A wide range of applications in marine ecology extensively uses underwater cameras. Still, to efficiently process the vast amount of data generated, we need to develop tools that can automatically detect and recognize species captured on film. Classifying fish species from videos and images in natural environments can be challenging because of noise and variation in illumination and the surrounding habitat. In this paper, we propose a two-step deep learning approach for the detection and classification of temperate fishes without pre-filtering. The first step is to detect each single fish in an image, independent of species and sex. For this purpose, we employ the You Only Look Once (YOLO) object detection technique. In the second step, we adopt a Convolutional Neural Network (CNN) with the Squeeze-and-Excitation (SE) architecture for classifying each fish in the image without pre-filtering. We apply transfer learning to overcome the limited training samples of temperate fishes and to improve the accuracy of the classification. This is done by training the object detection model with ImageNet and the fish classifier via a public dataset (Fish4Knowledge), whereupon both the object detection and classifier are updated with temperate fishes of interest. The weights obtained from pre-training are applied to post-training as a priori. Our solution achieves the state-of-the-art accuracy of 99.27\% on the pre-training. The percentage values for accuracy on the post-training are good; 83.68\% and 87.74\% with and without image augmentation, respectively, indicating that the solution is viable with a more extensive dataset.
Image Understanding is becoming a vital feature in ever more applications ranging from medical diagnostics to autonomous vehicles. Many applications demand for embedded solutions that integrate into existing systems with tight real-time and power constraints. Convolutional Neural Networks (CNNs) presently achieve record-breaking accuracies in all image understanding benchmarks, but have a very high computational complexity. Embedded CNNs thus call for small and efficient, yet very powerful computing platforms. This master thesis explores the potential of FPGA-based CNN acceleration and demonstrates a fully functional proof-of-concept CNN implementation on a Zynq System-on-Chip. The ZynqNet Embedded CNN is designed for image classification on ImageNet and consists of ZynqNet CNN, an optimized and customized CNN topology, and the ZynqNet FPGA Accelerator, an FPGA-based architecture for its evaluation. ZynqNet CNN is a highly efficient CNN topology. Detailed analysis and optimization of prior topologies using the custom-designed Netscope CNN Analyzer have enabled a CNN with 84.5% top-5 accuracy at a computational complexity of only 530 million multiplyaccumulate operations. The topology is highly regular and consists exclusively of convolutional layers, ReLU nonlinearities and one global pooling layer. The CNN fits ideally onto the FPGA accelerator. The ZynqNet FPGA Accelerator allows an efficient evaluation of ZynqNet CNN. It accelerates the full network based on a nested-loop algorithm which minimizes the number of arithmetic operations and memory accesses. The FPGA accelerator has been synthesized using High-Level Synthesis for the Xilinx Zynq XC-7Z045, and reaches a clock frequency of 200MHz with a device utilization of 80% to 90 %.
Novel Coronavirus disease (COVID-19) is an extremely contagious and quickly spreading Coronavirus disease. Severe Acute Respiratory Syndrome (SARS)-CoV, Middle East Respiratory Syndrome (MERS)-CoV outbreak in 2002 and 2011 and current COVID-19 pandemic all from the same family of Coronavirus. The fatality rate due to SARS and MERS were higher than COVID-19 however, the spread of those were limited to few countries while COVID-19 affected more than two-hundred countries of the world. In this work, authors used deep machine learning algorithms along with innovative image pre-processing techniques to distinguish COVID-19 images from SARS and MERS images. Several deep learning algorithms were trained, and tested and four outperforming algorithms were reported: SqueezeNet, ResNet18, Inceptionv3 and DenseNet201. Original, Contrast limited adaptive histogram equalized and complemented image were used individually and in concatenation as the inputs to the networks. It was observed that inceptionv3 outperforms all networks for 3-channel concatenation technique and provide an excellent sensitivity of 99.5%, 93.1% and 97% for classifying COVID-19, MERS and SARS images respectively. Investigating deep layer activation mapping of the correctly classified images and miss-classified images, it was observed that some overlapping features between COVID-19 and MERS images were identified by the deep layer network. Interestingly these features were present in MERS images and 10 out of 144 images were miss-classified as COVID while only one out of 423 COVID-19 images was miss-classified as MERS. None of the MERS images was miss-classified to SARS and only one COVID-19 image was miss-classified as SARS. Therefore, it can be summarized that SARS images are significantly different from MERS and COVID-19 in the eyes of AI while there are some overlapping feature available between MERS and COVID-19.
In conditional Generative Adversarial Networks (cGANs), when two different initial noises are concatenated with the same conditional information, the distance between their outputs is relatively smaller, which makes minor modes likely to collapse into large modes. To prevent this happen, we proposed a hierarchical mode exploring method to alleviate mode collapse in cGANs by introducing a diversity measurement into the objective function as the regularization term. We also introduced the Expected Ratios of Expansion (ERE) into the regularization term, by minimizing the sum of differences between the real change of distance and ERE, we can control the diversity of generated images w.r.t specific-level features. We validated the proposed algorithm on four conditional image synthesis tasks including categorical generation, paired and un-paired image translation and text-to-image generation. Both qualitative and quantitative results show that the proposed method is effective in alleviating the mode collapse problem in cGANs, and can control the diversity of output images w.r.t specific-level features.
Encoded (or ciphered) manuscripts are a special type of historical documents that contain encrypted text. The automatic recognition of this kind of documents is challenging because: 1) the cipher alphabet changes from one document to another, 2) there is a lack of annotated corpus for training and 3) touching symbols make the symbol segmentation difficult and complex. To overcome these difficulties, we propose a novel method for handwritten ciphers recognition based on few-shot object detection. Our method first detects all symbols of a given alphabet in a line image, and then a decoding step maps the symbol similarity scores to the final sequence of transcribed symbols. By training on synthetic data, we show that the proposed architecture is able to recognize handwritten ciphers with unseen alphabets. In addition, if few labeled pages with the same alphabet are used for fine tuning, our method surpasses existing unsupervised and supervised HTR methods for ciphers recognition.
Dynamic patterns are characterized by complex spatial and motion patterns. Understanding dynamic patterns requires a disentangled representational model that separates the factorial components. A commonly used model for dynamic patterns is the state space model, where the state evolves over time according to a transition model and the state generates the observed image frames according to an emission model. To model the motions explicitly, it is natural for the model to be based on the motions or the displacement fields of the pixels. Thus in the emission model, we let the hidden state generate the displacement field, which warps the trackable component in the previous image frame to generate the next frame while adding a simultaneously emitted residual image to account for the change that cannot be explained by the deformation. The warping of the previous image is about the trackable part of the change of image frame, while the residual image is about the intrackable part of the image. We use a maximum likelihood algorithm to learn the model that iterates between inferring latent noise vectors that drive the transition model and updating the parameters given the inferred latent vectors. Meanwhile we adopt a regularization term to penalize the norms of the residual images to encourage the model to explain the change of image frames by trackable motion. Unlike existing methods on dynamic patterns, we learn our model in unsupervised setting without ground truth displacement fields. In addition, our model defines a notion of intrackability by the separation of warped component and residual component in each image frame. We show that our method can synthesize realistic dynamic pattern, and disentangling appearance, trackable and intrackable motions. The learned models are useful for motion transfer, and it is natural to adopt it to define and measure intrackability of a dynamic pattern.
Several bandwise total variation (TV) regularized low-rank (LR)-based models have been proposed to remove mixed noise in hyperspectral images (HSIs). Conventionally, the rank of LR matrix is approximated using nuclear norm (NN). The NN is defined by adding all singular values together, which is essentially a $L_1$-norm of the singular values. It results in non-negligible approximation errors and thus the resulting matrix estimator can be significantly biased. Moreover, these bandwise TV-based methods exploit the spatial information in a separate manner. To cope with these problems, we propose a spatial-spectral TV (SSTV) regularized non-convex local LR matrix approximation (NonLLRTV) method to remove mixed noise in HSIs. From one aspect, local LR of HSIs is formulated using a non-convex $L_{\gamma}$-norm, which provides a closer approximation to the matrix rank than the traditional NN. From another aspect, HSIs are assumed to be piecewisely smooth in the global spatial domain. The TV regularization is effective in preserving the smoothness and removing Gaussian noise. These facts inspire the integration of the NonLLR with TV regularization. To address the limitations of bandwise TV, we use the SSTV regularization to simultaneously consider global spatial structure and spectral correlation of neighboring bands. Experiment results indicate that the use of local non-convex penalty and global SSTV can boost the preserving of spatial piecewise smoothness and overall structural information.
Answering questions that require reading texts in an image is challenging for current models. One key difficulty of this task is that rare, polysemous, and ambiguous words frequently appear in images, e.g., names of places, products, and sports teams. To overcome this difficulty, only resorting to pre-trained word embedding models is far from enough. A desired model should utilize the rich information in multiple modalities of the image to help understand the meaning of scene texts, e.g., the prominent text on a bottle is most likely to be the brand. Following this idea, we propose a novel VQA approach, Multi-Modal Graph Neural Network (MM-GNN). It first represents an image as a graph consisting of three sub-graphs, depicting visual, semantic, and numeric modalities respectively. Then, we introduce three aggregators which guide the message passing from one graph to another to utilize the contexts in various modalities, so as to refine the features of nodes. The updated nodes have better features for the downstream question answering module. Experimental evaluations show that our MM-GNN represents the scene texts better and obviously facilitates the performances on two VQA tasks that require reading scene texts.
Bio-inspired spiking neural networks (SNNs), operating with asynchronous binary signals (or spikes) distributed over time, can potentially lead to greater computational efficiency on event-driven hardware. The state-of-the-art SNNs suffer from high inference latency, resulting from inefficient input encoding, and sub-optimal settings of the neuron parameters (firing threshold, and membrane leak). We propose DIET-SNN, a low latency deep spiking network that is trained with gradient descent to optimize the membrane leak and the firing threshold along with other network parameters (weights). The membrane leak and threshold for each layer of the SNN are optimized with end-to-end backpropagation to achieve competitive accuracy at reduced latency. The analog pixel values of an image are directly applied to the input layer of DIET-SNN without the need to convert to spike-train. The information is converted into spikes in the first convolutional layer where leaky-integrate-and-fire (LIF) neurons integrate the weighted inputs and generate an output spike when the membrane potential crosses the trained firing threshold. The trained membrane leak controls the flow of input information and attenuates irrelevant inputs to increase the activation sparsity in the convolutional and linear layers of the network. The reduced latency combined with high activation sparsity provides large improvements in computational efficiency. We evaluate DIET-SNN on image classification tasks from CIFAR and ImageNet datasets on VGG and ResNet architectures. We achieve top-1 accuracy of 66.52% with 25 timesteps (inference latency) on the ImageNet dataset with 3.1X less compute energy than an equivalent standard ANN. Additionally, DIET-SNN performs 5-100X faster inference compared to other state-of-the-art SNN models.
Self-supervised depth estimators have recently shown results comparable to the supervised methods on the challenging single image depth estimation (SIDE) task, by exploiting the geometrical relations between target and reference views in the training data. However, previous methods usually learn forward or backward image synthesis, but not depth estimation, as they cannot effectively neglect occlusions between the target and the reference images. Previous works rely on rigid photometric assumptions or the SIDE network to infer depth and occlusions, resulting in limited performance. On the other hand, we propose a method to "Forget About the LiDAR" (FAL), for the training of depth estimators, with Mirrored Exponential Disparity (MED) probability volumes, from which we obtain geometrically inspired occlusion maps with our novel Mirrored Occlusion Module (MOM). Our MOM does not impose a burden on our FAL-net. Contrary to the previous methods that learn SIDE from stereo pairs by regressing disparity in the linear space, our FAL-net regresses disparity by binning it into the exponential space, which allows for better detection of distant and nearby objects. We define a two-step training strategy for our FAL-net: It is first trained for view synthesis and then fine-tuned for depth estimation with our MOM. Our FAL-net is remarkably light-weight and outperforms the previous state-of-the-art methods with 8x fewer parameters and 3x faster inference speeds on the challenging KITTI dataset. We present extensive experimental results on the KITTI, CityScapes, and Make3D datasets to verify our method's effectiveness. To the authors' best knowledge, the presented method performs the best among all the previous self-supervised methods until now.