Yuhang Dong

A tremendous amount of image and video data are being generated and shared in our daily lives. Image and video data are typically stored and transmitted in compressed form in order to reduce storage space and transmission time. The processing and analysis of compressed image and video data can greatly reduce input data size and eliminate the need for decompression and recompression, thereby achieving significant savings in memory and computation time. There exists a body of research on compression domain data processing and analysis. This survey focuses on the work related to image and video data. The papers cited are categorized based on their target applications, including image and video resizing and retrieval, information hiding and watermark embedding, image and video enhancement and segmentation, object and motion detection, as well as pattern classification, among several other applications. Key methods used for these applications are explained and discussed. Comparisons are drawn among similar approaches. We then point out possible directions of further research.

Digital images are usually stored in compressed format. However, image classification typically takes decompressed images as inputs rather than compressed images. Therefore, performing image classification directly in the compression domain will eliminate the need for decompression, thus increasing efficiency and decreasing costs. However, there has been very sparse work on image classification in the compression domain. In this paper, we studied the feasibility of classifying images in their JPEG compression domain. We analyzed the underlying mechanisms of JPEG as an example and conducted classification on data from different stages during the compression. The images we used were malaria-infected red blood cells and normal cells. The training data include multiple combinations of DCT coefficients, DC values in both decimal and binary forms, the “scan” segment in both binary and decimal form, and the variable length of the entire bitstream. The result shows that LSTM can successfully classify the image in its compressed form, with accuracies around 80%. If using only coded DC values, we can achieve accuracies higher than 90%. This indicates that images from different classes can still be well separated in their JPEG compressed format. Our simulations demonstrate that the proposed compression domain-processing method can reduce the input data, and eliminate the image decompression step, thereby achieving significant savings on memory and computation time.

Malaria is a severe public health problem worldwide, with some developing countries being most affected. Reliable remote diagnosis of malaria infection will benefit from efficient compression of high-resolution microscopic images. This paper addresses a lossless compression of malaria-infected red blood cell images using deep learning. Specifically, we investigate a practical approach where images are first classified before being compressed using stacked autoencoders. We provide probabilistic analysis on the impact of misclassification rates on compression performance in terms of the information-theoretic measure of entropy. We then use malaria infection image datasets to evaluate the relations between misclassification rates and actually obtainable compressed bit rates using Golomb–Rice codes. Simulation results show that the joint pattern classification/compression method provides more efficient compression than several mainstream lossless compression techniques, such as JPEG2000, JPEG-LS, CALIC, and WebP, by exploiting common features extracted by deep learning on large datasets. This study provides new insight into the interplay between classification accuracy and compression bitrates. The proposed compression method can find useful telemedicine applications where efficient storage and rapid transfer of large image datasets is desirable.

Malaria is a life-threatening disease caused by parasites that are transmitted to people through the bites of infected mosquitoes. Automation of the diagnosis process will enable accurate diagnosis of the disease and hence holds the promise of delivering reliable health-care to resource-scarce areas. Machine learning technologies have been used for automated diagnosis of malaria. We present some of our recent progresses on highly accurate classification of malaria-infected cells using deep convolutional neural networks. First, we describe image processing methods used for segmentation of red blood cells from wholeslide images. We then discuss the procedures of compiling a pathologists-curated image dataset for training deep neural network, as well as data augmentation methods used to significantly increase the size of the dataset, in light of the overfitting problem associated with training deep convolutional neural networks. We will then compare the classification accuracies obtained by deep convolutional neural networks through training, validating, and testing with various combinations of the datasets. These datasets include the original dataset and the significantly augmented datasets, which are obtained using direct interpolation, as well as indirect interpolation using automatically extracted features provided by stacked autoencoders. This chapter ends with a discussion of further research.

While Golomb-Rice codes are optimal for geometrically distributed source, the practically achievable coding efficiency depends on the accuracy of the coding parameter estimated from the input data. Most existing methods are based on the assumption of geometric distribution and thus would suffer from a loss in coding efficiency if the underlying distribution deviates from the geometric distribution, which is usually the case in practice. We proposed a data-driven parameter estimation method without assuming the underlying distribution. We formulated the problem of choosing the best coding parameter for the given input data as a pattern classification problem. To this end, we trained a deep belief network using the data segments to be coded, along with their “labels”, which are the optimal coding parameters that yield the shortest codewords. Simulations on data synthesized using statistical models, as well as data in hyperspectral image coding showed that the proposed deep learning method tended to be more robust than several state-of-the-art parameter estimation methods, with the capability to further improve the accuracies of these methods.

In many biomedical applications, images are stored and transmitted in the form of compressed images. However, typical pattern classifiers are trained using original images. There has been little prior study on how lossily decompressed images would impact the classification performance. In a case study of automatic classification of malaria infected cells, we used decompressed cell images as the inputs to deep convolutional neural networks. We evaluated how various lossy image compression methods and varying compression ratios would impact the classification accuracies. Specifically, we compared four compression methods: lossy compression via bitplane reduction, JPEG and JPEG 2000, and sparse autoencoders. Decompressed images were fed into LeNet-5 for training and testing. Simulation results showed that for similar compression ratios, the bitplane reduction method had the lowest classification accuracy, while JPEG and JPEG 2000 methods could maintain good accuracies. In particular, JPEG 2000 decompressed images could achieve about 95% accuracy even after 30 to 1 compression. We also provide classification results based on the widely used MNIST dataset, where handwritten digits were found to be much easier to classify using decompressed images, with about 90% accuracy still achievable using only one single bitplane. As a lossy compression method, Autoencoder was also applied to the MNIST dataset, achieving about 85% accuracy with a compression ratio much higher than the other three lossy image compression methods. Autoencoders were also found to provide more scalable compression ratios, while capable of maintaining good classification accuracies.

This paper studied automatic identification of malaria infected cells using deep learning methods. We used whole slide images of thin blood stains to compile an dataset of malaria-infected red blood cells and non-infected cells, as labeled by a group of four pathologists. We evaluated three types of well-known convolutional neural networks, including the LeNet, AlexNet and GoogLeNet. Simulation results showed that all these deep convolution neural networks achieved classification accuracies of over 95%, higher than the accuracy of about 92% attainable by using the support vector machine method. Moreover, the deep learning methods have the advantage of being able to automatically learn the features from the input data, thereby requiring minimal inputs from human experts for automated malaria diagnosis.

This paper aims at investigating the best feature selection method for optimized and automated machine learning based detection of malarial parasite in wholeslide images of peripheral blood smears. We do this by extracting samples from the wholeslide images and performing feature extraction. A host of feature selection methods are used to judge the performance of the Support Vector Machine as a binary classifier. For each feature selection method, the Support Vector Machine is trained using the significant features. We perform cross-validation and grid-search for finding the best SVM parameters. The trained SVM is subsequently used to classify known instances of "Normal" and "Infected" samples that are taken randomly from the wholeslide image but unknown to the SVM. Confusion matrices are generated for each classification performed. Various performance measures of each classification task are reported. We conclude that based on our experiments, the binary SVM classifier yields a superlative accuracy of 95.5% if the feature-selection is based on Kullback-Leibler distance between the two classes.

In this paper, we presented an integrated graphical user interface (GUI) tool, which supports interactive regions of interest (ROIs) extraction from whole slide imagery and provides lossless compression of the ROIs extracted, leading to over 40 times compression on the original image. As a case study to demonstrate the usefulness of this tool, we applied a simple block-matching technique to detect malaria infected cells from a whole slide image based on a template provided by the user, who can control the percentage of the extracted pixels by varying the acceptable similarity threshold. The tool allowed us to conduct a comparative study on the efficiencies of several compression methods (including Bzip2, Winzip and Shape-Adaptive JPEG 2000) on image subsets that are continuously variable. This case study provides new insights into the interactions between the geometry of ROIs and the amount of compression achievable on medical images.