Abstract

In image processing and computer vision tasks such as object of interest image segmentation, adaptive image compression, object based image retrieval, seam carving, and medical imaging, the cost of information storage and computational complexity is generally a great concern. Therefore, for these and other applications, identifying and focusing only on the parts of the image that are visually most informative is much desirable. These most informative parts or regions that also have more contrast with the rest of the image are called the salient regions of the image, and the process of identifying them is referred to as salient object detection. The main challenges in devising a salient object detection scheme are in extracting the image features that correctly differentiate the salient objects from the non-salient ones, and then utilizing them to detect the salient objects accurately.
Several salient object detection methods have been developed in the literature using spatial domain image features. However, these methods generally cannot detect the salient objects uniformly or with clear boundaries between the salient and non-salient regions. This is due to the fact that in these methods, unnecessary frequency content of the image get retained or the useful ones from the original image get suppressed. Frequency domain features can address these limitations by providing a better representation of the image. Some salient object detection schemes have been developed based on the features extracted using the Fourier or Fourier like transforms. While these methods are more successful in detecting the entire salient object in images with small salient regions, in images with large salient regions these methods have a tendency to highlight the boundaries of the salient region rather than doing so for the entire salient region. This is due to the fact that in the Fourier transform of an image, the global contrast is more dominant than the local ones. Moreover, it is known that the Fourier transform cannot provide simultaneous spatial and frequency localization.
It is known that multi-resolution feature extraction techniques can provide more accurate features for different image processing tasks, since features that might not get extracted at one resolution may be detected at another resolution. However, not much work has been done to employ multi-resolution feature extraction techniques for salient object detection. In view of this, the objective of this thesis is to develop schemes for image salient object detection using multi-resolution feature extraction techniques both in the frequency domain and the spatial domain.
The first part of this thesis is concerned with developing salient object detection methods using multi-resolution frequency domain features. The wavelet transform has the ability of performing multi-resolution simultaneous spatial and frequency localized analysis, which makes it a better feature extraction tool compared to the Fourier or other Fourier like transforms. In this part of the thesis, first a salient object detection scheme is developed by extracting features from the high-pass coefficients of the wavelet decompositions of the three color channels of images, and devising a scheme for the weighted linear combination of the color channel features. Despite the advantages of the wavelet transform in image feature extraction, it is not very effective in capturing line discontinuities, which correspond to directional information in the image. In order to circumvent the lack of directional flexibility of the wavelet-based features, in this part of the thesis, another salient object detection scheme is also presented by extracting local and global features from the non-subsampled contourlet coefficients of the image color channels. The local features are extracted from the local variations of the low-pass coefficients, whereas the global features are obtained based on the distribution of the subband coefficients afforded by the directional flexibility provided by the non-subsampled contourlet transform.
In the past few years, there has been a surge of interest in employing deep convolutional neural networks to extract image features for different applications. These networks provide a platform for automatically extracting low-level appearance features and high-level semantic features at different resolutions from the raw images. The second part of this thesis is, therefore, concerned with the investigation of salient object detection using multiresolution deep convolutional features. The existing deep salient object detection schemes are based on the standard convolution. However, performing the standard convolution is computationally expensive specially when the number of channels increases through the layers of a deep network. In this part of the thesis, using a lightweight depthwise separable convolution, a deep salient object detection network that exploits the fusion of multi-level and multi-resolution image features through judicious skip connections between the layers is developed. The proposed deep salient object detection network is aimed at providing good performance with a much reduced complexity compared to the existing deep salient object detection methods.
Extensive experiments are conducted in order to evaluate the performance of the proposed salient object detection methods by applying them to the natural images from several datasets. It is shown that the performance of the proposed methods are superior to that of the existing methods of salient object detection.