Image Rotation Research Articles

Spatio-Temporal Deep Learning-Based Undersampling Artefact Reduction for 2D Radial Cine MRI With Limited Training Data.

In this work we reduce undersampling artefacts in two-dimensional (2D) golden-angle radial cine cardiac MRI by applying a modified version of the U-net. The network is trained on 2D spatio-temporal slices which are previously extracted from the image sequences. We compare our approach to two 2D and a 3D deep learning-based post processing methods, three iterative reconstruction methods and two recently proposed methods for dynamic cardiac MRI based on 2D and 3D cascaded networks. Our method outperforms the 2D spatially trained U-net and the 2D spatio-temporal U-net. Compared to the 3D spatio-temporal U-net, our method delivers comparable results, but requiring shorter training times and less training data. Compared to the compressed sensing-based methods kt-FOCUSS and a total variation regularized reconstruction approach, our method improves image quality with respect to all reported metrics. Further, it achieves competitive results when compared to the iterative reconstruction method based on adaptive regularization with dictionary learning and total variation and when compared to the methods based on cascaded networks, while only requiring a small fraction of the computational and training time. A persistent homology analysis demonstrates that the data manifold of the spatio-temporal domain has a lower complexity than the one of the spatial domain and therefore, the learning of a projection-like mapping is facilitated. Even when trained on only one single subject without data-augmentation, our approach yields results which are similar to the ones obtained on a large training dataset. This makes the method particularly suitable for training a network on limited training data. Finally, in contrast to the spatial 2D U-net, our proposed method is shown to be naturally robust with respect to image rotation in image space and almost achieves rotation-equivariance where neither data-augmentation nor a particular network design are required.

Use of computed tomography volumetric measurements to predict operative techniques in paraesophageal hernia repair.

Despite advances in diagnostic imaging capabilities, little information exists concerning the impact of physical dimensions of a paraesophageal hernia (PEH) on intraoperative decision making. The authors hypothesized that computerized volumetric analysis and multidimensional visualization to measure hiatal defect area (HDA) and intrathoracic hernia sac volume (HSV) would correlate to operative findings and required surgical techniques performed. Using volumetric analysis software (Aquarius iNtuition, TeraRecon, Inc), HDA and HSV were measured in PEH patients with preoperative computerized tomography (CT) scans, and used to predict the likelihood of intraoperative variables. Multidimensional rotation of images enabled visualization of the entire hiatal defect in a plane mimicking the surgeon's view during repair. The intrathoracic hernia sac was outlined producing volume measurements based on a summation of exact dimensions. A total of 213 PEHR patients had preoperative CT imaging, with 14.1% performed emergently. Primary cruroplasty was performed in 89.2%, salvage gastropexy in 10.3%, and diaphragmatic relaxing incisions in 4.2%. Median HDA was 25.7cm2 (IQR17.8-35.6cm2); median HSV was 365.0cm3 (IQR150.0-611.0cm3). Incremental 5cm2 increase in HDA was associated with greater likelihood of presenting emergently (OR 1.27; 95%CI 1.124-1.428, p = 0.0001), incarceration (OR 1.27; 1.074-1.499, p = 0.005), gastric volvulus (OR 1.13; 1.021-1.248, p = 0.02), and requiring either relaxing incision (OR 1.43; 1.203-1.709, p < 0.0001) or salvage gastropexy (OR 1.13; 1.001-1.274, p = 0.04). Similarly, HSV increases of 100cm3 were associated with 23% greater likelihood of emergent repair (CI 1.121-1.353, p < 0.0001), and were more likely to require a relaxing incision (OR 1.18; 1.043-1.339, p = 0.009) or salvage gastropexy (1.19; 1.083-1.312, p = 0.0003). Utilization of CT volumetric measurements is a valuable adjunct in preoperative planning, allowing the surgeon to anticipate complexity of repair and operative approach, as incremental increases in HSV by 100cm3 and HDA by 5cm2 are more likely to require complex techniques or bailout procedures and/or present emergently.

Phase recovery with the optical vortex microscope

In this work we present the principles of recovering the phase shift introduced by the phase sample in an optical vortex scanning microscope. In this microscope a Gaussian beam with an embedded vortex illuminates the sample and then is imaged onto the observation plane of the microscope. When the sample interacts with the vortex beam, its internal structure is modified. In the case of a uniform phase change inside the focused vortex spot, the modification manifests itself as a rotation of the phase image. We show that using a vortex microscope setup we can obtain a high accuracy of the measured phase shift. The second advantage of our system is its very precise tracking of the vortex point.

A Deep Learning Framework for Joint Image Restoration and Recognition

Image restoration and recognition are important computer vision tasks representing an inherent part of autonomous systems. These two tasks are often implemented in a sequential manner, in which the restoration process is followed by a recognition. In contrast, this paper proposes a joint framework that simultaneously performs both tasks within a shared deep neural network architecture. This joint framework integrates the restoration and recognition tasks by incorporating: (i) common layers, (ii) restoration layers and (iii) classification layers. The total loss function combines the restoration and classification losses. The proposed joint framework, based on capsules, provides an efficient solution that can cope with challenges due to noise, image rotations and occlusions. The developed framework has been validated and evaluated on a public vehicle logo dataset under various degradation conditions, including Gaussian noise, rotation and occlusion. The results show that the joint framework improves the accuracy compared with the single task networks.

Optimal Threshold Based Brain Image Fusion for Brain Cancer Detection using Firefly algorithm

In this paper an attempt is made to diagnose brain disease like neoplastic disease, cerebrovascular disease, Alzheimer disease, fatal disease, Sarcoma disease by effective fusion of two images. Two images are fused in three steps: Step 1.Segmentation: The images are segmented on the basis of optimal thresholding; thresholds are optimized with natural inspired firefly algorithm by assuming fuzzy entropy as objective function. Image thresholding is one of the segmentation techniques which is flexible, simple and has less convergence time as compared to others. Step 2: the segmented features are extracted with Scale Invariant Feature Transform (SIFT) algorithm. The SIFT algorithm is good in extracting the features even after image rotation and scaling. Step 3: Finally fusion rules are made on the basis of interval type-2 fuzzy (IT2FL), where uncertainty effects are minimized unlike type-1. The novelty of the proposed work is tested on different benchmark Image fusion data set and proved better in all measuring parameters.

Deep cascaded convolutional models for cattle pose estimation

Cattle pose estimation is a key step analyzing cattle behaviors and evaluating cattle health, hence, greatly significant for intelligent breeding of cattle. Computer vision based on cattle pose automatic estimation techniques were investigated in this paper and three deep cascaded convolutional neural networks models, including the convolutional pose machine model, the stacked hourglass model and the convolutional heatmap regression model were developed to perform robust cattle pose estimation, with RGB images captured under real cattle farm conditions. A cattle image dataset was also constructed for data modeling and method evaluation, which contains 2134 images of 33 dairy cattle and 30 beef cattle with various poses under natural conditions. In order to enlarge the dataset and decrease the chances of overfitting, data augmentation techniques such as image horizontal flip, rotation and color conversion were used in the data training process of these deep models. The experimental results show that the stacked hourglass model has achieved better performance than the other two, reaching a 90.39% PCKh mean score at the threshold of 0.5 for 16 joints. The high detection accuracy would make the model a very helpful tool for the subsequent cattle behavior recognition and understanding.

Modeling global geometric spatial information for rotation invariant classification of satellite images.

The classification of high-resolution satellite images is an open research problem for computer vision research community. In last few decades, the Bag of Visual Word (BoVW) model has been used for the classification of satellite images. In BoVW model, an orderless histogram of visual words without any spatial information is used as image signature. The performance of BoVW model suffers due to this orderless nature and addition of spatial clues are reported beneficial for scene and geographical classification of images. Most of the image representations that can compute image spatial information as are not invariant to rotations. A rotation invariant image representation is considered as one of the main requirement for satellite image classification. This paper presents a novel approach that computes the spatial clues for the histograms of BoVW model that is robust to the image rotations. The spatial clues are calculated by computing the histograms of orthogonal vectors. This is achieved by calculating the magnitude of orthogonal vectors between Pairs of Identical Visual Words (PIVW) relative to the geometric center of an image. The comparative analysis is performed with recently proposed research to obtain the best spatial feature representation for the satellite imagery. We evaluated the proposed research for image classification using three standard image benchmarks of remote sensing. The results and comparisons conducted to evaluate this research show that the proposed approach performs better in terms of classification accuracy for a variety of datasets based on satellite images.

Face perception inherits low-level binocular adaptation.

In previous work (May & Zhaoping, 2016; May, Zhaoping, & Hibbard, 2012), we have provided evidence that the visual system efficiently encodes binocular information using separately adaptable binocular summation and differencing channels. In that work, binocular test stimuli delivered different grating patterns to the two binocular channels; selective adaptation of one of the binocular channels made participants more likely to see the other channel's grating pattern. In the current study, we extend this paradigm to face perception. Our test stimuli delivered different face images to the two binocular channels, and we found that selective adaptation of one binocular channel biased the observer to perceive the other channel's face image. We show that the perceived identity, gender, emotional expression, or direction of 3-D rotation of a facial test image can be influenced by pre-exposure to binocular random-noise patterns that contain no meaningful spatial structure. Our results provide compelling evidence that face-processing mechanisms can inherit adaptation from low-level sites. Our adaptation paradigm targets the low-level mechanisms in such a way that any response bias or inadvertent adaptation of high-level mechanisms selective for face categories would reduce, rather than produce, the measured effects of adaptation.

SAR ATR with full‐angle data augmentation and feature polymerisation

Utilising neural networks to learn and extract valuable features has achieved satisfactory performance for synthetic aperture radar automatic target recognition (SAR ATR). However, such target recognition capability could be seriously limited by severe image rotation. To greatly improve the performance of convolutional neural networks-based SAR ATR, a data augmentation method combining region of interest (ROI) extraction and full-angle rotation method is firstly proposed in this study. Then, an inception-SAR NET is presented to polymerise multi-branch feature maps. The superior performance of inception-SAR NET structures is obtained by comprehensive experiments. Finally, the results based on MSTAR dataset demonstrate that authors’ methods could achieve the most advanced performance.

Weld image deep learning-based on-line defects detection using convolutional neural networks for Al alloy in robotic arc welding

Accurate on-line weld defects detection is still challenging for robotic welding manufacturing due to the complexity of weld defects. This paper studied deep learning–based on-line defects detection for aluminum alloy in robotic arc welding using Convolutional Neural Networks (CNN) and weld images. Firstly, an image acquisition system was developed to simultaneously collect weld images, which can provide more information of the real-time weld images from different angles including top front, top back and back seam. Then, a new CNN classification model with 11 layers based on weld image was designed to identify weld penetration defects. In order to improve the robustness and generalization ability of the CNN model, weld images from different welding current and feeding speed were captured for the CNN model. Based on the actual industry challenges such as the instability of welding arc, the complexity of the welding environment and the random changing of plate gap condition, two kinds of data augmentation including noise adding and image rotation were used to boost the CNN dataset while parameters optimization was carried out. Finally, non-zero pixel method was proposed to quantitatively evaluate and visualize the deep learning features. Furthermore, their physical meaning were clearly explained. Instead of decreasing the interference from arc light as in traditional way, the CNN model has taken full use of those arc lights by combining them in a various way to form the complementary features. Test results shows that the CNN model has better performance than our previous work with the mean classification accuracy of 99.38%. This paper can provide some guidance for on-line detection of manufacturing quality in metal additive manufacturing (AM) and laser welding.

Dependence of the Stereokinetic Effect on the Rotational Speed and Eccentricity of the Test Image in Children with Partial Optic Atrophy

We studied the influence of the rotational speed and eccentricity of the ring image on the severity of the stereokinetic effect (SE) in a group of 36 children aged 8 to 16 years with partial optic atrophy and in the control group of 44 age-matched children with a normal fundus. Four variants of the ring image with the eccentricity of 0.2, 0.4, 0.6 or 0.8 were used as a test image. The rotational speed of the ring image was 2, 10, 30, 60 or 90 rpm. It was found that the stereokinetic effect was absent in the majority of children (63.9%) with optic atrophy at a combination of eccentricity of 0.2 and a rotational speed of 2 rpm, in contrast to the control group, where the stereokinetic effect was only absent in 9.1% of children under the same conditions of stimulus presentation. The stereokinetic effect was observed with all eccentricities when the speed of rotation of the test image was 30–90 rpm in children with optic atrophy and 10–90 rpm in children from the control group. The number of children with alternate perception of the virtual cone and the virtual funnel was significantly higher in the control group than in the group of children with optic atrophy at all the test image eccentricities and rotational speeds used. The binocular figure size estimates were equal to the monocular estimates in the group of children with optic atrophy for all the image eccentricities and rotation speeds in contrast to the control group with a general tendency for the virtual cone size estimate to increase under the monocular conditions of observation. An increase in the test image rotation speed from 2 to 60 rpm was accompanied by an increase in the size of a virtual figure and a significant decrease in the duration of alternate dominance of virtual cone and virtual funnel perception in both groups of subjects with all the eccentricities used. The difference between the results obtained in the group of children with POA and in the control group may be explained by reduction of afferent visual information flow. These differences are partially offset by increased speed and increased eccentricity of the test image.

A Novel Zero-Watermarking Based Scheme for Copyright Protection of Gray scale Images

Zero-watermarking of digital images is a powerful method with respect to transparency in the watermarked image. However, robustness is still a challenging characteristic for researchers. The proposed method of zero-watermarking provides a novel solution for increasing robustness by obtaining resident features of gray scale image that are robust against common signal processing operations. The proposed solution is based on image scanning to produce NDD (Neighboring Distance Difference) profile. This scheme is used to extract image features for generating redundancy binary profile with the help of image scanning and identification of robust image areas for embedding a binary watermark. Redundant areas from binary profile show perceptually insignificant regions of gray scale image according to human visual system. Resident features from robust areas of image are collected to generate the zero watermarking binary key image using reversible XOR operation. The binary key is used for extraction of binary watermark. Experimental results of the proposed method have been compared with the results of various zero-watermarking schemes as well as traditional watermarking methods and found much better at slightly higher computational cost. The comparison analysis for testing robustness has been carried out against image processing attacks like Gaussian filtering, block average filtering, motion blur filtering, image resizing, image rotation, image compression and cropping. For each attack maximum correlated watermark from the set of recovered watermarks is selected to evaluate the performance of proposed zero-watermarking scheme. It has been recorded that perfect matching is observed between original and extracted watermarks for a number of signal processing attacks.

Fusing learned representations from Riesz Filters and Deep CNN for lung tissue classification.

A novel method to detect and classify several classes of diseased and healthy lung tissue in CT (Computed Tomography), based on the fusion of Riesz and deep learning features, is presented. First, discriminative parametric lung tissue texture signatures are learned from Riesz representations using a one-versus-one approach. The signatures are generated for four diseased tissue types and a healthy tissue class, all of which frequently appear in the publicly available Interstitial Lung Diseases (ILD) dataset used in this article. Because the Riesz wavelets are steerable, they can easily be made invariant to local image rotations, a property that is desirable when analyzing lung tissue micro-architectures in CT images. Second, features from deep Convolutional Neural Networks (CNN) are computed by fine-tuning the Inception V3 architecture using an augmented version of the same ILD dataset. Because CNN features are both deep and non-parametric, they can accurately model virtually any pattern that is useful for tissue discrimination, and they are the de facto standard for many medical imaging tasks. However, invariance to local image rotations is not explicitly implemented and can only be approximated with rotation-based data augmentation. This motivates the fusion of Riesz and deep CNN features, as the two techniques are very complementary. The two learned representations are combined in a joint softmax model for final classification, where early and late feature fusion schemes are compared. The experimental results show that a late fusion of the independent probabilities leads to significant improvements in classification performance when compared to each of the separate feature representations and also compared to an ensemble of deep learning approaches.

Projection-domain scatter correction for cone beam computed tomography using a residual convolutional neural network.

Scatter is a major factor degrading the image quality of cone beam computed tomography (CBCT). Conventional scatter correction strategies require handcrafted analytical models with ad hoc assumptions, which often leads to less accurate scatter removal. This study aims to develop an effective scatter correction method using a residual convolutional neural network (CNN). A U-net based 25-layer CNN was constructed for CBCT scatter correction. The establishment of the model consists of three steps: model training, validation, and testing. For model training, a total of 1800 pairs of x-ray projection and the corresponding scatter-only distribution in nonanthropomorphic phantoms taken in full-fan scan were generated using Monte Carlo simulation of a CBCT scanner installed with a proton therapy system. An end-to-end CNN training was implemented with two major loss functions for 100 epochs with a mini-batch size of 10. Image rotations and flips were randomly applied to augment the training datasets during training. For validation, 200 projections of a digital head phantom were collected. The proposed CNN-based method was compared to a conventional projection-domain scatter correction method named fast adaptive scatter kernel superposition (fASKS) method using 360 projections of an anthropomorphic head phantom. Two different loss functions were applied for the same CNN to evaluate the impact of loss functions on the final results. Furthermore, the CNN model trained with full-fan projections was fine-tuned for scatter correction in half-fan scan by using transfer learning with additional 360 half-fan projection pairs of nonanthropomorphic phantoms. The tuned-CNN model for half-fan scan was compared with the fASKS method as well as the CNN-based method without the fine-tuning using additional lung phantom projections. The CNN-based method provides projections with significantly reduced scatter and CBCT images with more accurate Hounsfield Units (HUs) than that of the fASKS-based method. Root mean squared error of the CNN-corrected projections was improved to 0.0862 compared to 0.278 for uncorrected projections or 0.117 for the fASKS-corrected projections. The CNN-corrected reconstruction provided better HU quantification, especially in regions near the air or bone interfaces. All four image quality measures, which include mean absolute error (MAE), mean squared error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM), indicated that the CNN-corrected images were significantly better than that of the fASKS-corrected images. Moreover, the proposed transfer learning technique made it possible for the CNN model trained with full-fan projections to be applicable to remove scatters in half-fan projections after fine-tuning with only a small number of additional half-fan training datasets. SSIM value of the tuned-CNN-corrected images was 0.9993 compared to 0.9984 for the non-tuned-CNN-corrected images or 0.9990 for the fASKS-corrected images. Finally, the CNN-based method is computationally efficient - the correction time for the 360 projections only took less than 5s in the reported experiments on a PC (4.20GHz Intel Core-i7 CPU) with a single NVIDIA GTX 1070 GPU. The proposed deep learning-based method provides an effective tool for CBCT scatter correction and holds significant value for quantitative imaging and image-guided radiation therapy.

Automated Detection of Lunar Rockfalls Using a Convolutional Neural Network

This paper implements a novel approach to automatically detect and classify rockfalls in Lunar Reconnaissance Orbiter narrow angle camera (NAC) images using a single-stage dense object detector (RetinaNet). The convolutional neural network has been trained with a data set of 2932 original rockfall images. In order to avoid overfitting, the initial training data set has been augmented during training using random image rotation, scaling, and flipping. Testing images have been labelled by human operators and have been used for RetinaNet performance evaluation. Testing shows that RetinaNet is capable to reach recall values between 0.98 and 0.39, precision values between 1 and 0.25, and average precisions ranging from 0.89 to 0.69, depending on the used confidence threshold and intersection-over-union values. Mean processing time of a single NAC image in RetinaNet is around 10 s using a GeForce GTX 1080 Ti and GeForce Titan Xp, which is in orders of magnitudes faster than a human operator. The processing speed allows to efficiently exploit the currently available NAC data stack with more than 1 million images in a reasonable timeframe. The combination of speed and detection performance can be used to produce lunar rockfall distribution maps on large spatial scales for utilization by the scientific and engineering community.

Transfer learning with pre-trained deep convolutional neural networks for serous cell classification

Serous effusion is a condition of excess accumulation of fluids in serous cavities due to different underlying pathological conditions. The basis of cytopathological assessment of serous effusions is the identification of cells in the fluid based on their morphology and texture. This assessment is a physically and mentally laborious task, and it can also lead to variability among pathologists. In literature, only a small number of feature-based methods are conducted for automated serous cell classification. In this study, a transfer learning with pre-trained deep convolutional neural networks (ConvNets) is proposed to automatically identify 11 different categories of serous cells in effusion cytology. Unlike the methods which rely on the extraction of cellular features such as morphology and texture, this method is an appearance-based machine learning approach. We fine-tuned four pre-trained ConvNet architectures that are AlexNet, GoogleNet, ResNet and DenseNet on the serous cell dataset. To reduce the overfitting effect, we augmented the data by image rotation, translation, and mirroring. The proposed method was evaluated on both original and augmented sets of serous cells derived from a publicly available dataset. Among the four ConvNet models, ResNet and DenseNet obtained the highest accuracies of 93.44% and 92.90%. However, when two models were compared in terms of accuracy and model complexity, ResNet-TL was selected as the best network model. When compared to the results without data augmentation, data augmentation increased the accuracy rate approximately 10%. Results show that higher classification results were achieved than other traditional methods without requiring precise segmentation.

Brain Tumor Segmentation Based on Improved Convolutional Neural Network in Combination with Non-quantifiable Local Texture Feature.

Accurate and reliable brain tumor segmentation is a critical component in cancer diagnosis. According to deep learning model, a novel brain tumor segmentation method is developed by integrating fully convolutional neural networks (FCNN) and dense micro-block difference feature (DMDF) into a unified framework so as to obtain segmentation results with appearance and spatial consistency. Firstly, we propose a local feature to describe the rotation invariant property of the texture. In order to deal with the change of rotation and scale in texture image, Fisher vector encoding method is used to analyze the texture feature, which can combine with the scale information without increasing the dimension of the local feature. The obtained local features have strong robustness to rotation and gray intensity variation. Then, the non-quantifiable local feature is fused to the FCNN to perform fine boundary segmentation. Since brain tumors occupy a small portion of the image, deconvolutional layers are designed with skip connections to obtain a high quality feature map. Compared with the traditional MRI brain tumor segmentation methods, the experimental results show that the segmentation accuracy and stability has been greatly improved. Average Dice index can be up to 90.98%. And the proposed method has very high real-time performance, where brain tumor image can segment within 1s.

Identification and correction of microlens array rotation error in plenoptic imaging systems

Plenoptic imaging systems can acquire spatio-angular information of light rays, allowing complex scene measurements with a single exposure. The measurement accuracy is significantly affected by the orientation parameters of internal components in the system, especially the rotation error between the microlens array (MLA) and the image sensor. In this paper, we formulate an MLA rotation error model to derive the transformations between the error point and its ideal point of ray projection. Furthermore, we present a method for identifying and correcting the rotation error, including error angle estimation, raw image rotation, and refocused image reverse rotation. Through simulation experiments, the imaging characteristics of the MLA rotation error are analyzed and the performance of the proposed model and method are verified. The results demonstrate that the integral distribution of raw image distortion caused by the error is centrosymmetric. In the error range of 0.1–1.0°, the mean square error (MSE) and structural similarity index metric (SSIM) of the corrected refocused images are improved to 0.2146–3.8190 and 0.9986–0.9999, respectively.

Building Damage Detection from Post-Event Aerial Imagery Using Single Shot Multibox Detector

Using aerial cameras, satellite remote sensing or unmanned aerial vehicles (UAV) equipped with cameras can facilitate search and rescue tasks after disasters. The traditional manual interpretation of huge aerial images is inefficient and could be replaced by machine learning-based methods combined with image processing techniques. Given the development of machine learning, researchers find that convolutional neural networks can effectively extract features from images. Some target detection methods based on deep learning, such as the single-shot multibox detector (SSD) algorithm, can achieve better results than traditional methods. However, the impressive performance of machine learning-based methods results from the numerous labeled samples. Given the complexity of post-disaster scenarios, obtaining many samples in the aftermath of disasters is difficult. To address this issue, a damaged building assessment method using SSD with pretraining and data augmentation is proposed in the current study and highlights the following aspects. (1) Objects can be detected and classified into undamaged buildings, damaged buildings, and ruins. (2) A convolution auto-encoder (CAE) that consists of VGG16 is constructed and trained using unlabeled post-disaster images. As a transfer learning strategy, the weights of the SSD model are initialized using the weights of the CAE counterpart. (3) Data augmentation strategies, such as image mirroring, rotation, Gaussian blur, and Gaussian noise processing, are utilized to augment the training data set. As a case study, aerial images of Hurricane Sandy in 2012 were maximized to validate the proposed method’s effectiveness. Experiments show that the pretraining strategy can improve of 10% in terms of overall accuracy compared with the SSD trained from scratch. These experiments also demonstrate that using data augmentation strategies can improve mAP and mF1 by 72% and 20%, respectively. Finally, the experiment is further verified by another dataset of Hurricane Irma, and it is concluded that the paper method is feasible.

The dynamics and distribution of angular momentum in HiZELS star-forming galaxies atz = 0.8–3.3

We present adaptive optics assisted integral field spectroscopy of 34 star-forming galaxies at $z$ = 0.8-3.3 selected from the HiZELS narrow-band survey. We measure the kinematics of the ionised interstellar medium on $\sim$1 kpc scales, and show that the galaxies are turbulent, with a median ratio of rotational to dispersion support of $v$/$\sigma$=0.82$\pm$0.13. We combine the dynamics with high-resolution rest-frame optical imaging and extract emission line rotation curves. We show that high-redshift star-forming galaxies follow a similar power-law trend in specific angular momentum with stellar mass as that of local late type galaxies. We exploit the high resolution of our data and examine the radial distribution of angular momentum within each galaxy by constructing total angular momentum profiles. Although the stellar mass of a typical star-forming galaxy is expected to grow by a factor $\sim$8 in the $\sim$5 Gyrs between $z$$\sim$3.3 and $z$$\sim$0.8, we show that the internal distribution of angular momentum becomes less centrally concentrated in this period i.e the angular momentum grows outwards. To interpret our observations, we exploit the EAGLE simulation and trace the angular momentum evolution of star forming galaxies from $z$$\sim$3 to $z$$\sim$0, identifying a similar trend of decreasing angular momentum concentration. This change is attributed to a combination of gas accretion in the outer disk, and feedback that preferentially arises from the central regions of the galaxy. We discuss how the combination of the growing bulge and angular momentum stabilises the disk and gives rise to the Hubble sequence.