment, the wavelet analysis results indicate: (1) the city landscape image is mainly composed of three level structures whose spatial pattern characters appear at 2m, 16m, 128m and 256m accordingly; (2) the farmland landscape is mainly two scale spatial patterns appearing at the 2m, 128m and 256m. IKONOS Remote sensing, with the high spatial and spectral information, is a powerful tool that can use in many ecological systems research and sustainable management.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.15 no.3
• /
• pp.993-1014
• /
• 2021

This paper proposes a new image retrieval method called the 3D integrated multi-index to fuse SIFT (Scale Invariant Feature Transform) visual words with other features at the indexing level. The advantage of the 3D integrated multi-index is that it can produce finer subdivisions in the search space. Compared with the inverted indices of medium-sized codebook, the proposed method increases time slightly in preprocessing and querying. Particularly, the SIFT, contour and colour features are fused into the integrated multi-index, and the joint cooperation of complementary features significantly reduces the impact of false positive matches, so that effective image retrieval can be achieved. Extensive experiments on five benchmark datasets show that the 3D integrated multi-index significantly improves the retrieval accuracy. While compared with other methods, it requires an acceptable memory usage and query time. Importantly, we show that the 3D integrated multi-index is well complementary to many prior techniques, which make our method compared favorably with the state-of-the-arts.

• Journal of the Korea Society of Computer and Information
• /
• v.25 no.7
• /
• pp.47-55
• /
• 2020

In this paper, we propose one stage multi-scale face detector based Fully Convolution Network using anchor free method. Recently almost all state-of-the-art face detectors which predict location of faces using anchor-based methods rely on pre-defined anchor boxes. However this face detectors need to hyper-parameters and additional computation in training. The key idea of the proposed method is to eliminate hyper-parameters and additional computation using anchor free method. To do this, we apply two ideas. First, by eliminating the pre-defined set of anchor boxes, we avoid the additional computation and hyper-parameters related to anchor boxes. Second, our detector predicts location of faces using multi-feature maps to reduce foreground/background imbalance issue. Through Quantitative evaluation, the performance of the proposed method is evaluated and analyzed. Experimental results on the FDDB dataset demonstrate the effective of our proposed method.

• Journal of Korea Multimedia Society
• /
• v.17 no.12
• /
• pp.1383-1392
• /
• 2014

In this paper, we proposed a method to automatically determine the optimal radius through multi-scale LBP operation generalizing the size of radius variation and boosting learning in facial emotion recognition. When we looked at the distribution of features vectors, the most common was $LBP_{8.1}$ of 31% and sum of $LBP_{8.1}$ and $LBP_{8.2}$ was 57.5%, $LBP_{8.3}$, $LBP_{8.4}$, and $LBP_{8.5}$ were respectively 18.5%, 12.0%, and 12.0%. It was found that the patterns of relatively greater radius express characteristics of face well. In case of normal and anger, $LBP_{8.1}$ and $LBP_{8.2}$ were mainly distributed. The distribution of $LBP_{8.3}$ is greater than or equal to the that of $LBP_{8.1}$ in laugh and surprise. It was found that the radius greater than 1 or 2 was useful for a specific emotion recognition. The facial expression recognition rate of proposed multi-scale LBP method was 97.5%. This showed the superiority of proposed method and it was confirmed through various experiments.