Adaptive Convolution Research Articles

Deep Reference Generation With Multi-Domain Hierarchical Constraints for Inter Prediction

Inter prediction is an important module in video coding for temporal redundancy removal, where similar reference blocks are searched from previously coded frames and employed to predict the block to be coded. Although traditional video codecs can estimate and compensate for block-level motions, their inter prediction performance is still heavily affected by the remaining inconsistent pixel-wise displacement caused by irregular rotation and deformation. In this paper, we address the problem by proposing a deep frame interpolation network to generate additional reference frames in coding scenarios. First, we summarize the previous adaptive convolutions used for frame interpolation and propose a factorized kernel convolutional network to improve the modeling capacity and simultaneously keep its compact form. Second, to better train this network, multi-domain hierarchical constraints are introduced to regularize the training of our factorized kernel convolutional network. For spatial domain, we use a gradually down-sampled and up-sampled auto-encoder to generate the factorized kernels for frame interpolation at different scales. For quality domain, considering the inconsistent quality of the input frames, the factorized kernel convolution is modulated with quality-related features to learn to exploit more information from high quality frames. For frequency domain, a sum of absolute transformed difference loss that performs frequency transformation is utilized to facilitate network optimization from the view of coding performance. With the well-designed frame interpolation network regularized by multi-domain hierarchical constraints, our method surpasses HEVC on average 6.1% BD-rate saving and up to 11.0% BD-rate saving for the luma component under the random access configuration.

Adaptive Convolution Local and Global Learning for Class-Level Joint Representation of Facial Recognition With a Single Sample Per Data Subject

Due to the absence of training samples and intraclass variation, the extraction of discriminative facial features and construction of powerful classifiers have bottlenecks in improving the performance of facial recognition (FR) with a single sample per data subject (SSPDS). In this paper, we propose to learn regional adaptive convolution features that are locally and globally discriminative to facial identity and robust to facial variation. Then, a novel class-level joint representation framework is presented to exploit the distinctiveness and class-level commonality of different facial features. In the proposed class-level joint representation with regional adaptive convolution features (CJR-RACF), both discriminative facial features that are robust to facial variations and powerful representations for classification with generic facial variations have been fully exploited. Furthermore, the gallery discrimination is extracted by our proposed weight-embedded supervision in the training phase (denoted by CJR-RACF w ), which is conducive to more specific features for FR with SSPDS. CJR-RACF and CJR-RACF w have been evaluated on several popular databases, including the large-scale CMU Multi-PIE, LFW, Megaface, and VGGFace datasets. Experimental results demonstrate the much higher robustness and effectiveness of the proposed methods compared to the state-of-the-art methods.

Two-Terminal Fault Location Method of Distribution Network Based on Adaptive Convolution Neural Network

When a single-phase ground fault occurs in a distribution network, it is generally allowed to operate with faults for one to two hours, which may lead to further development of the fault and even threaten the safe operation of the power system. Therefore, when a small current system has a ground fault, it must be quickly diagnosed to shorten the time of operation with fault. In this paper, an adaptive convolutional neural network (ACNN)-based fault line selection method is proposed for a distribution network. This method improves the feature extraction ability of the network by improving the pooling model. Compared with deep belief network (DBN), it can improve the accuracy of fault classification by 7.86% and reduce the training time by 42.7%. On this basis, the secondary fault location is identified using the principle of two-terminal fault location. In this research, fault data obtained by Simulink simulation is used as training set, and ACNN model is built based on TensorFlow framework. The analysis of results proves that the model has a high fault recognition rate and fast convergence speed. It can be used as an auxiliary hand for fault diagnosis in distribution networks.

Independent Target Detection of PolSAR Image Joint Polarimetric and Spatial Features Based on Adaptive Convolution Sparse Representation

Target detection is of great significance for polarimetric SAR (PolSAR) image applications, and independent target detection in a large scene can be considered as a sparse problem. In high-resolution PolSAR, the independent targets generally present a cluster of similar pixels, and therefore, the detection principle should include not only the internal characteristics but also the spatial information. This letter proposes an unsupervised adaptive convolution sparse representation (ACSR) method for PolSAR image independent target detection. The proposed method updates the dictionary to realize adaptive spatial information injection into polarimetric features, and then the independent target can be detected through the iteration. Two sets of high-resolution, full PolSAR images of unmanned aerial vehicle synthetic aperture radar (UAVSAR) system are used to validate the performance of the proposed method. The results indicate the potential of the proposed method in target detection of PolSAR image.

Adaptive Convolution for Object Detection

It is quite challenging to detect objects, especially, small objects, in complex scenes. To solve this problem, we propose a novel module named as adaptive convolution block (ACB), which adaptively adjusts the parameters of convolutional filters according to the current feature maps, and then, filter these feature maps with the obtained adaptive convolutional filters to generate enhanced features. Due to such adaptive convolution, the enhanced features can pay more attention to the concerned objects, suppress the interference information caused by irrelevant surroundings, and efficiently improve the detection accuracy. The proposed ACB is light weight and fast. By directly embedding the ACB into the single shot detection framework, we construct a novel real-time adaptive convolutional detector (ACD). Experiments on PASCAL VOC and MS COCO benchmarks confirm that our ACD outperforms the existing state-of-the-art single-stage detection models, and achieves a better tradeoff between accuracy and speed.

4D Monte Carlo dose calculations for pre-treatment quality assurance of VMAT SBRT: a phantom-based feasibility study

Volumetric arc therapy (VMAT) for lung stereotactic body radiotherapy (SBRT) is challenging due to both breathing-induced motion and the dynamic components of the linear accelerator. In this study, a 4D Monte Carlo (4DMC) dose calculation method for VMAT SBRT is proposed and the feasibility of the method is evaluated. A rigidly-moving lung phantom was imaged using four dimensional computed tomography (4DCT). VMAT SBRT plans were generated on the average intensity projection dataset using the internal target volume (ITV) strategy (ITV-plan) and a single phase to simulate a dynamic treatment-couch tracking technique (TRACKING-plan). 4DMC simulations were performed and compared to 3D Monte Carlo (3DMC) and 3D- and 4D- calculations in the treatment planning system using the adaptive convolution (AC) algorithm. Dose metrics calculated for the ITV-plan showed an overestimation with 3D adaptive convolution (3DAC) for D (GTV) by 3.5% and by 2.0% for 3DMC, both compared to 4DMC. The TRACKING-plan D (GTV) calculated with the 3DAC method overestimated by 2.0% compared with 4DMC. Deviations between the calculation methods for Dmean (Lung) and D (PTV) were minimal. For both plans, measurements were taken with EBT3 film inside the phantom tumour. EBT3 film profiles showed good agreement with 4DMC for the TRACKING-plan giving a gamma pass rate of 97.2% for 3%/3 mm global and for 3DAC compared with measured, 95.8%. Whereas for the ITV-plan, the 3D profiles varied from film in the ITV periphery region with a pass rates of 50% and 48.6% for 3DAC and 3DMC, respectively. 4DMC agreed more closely to measurements for this plan with a pass rate of 95.8%. We have proposed an accurate method to perform 4D dose calculations for pre-treatment quality assurance of VMAT SBRT. The method was compared to experimental measurements and for both plans, 4DMC dose agreed with measurements more closely than other evaluated dose calculation methods. This study has demonstrated the feasibility of this 4DMC method.

New image denoising algorithm via improved deep convolutional neural network with perceptive loss

In order to remove the noise effectively and obtain the high quality denoised image, the new image denoising algorithm is proposed based on improved convolutional neural network with joint loss function (MP-DCNN). Firstly, the end to end adaptive residual convolutional neural network is constructed by using adaptive convolutional and residual learning for modeling the noisy image. Secondly, the input noise image features are extracted by adaptive convolution and leaky rectified linear units (Leaky ReLU) active function in the first half of the network, and then the image features reconstruction is performed by the other structure and convolutional operation of the network. Finally, the initial denoised image is obtained using the mean square error (MSE) loss function which can learn and optimize the extracted feature. Meanwhile, the initial denoising image is input to the pre-trained SegNet which can obtain a better learn of the image edge information. The final denoised image is obtained by using the perceptive loss function. To demonstrate the effectiveness of the proposed image denoising algorithm, the performance of the proposed algorithm is experimentally verified on a variety of noise levels. The experimental results show that the proposed algorithm can obtain superior performance both in visual quality and two evaluation indexes such as peak signal-to-noise ratio (PSNR), structural similarity (SSIM), compared with other state-of-the-art denoising algorithms. Especially for high noisy environment with a noise standard deviation of 60, the algorithm finally obtains 24.26 dB of average PSNR in the test dataset. Its average PSNR is about 0.5 dB and 1 dB higher than DnCNN and BM3D, respectively. At the same time, the proposed method obtain denoising image with no boundary artifact finally. The proposed algorithm can significantly improve the performance of the image denoising system from the perspective of expert and intelligent systems.

Object detection using convolutional networks with adaptively adjusting receptive field of convolutional filter

The receptive field size of a convolutional filter in a deep convolutional network is a crucial issue for object detection task, as the output must response to a suitable size of area in the image to capture proper information. Receptive field size of convolutional filter is fixed due to the inherently fixed geometric structure in its building module. However, objects of interest vary significantly in size within the images for object detection. Different locations of images correspond to objects with different scales, and high level convolutional layers encode semantic features over spatial positions, thus adaptive determination of receptive field size of convolutional filter is desirable for object detection. The authors propose a new module to adaptively determine the receptive field size of convolutional filter, named adaptive convolution. It is based on the idea of dilating the convolutional filter with multiple dilation values and choosing the maximum activation as output, without adding any other parameters. The plain counterparts in existing convolutional neural networks can be easily replaced by adaptive convolution, giving rise to adaptive convolutional networks. Adequate experiments have proven the effectiveness of authors’ method.

Temporal-Spatial Consistency and Sampling Neural Network Based Adaptive Hand-held Video Deblurring

Video deblurring for hand-held camera is vital in many high-level video enhancement applications. Although quite a few proposed approaches have achieved remarkable success in recent years, many technical challenges still prevail for video (from hand-held camera) containing wide-range scenes and highly-variable objects. In particular, we are lacking effective and versatile strategies to adaptively handle high-level alignment, parallax diminution, unclear area detection, sharpness enhancement in a consistent fashion. To ameliorate, this paper develops a novel method to successively respect temporal-spatial alignment and precise deblurring. Our aim is to devise a new adaptive video deblurring technique by resorting to new modeling strategies. This paper's key originality is hinged upon the joint utility of both self-adaptive intrinsic mode functions (IMFs) based on empirical mode decomposition (EMD) in the temporal domain for video signal and mesh-structure constraint enforcement in the spatial domain, beside, this paper innovatively jointly utilities the subsampling convolution, upsampling convolution and luckiness threshold value based back propagation, which takes video deblurring as a high-precision adaptive convolution. As a result, our new approach can adaptively improve the clarity of the video, and synchronously optimize the camera trajectory of wobbly video. To validate our optimization approach for adaptive video deblurring, we conduct comprehensive experiments on public benchmarks, and make extensive and quantitative evaluations with available state-of-the-art methods as well as popular commercial software. All of our experiments demonstrate the advantages of the optimization method in terms of versatility, accuracy, and efficiency.

The Application of A-CNN in Crowd Counting of Scenic Spots

In places where people are concentrated, such as scenic spots, the statistical accuracy of existing crowd statistics algorithms is not enough. In order to solve this problem, a crowd counting algorithm based on adaptive convolution neural network (A-CNN) is proposed, which is based on video monitoring technology. The process of its pooling is dynamically adjusted according to different feature graphs. Then the pooled weights are adjusted adaptively according to the contents of each pooled domain. Therefore, CNN can extract more accurate features when processing different pooled domains under different iteration times, so as to achieve adaptive effect finally. The experimental results show that the proposed A-CNN algorithm has improved the recognition accuracy.

Grasping Point Detection of Randomly Placed Fruit Cluster Using Adaptive Morphology Segmentation and Principal Component Classification of Multiple Features

Precise and rapid grasping point detection based on machine vision is one of the challenging problems in automatic sorting of randomly placed fruit clusters by robot. Grasping stalk of fruit cluster can improve grasping success probability and reduce fruit damage. For the problem that the segmentation of stalk candidates for randomly placed fruit cluster based on existing morphology algorithm tends to be low precision, an improved image segmentation algorithm based on adaptive morphology is proposed. According to edge distances defined based on minimum distance between edge point and unconnected components in minimum domain, the adaptive convolution kernel is constructed. In addition, a run analysis method with different and unordered labels is designed to reduce calculation time of edge distances. For the problem that it is difficult to describe and classify unconstraint stalk by existing features, an improved region classification algorithm based on principal components of multiple features is proposed. The descriptors based on features of object region are designed and principal components of multiple features are extracted based on variance contribution to improve precision and speed of stalk extraction. The proposed grasping point detection method of randomly placed fruit cluster based on improved morphology image segmentation and region classification algorithms is verified by experiments with grape clusters based on parallel robot sorting system. The results show that, compared with existing methods, the average precisions of segmentation and extraction for stalk increase by 9.89% and 2.17% respectively, the average precision and time of grasping point detection reach 94.50% and 2.01s respectively.

A posterior evaluation algorithm of steganalysis accuracy inspired by residual co-occurrence probability

Steganalysis research is committed to distinguishing the steganography media from the normal one correctly. However, individual differences of carriers disturb the detection inevitably and greatly. Existing methods treat all detection results with the same confidence level, or prior accuracy, which may make the prior accuracy overestimate or underestimate the real result. This paper presents a novel performance evaluation method of steganalysis based on posterior accuracy. Adaptive Convolution Feature (ACF) is calculated by the adaptive convolution, then a quantitative value S based on the ACF is modeled to posterior testify and estimate the detection confidence of the image under test. By clustering of carrier noise, we classify the images which we believe they have a similar confidence of the same cluster. The distribution of ACF from each cluster indicates that the S value works well for confidence evaluation. The experimental results show that the S value can distinguish and identify high-confidence samples from the low-confidence, which will greatly improve the real performance of steganalysis. Besides, it improves the steganalysis accuracy of the whole image set by matching S values between training and prediction samples.

An Adaptive Scheme to Achieve Fine Grained Video Scaling

<p>A robust Adaptive Reconstruction Error Minimization Convolution Neural Network (<strong> ARemCNN</strong>) architecture introduced to provide high reconstruction quality from low resolution using parallel configuration. Our proposed model can easily train the bulky datasets such as YUV21 and Videoset4.Our experimental results shows that our model outperforms many existing techniques in terms of PSNR, SSIM and reconstruction quality. The experimental results shows that our average PSNR result is 39.81 considering upscale-2, 35.56 for upscale-3 and 33.77 for upscale-4 for Videoset4 dataset which is very high in contrast to other existing techniques. Similarly, the experimental results shows that our average PSNR result is 38.71 considering upscale-2, 34.58 for upscale-3 and 33.047 for upscale-4 for YUV21 dataset.</p>

The characteristics of dose at mass interface on lung cancer Stereotactic Body Radiotherapy (SBRT) simulation

In lung cancer cases, there exists a difficulty for the Treatment Planning System (TPS) to predict the dose at or near the mass interface. This error prediction might influence the minimum or maximum dose received by lung cancer. In addition to target motion, the target dose prediction error also contributes in the combined error during the course of treatment. The objective of this work was to verify dose plan calculated by adaptive convolution algorithm in Pinnacle3 at the mass interface against a set of measurement. The measurement was performed using Gafchromic EBT 3 film in static and dynamic CIRS phantom with amplitudes of 5 mm, 10 mm, and 20 mm in superior-inferior motion direction. Static and dynamic phantom were scanned with fast CT and slow CT before planned. The results showed that adaptive convolution algorithm mostly predicted mass interface dose lower than the measured dose in a range of −0,63% to 8,37% for static phantom in fast CT scanning and −0,27% to 15,9% for static phantom in slow CT scanning. In dynamic phantom, this algorithm was predicted mass interface dose higher than measured dose up to −89% for fast CT and varied from −17% until 37% for slow CT. This interface of dose differences caused the dose mass decreased in fast CT, except for 10 mm motion amplitude, and increased in slow CT for the greater amplitude of motion.

Joint and collaborative representation with local adaptive convolution feature for face recognition with single sample per person

With the aid of a universal facial variation dictionary, sparse representation based classifier (SRC) has been naturally extended for face recognition (FR) with single sample per person (SSPP) and achieved promising performance. However, extracting discriminative facial features and building powerful representation framework for classifying query face images are still the bottlenecks of improving the performance of FR with SSPP. In this paper, by densely sampling and sparsely detecting facial points, we extract complete and robust local regions and learn convolution features adaptive to the local regions and discriminative to the face identity by using convolutional neural networks (CNN). With this powerful facial description and a generic face dataset with common facial variations, a joint and collaborative representation framework, which performs representation for each local region of the query face image while requires all regions of the query face image to have similar representation coefficients, is presented to exploit the distinctiveness and commonality of different local regions. In the proposed joint and collaborative representation with local adaptive convolution feature (JCR-ACF), both discriminative local facial features that are robust to various facial variations and powerful representation dictionaries of facial variations that can overcome the small-sample-size problem are fully exploited. JCR-ACF has been extensively evaluated on several popular databases including AR, CMU Multi-PIE, LFW and the large-scale CASIA-WebFace databases. Experimental results demonstrate the much higher robustness and effectiveness of JCR-ACF to complex facial variations compared to the state-of-the-art methods.

SU‐F‐T‐273: Using a Diode Array to Explore the Weakness of TPS DoseCalculation Algorithm for VMAT and Sliding Window Techniques

Purpose:To identify the weakness of dose calculation algorithm in a treatment planning system for volumetric modulated arc therapy (VMAT) and sliding window (SW) techniques using a two‐dimensional diode array.Methods:The VMAT quality assurance(QA) was implemented with a diode array using multiple partial arcs that divided from a VMAT plan; each partial arc has the same segments and the original monitor units. Arc angles were less than ± 30°. Multiple arcs delivered through consecutive and repetitive gantry operating clockwise and counterclockwise. The source‐toaxis distance setup with the effective depths of 10 and 20 cm were used for a diode array. To figure out dose errors caused in delivery of VMAT fields, the numerous fields having the same segments with the VMAT field irradiated using different delivery techniques of static and step‐and‐shoot. The dose distributions of the SW technique were evaluated by creating split fields having fine moving steps of multi‐leaf collimator leaves. Calculated doses using the adaptive convolution algorithm were analyzed with measured ones with distance‐to‐agreement and dose difference of 3 mm and 3%..Results:While the beam delivery through static and step‐and‐shoot techniques showed the passing rate of 97 ± 2%, partial arc delivery of the VMAT fields brought out passing rate of 85%. However, when leaf motion was restricted less than 4.6 mm/°, passing rate was improved up to 95 ± 2%. Similar passing rate were obtained for both 10 and 20 cm effective depth setup. The calculated doses using the SW technique showed the dose difference over 7% at the final arrival point of moving leaves.Conclusion:Error components in dynamic delivery of modulated beams were distinguished by using the suggested QA method. This partial arc method can be used for routine VMAT QA. Improved SW calculation algorithm is required to provide accurate estimated doses.

SU-F-T-585: A Novel Phantom for Dosimetric Validation of SBRT for Spinal Lesions

Purpose: SBRT is proving to be a very efficacious treatment modality for an increasing number of indications, including spine lesions. We have developed a novel phantom to serve as an end-to-end QA tool for either patient specific QA or commissioning QA of SBRT for spine lesions. Methods: In this feasibility study, we have selected a patient with a single metastatic lesion in the L5 vertebral body. The patient's CT simulation scan was used to develop a VMAT treatment plan delivering 18Gy to at least 90% of the target volume, following the guidelines of RTOG 0631. The treatment plan was developed with the Pinnacle planning system using the adaptive convolution superposition calculation mode. The approved plan was re-calculated using the Monaco planning system. We performed a pseudo-in-vivo study whereby we manufactured two copies of a phantom to the exact shape and anatomy of the patient. The phantom was made from the CT images of the patient using a 3D printer with sub-millimeter accuracy. One phantom was filled with a gel dosimeter and the other was made with two ion chamber inserts to allow us to obtain point dose measurements in the target's center and the spinal cord. Results: The prescribed dose of 18Gy was planned for the target while keeping the maximum spinal cord dose to less than 14Gy in 0.03cc of the cord. The VMAT plan was delivered to both the gel dosimeter filed phantom and the phantom with the ion chambers. The 3D gel dosimetry revealed a very good agreement between the monte carlo and measured point and volumetric dose. Conclusion: A patient like phantom was developed and validated for use as an end-to-end tool of dose verification for SBRT of spine lesions. We found that gel dosimetry is ideally suited to assess positional and dosimetric accuracy in 3D. RTsafe provided the phantoms and the gel dosimeter used for this study

Autonomous crowds tracking with box particle filtering and convolution particle filtering

Autonomous systems such as Unmanned Aerial Vehicles (UAVs) need to be able to recognise and track crowds of people, e.g. for rescuing and surveillance purposes. Large groups generate multiple measurements with uncertain origin. Additionally, often the sensor noise characteristics are unknown but measurements are bounded within certain intervals. In this work we propose two solutions to the crowds tracking problem— with a box particle filtering approach and with a convolution particle filtering approach. The developed filters can cope with the measurement origin uncertainty in an elegant way, i.e. resolve the data association problem. For the box particle filter (PF) we derive a theoretical expression of the generalised likelihood function in the presence of clutter. An adaptive convolution particle filter (CPF) is also developed and the performance of the two filters is compared with the standard sequential importance resampling (SIR) PF. The pros and cons of the two filters are illustrated over a realistic scenario (representing a crowd motion in a stadium) for a large crowd of pedestrians. Accurate estimation results are achieved.

SU‐E‐T‐220: Computational Accuracy of Adaptive Convolution (AC) and Collapsed Cone Convolution (CCC) Algorithms in the Presence of Air Gaps

Purpose:To compare the percentage depth dose (PDD) computational accuracy of Adaptive Convolution (AC) and Collapsed Cone Convolution (CCC) algorithms in the presence of air gaps.Methods:A 30×30×30 cm3 solid water phantom with two 5cm air gaps was scanned with a CT simulator unit and exported into the Phillips Pinnacle™ treatment planning system. PDDs were computed using the AC and CCC algorithms. Photon energy of 6 MV was used with field sizes of 3×3 cm2, 5×5 cm2, 10×10 cm2, 15×15 cm2, and 20×20 cm2. Ionization chamber readings were taken at different depths in water for all the field sizes. The percentage differences in the PDDs were computed with normalization to the depth of maximum dose (dmax). The calculated PDDs were then compared with measured PDDs.Results:In the first buildup region, both algorithms overpredicted the dose for all field sizes and under‐predicted for all other subsequent buildup regions. After dmax in the three water media, AC under‐predicted the dose for field sizes 3×3 and 5×5 cm2 and overpredicted for larger field sizes, whereas CCC under‐predicted for all field sizes. Upon traversing the first air gap, AC showed maximum differences of –3.9%, −1.4%, 2.0%, 2.5%, 2.9% and CCC had maximum differences of −3.9%, −3.0%,–3.1%, −2.7%, −1.8% for field sizes 3×3, 5×5, 10×10, 15×15, and 20×20 cm2 respectively.Conclusion:The effect of air gaps causes a significant difference in the PDDs computed by both the AC and CCC algorithms in secondary build‐up regions. AC computed larger values for the PDDs except at smaller field sizes. For CCC, the size of the errors in prediction of the PDDs has an inverse relationship with respect to field size. These effects should be considered in treatment planning where significant air gaps are encountered.

Wavelet balancing for residual moveouts

Conventional moveout analysis stretches and squeezes traces to increase the coherence of reflected amplitudes in prestack seismic gathers. Higher order residual moveouts require increasingly difficult scans of semblances with extra dimensions or picking from correlations with many local minima. Alternatively, we can model our data with an adaptive convolution that assumes consistent reflectivities at all offsets (or angles). Short, convolutional wavelets can adjust residual moveouts arbitrarily with offset, but slowly with time (or depth). A Gauss-Newton optimization easily inverts this transform by minimizing a least-squares objective function. With estimated and normalized wavelets, we deconvolved the original data to remove phase and spectral distortions that affected more than one reflection. By constraining how slowly wavelets adapt, we retained phase and amplitude changes distinctive to individual reflections. Deconvolution also avoided any explicit smoothing or mixing of amplitudes among traces. Estimated wavelets captured residual coherence and were easier to track visually than individually weak reflections. By adjusting the length and number of independent dynamic wavelets, we can adjust the resolution to the redundancy supported by the data.