Data augmentation has improved the accuracy and robustness of deep neural networks. Research has focused on finding an optimal augmentation policy that generates good quality training images to improve classification accuracy. However, searching for this optimal augmentation policy is computationally expensive and is dependent on the neural architecture. In this work, we design a dynamic augmentation approach that automatically adjusts the number of transformation operations and their magnitudes during the training of deep neural networks. We also address the shift in the test data distribution by proposing to perform augmentation on the test data. We validate the effectiveness of our solution on CIFAR-10, CIFAR-100, ImageNet, and the perturbed datasets including CIFAR-10-C, CIFAR-100-C, ImageNet-A, ImageNet-C and ImageNet-P. Experiment results show that our proposed dynamic augmentation approach is scalable and gives good performances on clean, adversarial and corrupt datasets, reducing the best published results by a significant margin.

Ischemic stroke is a leading cause of death and long-term disability that is difficult to predict reliably. Retinal fundus photography has been proposed for stroke risk assessment, due to its non-invasiveness and the similarity between retinal and cerebral microcirculations, with past studies claiming a correlation between venular caliber and stroke risk. However, it may be that other retinal features are more appropriate. In this paper, extensive experiments with deep learning on six retinal datasets are described. Feature isolation involving segmented vascular tree images is applied to establish the effectiveness of vessel caliber and shape alone for stroke classification, and dataset ablation is applied to investigate model generalizability on unseen sources. The results suggest that vessel caliber and shape could be indicative of ischemic stroke, and sourcespecific features could influence model performance.

Recent deep neural networks (DNN) utilize identity mappings involving either element-wise addition or channel-wise concatenation for the propagation of these identity mappings. In this paper, we propose a new propagation mechanism called channel-wise addition (cAdd) to deal with the vanishing gradients problem without sacrificing the complexity of the learned features. Unlike channel-wise concatenation, cAdd is able to eliminate the need to store feature maps thus reducing the memory requirement. The proposed cAdd mechanism can deepen and widen existing neural architectures with fewer parameters compared to channel-wise concatenation and element-wise addition. We incorporate cAdd into state-of-the-art architectures such as ResNet, WideResNet, and CondenseNet and carry out extensive experiments on CIFAR10, CIFAR100, SVHN and ImageNet to demonstrate that cAdd-based architectures are able to achieve much higher accuracy with fewer parameters compared to their corresponding base architectures.

Over-fitting is a common issue of training deep convolutional neural network especially when the dataset is limited. In this work, we propose a patch-level regularizer to force a convolutional neural network to learn many sub-models during training, and aggregate these models during testing. This approach proves to be robust and noise tolerant as our regularizer exposes small patches of an image to the network to learn all the features equally during training. The regularizer can be easily applied to the convolutional neural networks to further improve their classification performance. Experiment results on publicly available datasets demonstrate consistent improvement over existing regularizers.

Vessel type classification is a preliminary step in quantifying the severity of various diseases. This paper proposes DBA, a simple yet effective vessel type classification method based on the principle that arteries are brighter than veins at the local scale. The weighted local difference of the red channel intensity of the main trunk of each vessel is compared with that of its two immediately neighbouring vessels, a feature that is highly correlated with vessel-rectified oxygen capacity, and in turn, vessel type. Experiments on the publicly-available INSPIRE-AVR and DRIVE datasets obtained average vessel accuracies of 0.9217/0.9071, and average pixel accuracies of 0.9602/0.9634 respectively, with particular effectiveness on images with low contrast, non-uniform illumination and colour variation confirmed on the SiMES 1 dataset.

Previous image clutter metrics were proposed on the thought that clutter was just a perceptual effect, while we identify clutter as both perceptual and cognitive effects. Under this identification, we give a new definition of image clutter metric by analyzing the research results in the fields of visual psychology and psychophysics. According to the definition, we further put forward a DisSIMilarity (DSIM) based image clutter metric, which can also be taken as a kind of HVS-based signal-to-clutter ratio. The earlier image clutter metrics produced limited success in predicting targeting performance mainly since they did not consider brain cognitive characteristics. We develop a brain cognitive dissimilarity measure (BCDM) as a quantitative estimate of the selection weights which are allocated by brain attentional mechanism to affect visual selection processes. A human vision perceptual dissimilarity measure (VPDM), fully embodying vision perceptual properties, is first established between the target and clutter images, and then we utilize the BCDM between the two images as selection weights to pool the VPDM to be a clutter metric, which can be called DSIM metric. The metric is tested in Search_2 dataset provided by TNO Human Factors Research Institute of Netherlands. Error analysis and correlation tests demonstrate that the DSIM metric makes a more significant improvement than previously proposed metrics in predicting 62 observers' targeting performances including detection probability, false alarm probability and search time.