This study addresses the issue of blastomere cleavage timing, which is traditionally deemed the most important indicator in embryo quality assessment. We propose a novel quantization method for analyzing Time-lapse Incubator (TLI) images of embryos. This method incorporates five image classifiers, an instance segmentor, and the Expectation-Maximization (EM) algorithm to provide a cost-effective and reliable tool and to offer interpretable data for further research in the reproductive medicine field. The method is characterized in that the EM algorithm is employed to enhance the performance of classifiers and segmentor in fitting different patient data sources. The method enables (1) to identify the absolute cleavage timing of blastomere, i.e., tPNa, tPNf, t2, t3, t4, t8, t6, t7, and t8, (2) tracking the cleavage process of blastomere to highlight the irregular cleavage (IRC) patterns, (3) to quantitatively assess the symmetry degree of blastomere. Our method reduces the average time difference between Ground Truth and CTP to just 2.9 hours, a significant improvement over the 6.13 hours reported by Vitrolife™. Moreover, our method identifies up to eight cleavage time points, surpassing the six cleavage time points provided by Vitrolife™.

This study proposed a paddy rice mapping based on hand-crafted features combined with deep learning methods. In this research, the rice growth-related features were extracted from time-series SAR data provided by C-band Sentinel-1A images. Four rice features were first extracted from the rice growth curve, including Average Normalized Backscatter (ANB), Backscatter Variation Rate (BVR), Time Interval (TI) and Backscatter Difference (BD). Then, the deep learning U-Net model with four combined rice features was used to obtain the mapping distribution of paddy rice-fields. In the study, the experimental areas were composed of two important rice growth counties in central Taiwan, including Yunlin and Changhua counties. The experimental results showed that the detection accuracy of the proposed method for rice and non-rice can achieve 92.3% and 98.2%, respectively. These results demonstrate the great potential of SAR data in mapping paddy fields using the U-Net model with proposed feature inputs.

Ship detection is one of the most important researches in the field of navigation safety and marine environment monitoring. Synthetic aperture radar (SAR) imagery has been used as a promising data source for monitoring maritime activities. However, the resolution of SAR images is limited, and it cannot effectively detect densely distributed and small ships, especially near harbors. In order to effectively manage ships during the day and night, this research uses visible and infrared images for ship detection. In this study, an improved architecture based on You only look once version 3 (Yolov3) is proposed for ship detection. Yolov3 provided multi-scale feature extraction to enhance the recognition of small targets. In addition, the study also considered the influence of Yolov3 parameters on ship detection, including the input image size, the number of filters in convolution layers and the detection scales. The experiment is based on a data set containing six types of ships, and a total of 5, 513 visible and infrared images from the harbors in northern Taiwan. The experimental results show that when the model parameters are selected as: 352x352 input size, the scale of large target deleted and the convolution filters reduced by 30%, Yolov3 has better ship detection performance and computational efficiency. Compared with the original Yolov3 with 87.9% mean average precision (mAP) and 87.0 billion floating point operations per second (BFLOPs), the proposed architecture can achieve 89.1 % mAP and 24.3 BFLOPs.

This study presents a novel method based on entropy‐driven gradient evaluation ( P‐EdGE ) for detecting perceptual edges that represent boundaries of objects as perceived by human eyes. P‐EdGE is characterised by iteratively employing a shape‐changeable mask centred at a target pixel to sample gradient orientations of neighbouring pixels for measuring the directivity of the target pixel. The mask deforms to automatically cover pixels most suitable for exhibiting the directivity of the target pixel. The authors show that such an iterative scheme satisfies the similarity and proximity laws in Gestalt theory. The converged directivity in conjunction with the gradient magnitude are subjected to a Bayesian process, they show that doing so can help effectively determine whether the target pixel belongs to a perceptual edge. Experimental results are presented to justify the superiority of P‐EdGE over other methods in detecting perceptual edges.

This paper presents an eavesdropper-proof algorithm that is capable of fast generating symmetric (secret) keys. Instead of literally exchanging secret keys, both the sender and receiver adopt a mirroring process based on an improved Hebbian rule that uses identical random inputs to separately train on their reciprocal outputs to generate a pair of exactly identical secret key strings. Important parameters are elaborately characterized to implement a fast information transmission for ephemeral key exchanging. We show that performance optimization can be achieved by coordinating the parameters. One essential feature of the proposed algorithm is that even an eavesdropper who acquires entire structure of the algorithm and the transmission data still has no chance to decrypt the encrypted message, thus ensuring security in the subsequent encryption task. Moreover, computation load is well bounded in an acceptable range despite the increasing key length.

This paper presents a novel approach which incorporates dimension extension and generalized inverse transformation (DEGIT) to realize data clustering. Unlike k-means algorithm, DEGIT needs not pre-specify the number of clusters k, centroid locations are updated and redundant centroids eliminated automatically during iterative training process. The essence of DEGIT is that clustering is performed by generalized inverse transforming the input data such that each data point is represented by a linear combination of bases with extended dimension, with each basis corresponding to a centroid and its coefficient representing the closeness between the data point and the basis. Issue of clustering validation is also addressed in this paper. First, principal component analysis is applied to detect if there exists a dominated dimension, if so, the original input data will be rotated by a certain angle w.r.t. a defined center of mass, and the resulting data undergo another run of iterative training process. After plural runs of rotation and iterative process, the labeled results from various runs are compared, a data point labeled to a centroid more times than others will be labeled to the class indexed by that wining centroid.

This paper presents an improved multi-object segmentation algorithm. First, a critical look is focused on utilizing vector calculus operator and combinational operator to rewrite Dirichlet integral into a matrix form, and boundary condition is defined to obtain the needed harmonic function, from which a set of probabilistic values for each pixel are calculated and the maximum is used to label the pixel accordingly. The only unique parameter that dominantly affects the segmentation performance is characterized, and the result of which is used to derive a formula that adjusts the value of the unique parameter according to intensity difference between neighboring pixels. Furthermore, a pre-process involving the use of watershed analysis is applied to smooth the effect of high frequency components in the input image, so that better noise tolerance and more accurate object contours can be obtained.

This paper presents an improved snake model (ISM) effective for performing fast image segmentation. The work takes advantage of two well-known snake models of Balloon (Cohen, 1991) and GGVF (generalized gradient vector flow) (Xu and prince, 1998). The Balloon model is well known for its capability of fast locating object boundary using an extra pressure force field, but it may easily cause the contour to overwhelm the boundary. The more accurate GGVF model, however, requires lengthy computation time due to the need to specify a relatively large range in order to cover all the control points inside the range. Our goal is to obtain the exact field provided by GGVF while at the same time to expand the capture range with the Balloon pressure force. To this end, in ISM the force fields of Balloon and GGVT are integrated and a dynamic scheme for setting the control points is employed to avoid the time-consuming "re-sampling process" in both GVF (gradient vector flow) and GGVF. The key attribute of ISM is that it can reduce the number of training iterations while maintaining the capability of pushing the capture range toward the gradient map. Empirical results show that the proposed model can deliver the same performance level of segmentation using less computation time than "GGVF" does.

In this paper, a novel method called double-layer infiltration (DLI) is presented to perform fast region-based image segmentation. The method mainly comprises two successive processes, namely updating the feature of each region followed by merging the immediate regions as their updated features satisfy a local statistic criterion. The proposed segmentation method is fast as features of all regions are updated simultaneously and accurately effective as regions are merged by considering not only the immediate adjacent layer but also the regions next to the adjacent layer. The method starts with an initial set of seed regions until a well-defined terminal condition is satisfied. DLI is self-terminated and needs not pre-specify the final number of regions. Experimental results on real images are provided to verify the effectiveness of the method.

This paper presents a novel network, called scale equalized higher order neural network (SEHNN) based on concept of scale equalization (SE). We show that SE is particularly useful in alleviating the scale divergence problem that plagues higher order networks. SE comprises two main processes: setting the initial weight vector and conducting the matrix transformation. An illustrative embodiment of SEHNN is built on the Sigma-Pi network (SPN) applied to task of function approximation. Empirical results verify that SEHNN outperforms other higher order networks in terms of computation efficiency. Compared to SPN, and Pi-Sigma network (PSN), SEHNN requires less number of epochs to complete the training process.

This paper presents a neural solution for clustering without pre-specifying the number of clusters. The presented clustering approach incorporates a harmonic neural network (HNN) and a codeword-agglomerating neural network (ANN). The HNN harmonizes criteria of mean squared error and information entropy to exploit the substructure in the input data and find the fittest codewords. Following the HNN, ANN undergoes an unsupervised process that agglomerates the codewords in accordance with the input nature, and the input partition is autonomously determined when the agglomeration operation of ANN converges to a certain number of centroids that represent the clusters.

Image segmentation is a partitioning of an image into constituent parts using attributes such as pixel intensity, spectral values, and/or textural properties. As a key role in several approaches to image compression and image analysis, image segmentation produces an image representation in terms of boundaries and regions of various shapes and interrelationships. This paper presents a new approach called Double Pulse Coupled Neural Network (DPCNN) to perform fast image segmentation. Currently, most segmentation methods merge regions one by one to alleviate the over-segmentation problem. However, sequential merging would inevitably incur lengthy computation time. DPCNN simultaneously updates features of regions by referring to adjacent regions. Due to the use of synchronous update strategy, DPCNN achieves fast merging and provides great potentiality for a fully parallel hardware implementation. The iterative operation terminates when the numbers of regions in consecutive iterations are identical. Empirical results show that DPCNN outperforms other methods in terms of computation efficiency and segmentation accuracy.

This paper presents a method to enhance ultrasonic images that are commonly plagued with a special type of acoustic noise called speckles. To reduce the noise effect, filters such as the weighted median filter, adaptive trimmed mean filter, two dimensional weighted Savitzky-Golay filter (2D-WSGF) have been studied. We have found that 2D-WSGF has the best image restoration quality. Based on the least squares fitting of a polynomial function to image intensities, 2D-WSGF can preserve edges while performing noise reduction. However, 2D-WSGF requires a relatively large fixed mask, which inevitably incurs unnecessary computation cost. In this paper, a variable-mask approach is presented to improve the computation performance and the restored image quality of the 2D-WSGF. Advantage of using variable mask size is that computation time for filtering can be greatly saved at the expense of relatively little computations needed for detecting bumpy regions and estimating the corresponding mask size.

This paper presents a two-stage approach that is effective for performing fast clustering. First, a competitive neural network (CNN) that can harmonize mean squared error and information entropy criteria is employed to exploit the substructure in the input data by identifying the local density centers. A Gravitation neural network (GNN) then takes the locations of these centers as initial weight vectors and undergoes an unsupervised update process to group the centers into clusters. Each node (called gravi-node) in the GNN is associated with a finite attraction radius and would be attracted to a nearby centroid simultaneously during the update process, creating the Gravitation-like behavior without incurring complicated computations. This update process iterates until convergence and the converged centroid corresponds to a cluster. Compared to other clustering methods, the proposed clustering scheme is free of initialization problem and does not need to pre-specify the number of clusters. The two-stage approach is computationally efficient and has great flexibility in implementation. A fully parallel hardware implementation is very possible.

In this paper, we present a novel approach to the restoration of noise-corrupted image, which is particularly effective at removing highly impulsive noise while preserving image details. This is accomplished through a fuzzy smoothing filter constructed from a set of fuzzy membership functions for which the initial parameters are derived in accordance with input histogram. A principle of conservation in histogram potential is incorporated with input statistics to adjust the initial parameters so as to minimize the discrepancy between a reference intensity and the output of defuzzification process. Similar to median filters (MF), the proposed filter has the benefits that it is simple and it assumes no a priori knowledge of specific input image, yet it shows superior performance over conventional filters (including MF) for the full range of impulsive noise probability. Unlike in many neuro-fuzzy or fuzzy-neuro filters where random strategy is employed to choose initial membership functions for subsequent lengthy training, the proposed filter can achieve satisfactory performance without any training.

Over-segmentation is a serious problem in conventional watershed analysis owing to the topographic relief inherent in the input image. Currently in watershed segmentation methods, merging regions one by one is the most well known cure for the over-segmentation problem, K. Haris et al.(1998) proposed a merging algorithm called fast nearest neighbor region merging based on the observation that it is not necessary to keep all RAG in the heap, only a small portion of them is used to construct nearest neighbor graph (NNG). Although the performance of merging is greatly improved, the required computation time is in proportional to the initial number of regions in NNG. In this paper we propose the fuzzy-based feature tuning (FFT) algorithm that can simultaneously adjust ?/sub i/ of all region by referencing their adjacent neighboring regions, where ?/sub i/ is defined as the average gray value over region i.

The authors present a novel clustering approach called C-spheres Agglomeration (CA) that need not pre-specify the number of clusters; the input partition is autonomously determined by the input nature. The input data set first undergoes a preprocess and is partitioned into arbitrarily specified M/sub f/ subsets, each subset contains a portion of input items represented by a unique prototype c/sub j/, j=1,2,.. M/sub f/. Following the pre-process, the set {c/sub j/, j=1,2,.. M/sub f/} undergoes an iterative agglomeration process to extract clusters. Our idea to clustering is to treat each prototype c/sub j/ as a neutral particle in the feature space with mass m/sub j/ = number of input vectors represented by c/sub j/. Also associated with each c/sub j/ is a C-sphere, the radius of which is defined in accordance with a local density measure around c/sub j/. Prototypes are considered as neighbors if their C-spheres intersect. During one agglomeration iteration, every C-sphere is required to move directly to the corresponding centroid of neighborhoods.

A self-organizing mountain method (SOMM) is presented. SOMM incorporates the Possibilistic C-Means (PCM) technique and the concept of the mountain method to perform clustering. By self-organizing we mean that parameters are data-driven, the terrain of each cluster (or mountain) is estimated, the precise center of each cluster and the terminating condition are determined by the input nature. In addition, SOMM is robust even when a large number of outliers/noises is presented. The simulation results show that the robust clustering can be obtained for various Gaussian clusters and uniform clusters, respectively.

The paper proposes an algorithm to solve the difficult problem of stereo correspondence in stereo vision. This problem can be considered as an optimization process or a constraint satisfaction process. It is found that in the course of the constraint satisfaction process, the more constraints the cost function specifies, the more accurate the result will be. The authors propose a similarity constraint coupled with the uniqueness, smoothing, ordering constraints to formulate the energy function of a Hopfield network to perform the optimum task. Simulation results have verified the effectiveness of our approach.

We describe a method for object segmentation that uses a modified version of the anisotropic diffusion algorithm. The original anisotropic diffusion is performed in the original image, and it is not suitable for texture or noisy images. Our method achieves accurate object detection by using anisotropic diffusion in another 2D space whose size is the same as the processed image. This space is called energy space. The idea is simple. The energy will center on the pixels in the energy space corresponding to the edge pixels of the real image. It can simplify the original overly complex schemes to save computation time. The authors also propose a novel edge linking method by edge directional vector based on the energy space. By experimental evidence, our method can increase the noise tolerance.

Topology preservation is mainly used to analyze the structure of an input distribution. In some implementations, it refers to a data visualization process by means of which high-dimensional input data can be mapped onto a lower-dimensional space where the spatial features of the original input data can be visually revealed. In this paper, we propose a powerful topology-preserving method based on a self-creating model called the harmonic competitive neural network (HCNN). The HCNN is initialized as a triangular structure (i.e. three nodes connected to each other), as in the growing cell structure (GCS) of B. Fritzke (1994). In order to approximate the input distribution in a self-organizing manner, the training parameters are data-driven and the network size does not need to be pre-specified. Our goal is to map the topological structure of input data with less distortion error and lower computational cost in comparison with other networks, such as self-organizing feature maps (SOFMs) or topology-representing networks (TRNs).

A novel histogram manipulation technique, which re-assigns pixels of three sub-images based on a divided histogram, is presented. We use k-means to obtain two cluster centroids from the input image histogram, and then two thresholds are deduced to divide the input image into three sub-images. After mathematical morphology, connected components are obtained from sub-images separately. The proposed novel histogram manipulation technique applies, separately, histogram equalization to the three sub-images that were obtained by decomposing the input image based on two thresholds in the original histogram.

This paper optimizes the performance of the growing cell structures (GCS) model in learning topology and vector quantization. Each node in GCS is attached with a resource counter. During the competitive learning process, the counter of the best-matching node is increased by a defined resource measure after each input presentation, and then all resource counters are decayed by a factor alpha. We show that the summation of all resource counters conserves. This conservation principle provides useful clues for exploring important characteristics of GCS, which in turn provide an insight into how the GCS can be optimized. In the context of information entropy, we show that performance of GCS in learning topology and vector quantization can be optimized by using alpha=0 incorporated with a threshold-free node-removal scheme, regardless of input data being stationary or nonstationary. The meaning of optimization is twofold: (1) for learning topology, the information entropy is maximized in terms of equiprobable criterion and (2) for leaning vector quantization, the use is minimized in terms of equi-error criterion.

This paper presents a self-creating neural network in which a conservation principle is incorporated with the competitive learning algorithm to harmonize equi-probable and equi-distortion criteria. Each node is associated with a measure of vitality which is updated after each input presentation. The total amount of vitality in the network at any time is 1, hence the name conservation. Competitive learning based on a vitality conservation principle is near-optimum, in the sense that problem of trapping in a local minimum is alleviated by adding perturbations to the learning rate during node generation processes. Combined with a procedure that redistributes the learning rate variables after generation and removal of nodes, the competitive conservation strategy provides a novel approach to the problem of harmonizing equi-error and equi-probable criteria. The training process is smooth and incremental, it not only achieves the biologically plausible learning property, but also facilitates systematic derivations for training parameters. Comparison studies on learning vector quantization involving stationary and nonstationary, structured and nonstructured inputs demonstrate that the proposed network outperforms other competitive networks in terms of quantization error, learning speed, and codeword search efficiency.

This paper proposes an improved clustering method based on a neural network. The improved method does need not pre-specify the number of clusters. In order to obtain more accurate and computation-efficient clustering results, the proposed method adaptively computes the optimal threshold for each cluster separately, instead of calculating minimum spanning tree for determining a fixed global threshold. Simulation results show that in terms of accuracy and computation time, the proposed method provides superior performance to that of the traditional k-means method.

In this paper, we propose a novel method of encoding an image without blocky effects. The method incorporates the wavelet transform and a self-development neural network-the Vitality Conservation (VC) network to achieve significant improvement in image compression performance. The implementation consists of three steps. The image is first decomposed at different scales using wavelet transform to obtain an orthogonal wavelet representation of the image. Each band can be subsequently processed in parallel. In the second step, the discrete Karhunen-Loeve transform is used to extract the principal components of the wavelet coefficients. Thus, the processing speed can be much faster than otherwise. Finally, results of the second step are used as input to the VC network for vector quantization. Our simulation results show that such an implementation can, in much less time, achieve superior reconstructed images to other methods.

The existence of principal interconnections useful in solving the proliferation problem in higher order Hebbian-type associative memories is introduced. Among all legal interconnections, the author proves there exists a subset T/sub pr/ that carries more information than the others. Regardless of the network order p, the elements in T/sub pr/ are shown to be those interconnections T that fall within the range of /spl radic/m, /spl les/|T|/spl les/2/spl radic/m, where m/sub s/ equals the number of stored code words. Memories that use only T/sub pr/ can maintain original generalization performance, using less than 50 percent of the total number of interconnections.

A statistical method is applied to explore the characteristics of a certain class of quadratic order associative memories: the Synchronous Update Direct Convergence Memory (SQDM). The initial input vector is required to converge, with a given value of probability Pdc, to a stored codeword in just one synchronous update. The memory capacity ms is derived by a figure of merit N phi, and its tight asymptotic bound is found by using Tchebycheff's inequality. When the input contains erroneous bits, the maximum allowable number of attractors and their attraction radii are determined. The existence of principal connections Tpr useful in solving the problem of proliferation of connections in the SQDM is introduced. We proved that Tpr is a set of which an arbitrary connection Tijk satisfies square root of ms < or = [Tijk[ < or = 2 square root of ms.

A statistical method is applied to explore the unique characteristics of a certain class of neural network autoassociative memory with N neurons and first-order synaptic interconnections. The memory matrix is constructed to store M = αN vectors based on the outer-product learning algorithm. We theoretically prove that, by setting all the diagonal terms of the memory matrix to be M and letting the input error ratio ρ = 0, the probability of successful recall Pr steadily decreases as α increases, but as α increases past 1.0, Pr begins to increase slowly. When 0 < ρ ≤ 0.5, the network exhibits strong error-correction capability if α ≤ 0.15 and this capability is shown to rapidly decrease as α increases. The network essentially loses all its error-correction capability at α = 2, regardless of the value of ρ. When 0 < ρ ≤ 0.5, and under the constraint of Pr > 0.99, the tradeoff between the number of stable states and their attraction force is analyzed and the memory capacity is shown to be 0.15N at best.

Statistical parameters that can be used to estimate the convergence probability of arbitrary-order Hebbian-type neural network associative memories (HAMs) withNneurons andMstored patterns are developed. The principle involves using two figures of merit, ε/η and ηN, to determine the convergence probability for indirect (iterative) convergence and direct (one-step) convergence HAMs. Given η, the probability that a neuron changes to an incorrect bit after one update, the parameter ε/η determines the capability of converging iteratively to at most εNbits away from the stored vector after a stable state is reached, where 0Pic≈1.0 for all HAMs having ε/η>20. If precise convergence to the stored vector is required in one step, the parameter ηNis used to determine the probability of direct convergence,Pdc