Development of Improved Saliency Detection Techniques

Abstract

Saliency is the perceptual quality by which an observable thing or object stands out relative to its environment. It is that property by which some object stands out. Accurate detection of salient object is a challenging task. In this thesis, novel and robust saliency detection techniques and their different applications are proposed to detect salient objects from images. Different features of images are utilized to extract accurate saliency maps. The first proposed technique formalizes the saliency map using color and depth along with input’s histogram. The color saliency map assumes that the salient objects in an image are usually more colorful than objects that are not salient. Likewise, the salient objects are usually found closer to the camera during acquisition; therefore, the transmission map is also utilized for saliency detection. The third map is estimated from the histogram. These three maps are then fused and filtered in order to yield a smooth saliency map. The second proposed technique is based on texture contrast and convex hull center prior maps. Input image is first segmented into super-pixels. Primary saliency map is constructed using texture features. Convex-hull-based center prior map is extracted and refined with graph regularization to obtain secondary map. Primary and secondary maps are fused based on weight matrix. Thresholding and filtering are applied to obtain final map. The third proposed technique is frequency and gradient domain saliency detection using pyramid approach. The edges of the input image are sharpened first. Pyramid decomposition at different scales is applied. At each scale, the color space of the RGB image is converted into YIQ to extract the Q channel. In the frequency domain, Butterworth high-pass filtering is applied. Gradient and DoG filters are applied. Each channel’s saliency map is created via region filling. The saliency maps of pyramids at different scales are integrated and filled. The fourth proposed saliency detection technique is for underwater images. The input image is enhanced and gradient domain filtering is applied. The filtered image undergoes a pyramid decomposition. Saliency is computed on three different scales on two parallel steps. In first step, saliency is computed using cellular automata at each scale. The map is further optimized and filtered to obtain saliency at a given scale. A primary saliency map is computed by scale integration. In second step, saliency is computed using a homomorphic filter on each scale. Guided filter is applied to obtain saliency at a given scale. Secondary saliency map is computed by scale integration. Multiplicative fusion is applied on primary and secondary saliency maps to obtain final output. In the fifth proposed technique, saliency based fabric defect detection via bag-of-wordsmodel is proposed. The proposed methodology feeds the fabric images into the framework. The proposed saliency technique is used to create saliency maps of fabric images. The kmean clustering algorithm is used to extract local features. A bag of features is created using cluster centers. To detect defected fabric, feature vectors are used to train a K-Nearest Neighbor (K-NN) classifier. In the sixth proposed technique, it is proposed to direct the user’s attention to a specific region of interest (ROI). To begin with, a bilateral filter is used to smooth ROI by removing unwanted details from the defocus map. After that, the hue component of the ROI is adjusted, and the outline is extracted using neighboring data. Lastly, the hue component is used to direct the viewer’s attention to a specific region in the resultant image’s outline. Cutting-edge existing techniques are used to compare the proposed techniques. Experiments demonstrate the significance of proposed techniques on various datasets. The simulation results show that the proposed techniques produce better results that are noise-free and very close to the ground truth. These salient detection techniques can be used in various applications like image and video compression, object tracking, image segmentation, etc.