Characters pressed on the surface of the liquefied gas cylinder cap are used to mark each cylinder’s identity and record the information about the gas in the cylinder. An automatic character recognition based on a lightweight ResNet model is proposed for the online quality and safety management of cylinder-contained gas production. Firstly, images of pressed characters on the cylinder’s cap are collected under low angle lighting to improve the image contrast of character regions against the background. The connected component method is used to determine the position of each character and then segment the single characters. The pressed-character data set of gas cylinder is constructed by expanding data through image blurring, lightening, darkening and noise adding. According to the data scenario and scale, ResNet18 is selected as the backbone network. To implement the lightweight network model, by removing the last two residual modules in the backbone network, the number of the network layers is reduced to ten. In addition, the number of feature channels is trimmed to further reduce the redundancy of the convolution kernel and improve the calculation efficiency of the network model. Finally, the Max pooling layer is used for all channels to aggregate information and obtain character features. The experimental results show that the recognition accuracy of the proposed network is 94.7% on the constructed gas-cylinder character dataset. Compared with the comparison network, the proposed network not only has higher accuracy and robustness, but also reduces the capacity and computational complexity.
An automatic ID identification system for gas cylinders’ online production was developed based on the production conditions and requirements of the Technical Committee for Standardization of Gas Cylinders. A cylinder ID image acquisition system was designed to improve the image contrast of ID regions on gas cylinders against the background. Then the ID digits region was located by the CNN template matching algorithm. Following that, an adaptive threshold method based on the analysis of local average grey value and standard deviation was proposed to overcome defects of non-uniform background in the segmentation results. To improve the single digit identification accuracy, two BP neural networks were trained respectively for the identification of all digits and the easily confusable digits. If the single digit was classified as one of confusable digits by the former BP neural network, it was further tested by the later one, and the later result was taken as the final identification result of this single digit. At last, the majority voting was adopted to decide the final identification result for the 6-digit cylinder ID. The developed system was installed on a production line of a liquefied-petroleum-gas cylinder plant and worked in parallel with the existing weighing step on the line. Through the field test, the correct identification rate for single ID digit was 94.73%, and none of the tested 2000 cylinder ID was misclassified through the majority voting.
In this paper we study a multi-function and high precision submarine optical fiber monitorsystem. This design is based on the core technology of coherent optical time-domain reflectometry (COTDR) which is used for submarine optical fiber monitor. The design provides multi-functionwhich including: undersea amplifier gain monitor, output power monitor; undersea fiber events monitor (such as fiber cut, attenuation as well as reflection) and also the fiber fault location which may identify the accurate location of the failurepoint (distance from failure point to its upstream amplifier). With these functions, the designcan facilitate the staff to inquiry and debug, and it also improves the maintainability and operability of the undersea equipment.
Hilbert transform (HT) is widely used in temporal speckle pattern interferometry, but errors from low modulations might propagate and corrupt the calculated phase. A spatio-temporal method for phase retrieval using temporal HT and spatial phase unwrapping is presented. In time domain, the wrapped phase difference between the initial and current states is directly determined by using HT. To avoid the influence of the low modulation intensity, the phase information between the two states is ignored. As a result, the phase unwrapping is shifted from time domain to space domain. A phase unwrapping algorithm based on discrete cosine transform is adopted by taking advantage of the information in adjacent pixels. An experiment is carried out with a Michelson-type interferometer to study the out-of-plane deformation field. High quality whole-field phase distribution maps with different fringe densities are obtained. Under the experimental conditions, the maximum number of fringes resolvable in a 416×416 frame is 30, which indicates a 15λ deformation along the direction of loading.
An algorithm based on the histogram of edge difference between MS (model seal) and SS (sample seal) was proposed to
verify Chinese seal imprints in bank checks. Difference between MS and an elaborate fake SS may be slight, while that
between MS and a genuine SS may be not small due to the variety of imprinting conditions. Edge differences wholly
reflect geometrical differences between SS and MS. To evaluate similarities between MS and SS, edge difference was
quantified as two parameters, the distance between non-overlapped corresponding edges and the length of each piece of
non-overlapped seal edges. A histogram on the product of the two parameters is proposed as the input feature vectors of
SVM (support vector machine). SS was verified as true or false by Support Vector Machine (SVM). In Experiments,
4810 (2450 genuine and 2360 fake) seal imprints were verified, and the correct recognition rate is 99.42%. Moreover, the
classification results can be customized according to the requirements of users. When the false-acceptance error rate and
the false-rejection error rate are both required to be close to 0, the rejection rate is about 3%.
A SIFT (Scale Invariant Feature Transform) feature based registration algorithm is presented to prepare for the seal
verification, especially for the verification of high quality counterfeit sample seal. The similarities and the spatial
relationships between the matched SIFT features are combined for the registration. SIFT features extracted from the
binary model seal and sample seal images are matched according to their similarities. The matching rate is used to define
the similar sample seal that is similar with its model seal. For the similar sample seal, the false matches are eliminated
according to the position relationship. Then the homography between model seal and sample seal is constructed and
named HS . The theoretical homography is namedH . The accuracy of registration is evaluated by the Frobenius norm of
H-HS . In experiments, translation, filling and rotation transformations are applied to seals with different shapes, stroke
number and structures. After registering the transformed seals and their model seals, the maximum value of the
Frobenius norm of their H-HS is not more than 0.03. The results prove that this algorithm can accomplish accurate
registration, which is invariant to translation, filling, and rotation transformation, and there is no limit to the seal shapes,
stroke number and structures.
Dynamic speckle pattern interferometry has been widely applied to measure vibration or continuously-deformation. As a
promising technique, temporal phase analysis reduces the 2D phase retrieval task to 1D and gives wider measurement
range. In this paper, some classical and recently proposed temporal phase retrieval techniques, such as windowed Fourier
transform, wavelet transform and Hilbert transform, are comparatively studied. The advantages and drawbacks of each
algorithm are discussed and evaluated in simulation experiments.
Dynamic speckle interferometry using temporal phase analysis has larger measurement range and is easier to setup over
phase-shifting based speckle interferometry. Hilbert transform (HT) is a widely used approach to implement analytic
method based phase retrieval. To fulfill the requirements of HT based phase retrieval, EMD (Empirical Mode
Decomposition) can be applied to remove the bias intensity. With the low noise assumption, the first IMF was taken as
the input for HT. However, according to our experiments, some dynamic speckle signals are not as good as assumed. In
many cases, the first IMF is not the proper one. To find the proper IMF, we proposed to adaptively find the IMF of
largest similarity with the input signal. And the similarity is evaluated by mutual information. Simulation experiments
were given to verify the validity of the proposed algorithm.
This paper proposed an interactive image segmentation algorithm that can tolerate slightly incorrect user constraints.
Interactive image segmentation was formulated as a constrained spectral graph partitioning problem. Furthermore, it was
proven to equal to a supervised classification problem, where the feature space was formed by rows of the eigenvector
matrix that was computed by spectral graph analysis. ν-SVM (support vector machine) was preferred as the classifier.
Some incorrect labels in user constraints were tolerated by being identified as margin errors in ν-SVM. Comparison with
other algorithms on real color images was reported.
An algorithm quantifying edge difference between model seal (MS) and sample seal (SS) was proposed to verify
Chinese seal imprint in bank check. Differences between MS and deliberately faked SS may be slight, while there exists
differences between MS and genuine SS due to the variety of imprinting conditions. To evaluate similarities between MS
and SS, edge difference was quantified as two parameters, the distance between non-overlapped corresponding edges
and the length of each piece of non-overlapped seal edges. According to the two quantified parameters, SS was verified
as true, false or doubtful. 2000 seal imprints (1000 genuine and 1000 fake) were verified for experiments. Results
showed that all the fake seal imprints were verified accurately, even when their differences against MS were minute. 27
genuine seal imprints were misclassified as doubtful due to some serious distortions. The false-acceptance error rate was
0%. The false-rejection error rate was 2.7%, and the correct recognition rate was 98.65%.
KEYWORDS: Digital signal processing, Image segmentation, System identification, Image processing algorithms and systems, Signal processing, Image processing, Computing systems, Algorithm development, Feature extraction, Image acquisition
Automatic seal imprint identification is an important part of modern financial security. Accurate segmentation is the basis
of correct identification. In this paper, a DSP (digital signal processor) based identification system was designed, and an
adaptive algorithm was proposed to extract binary seal images from financial instruments. As the kernel of the
identification system, a DSP chip of TMS320DM642 was used to implement image processing, controlling and
coordinating works of each system module. The proposed algorithm consisted of three stages, including extraction of
grayscale seal image, denoising and binarization. A grayscale seal image was extracted by color transform from a
financial instrument image. Adaptive morphological operations were used to highlight details of the extracted grayscale
seal image and smooth the background. After median filter for noise elimination, the filtered seal image was binarized by
Otsu's method. The algorithm was developed based on the DSP development environment CCS and real-time operation
system DSP/BIOS. To simplify the implementation of the proposed algorithm, the calibration of white balance and the
coarse positioning of the seal imprint were implemented by TMS320DM642 controlling image acquisition. IMGLIB of
TMS320DM642 was used for the efficiency improvement. The experiment result showed that financial seal imprints,
even with intricate and dense strokes can be correctly segmented by the proposed algorithm. Adhesion and
incompleteness distortions in the segmentation results were reduced, even when the original seal imprint had a poor
quality.
In this paper, an adaptive morphological segmentation algorithm is proposed to extract a binary Chinese square seal from
a bank check image. The grayscale Chinese square seal is extracted from the color bank check image according to the
color information. Different Chinese characters have different stroke features and background evenness. To process each
character in the square seal respectively, the extracted square seal is divided into four sub-squares. The background
across each sub-square of the grayscale seal image is smoothed by top-hat transformation. The size of structuring
element in top-hat transformation might have a great influence on the segmentation. The optimal size of the structuring
element for the top-hat transformation on each sub-square is iteratively estimated according to the local foreground area.
Each top-hat processed sub-square is binarized by Otsu's method. In each binary sub-square, holes smaller than a
threshold are filled which is proportional to the ratio of the foreground area to the area of the whole sub-square. The
experiment result shows that the proposed algorithm can correctly segment Chinese characters with intricate and dense
strokes in a bank check square seal. Adhesion and incompleteness distortions in the segmentation results are reduced,
even when the original square seal has a poor quality.
This paper presents a new method to measure the characteristic parameter of the laser holographic anti-counterfeiting
label, with less dependency on the stability of source light. The He-Ne laser, having 5% relative instability, was used as
the light source. Based on the nature that one beam of light can be divided into two beams at a fixed ratio, the source
light was divided into the reflex and the transmission at the fixed ratio by a piece of glass. As the label was scanned
automatically, the intensities of the reflex and the diffraction were measured online and in real time to get the SNR. The
data table for contrast was given, which showed the differences between the real value and the calculated value of the
intensity of the transmission as the incident beam. The measurement results of the SNR of 8 different versions of labels
were also listed. The experimental data demonstrate that the relative error of the transmission intensity is not more than
0.29%, and that of the SNR is less than 2.0%, which is approximately half the error of the original method.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.