Non‐attention region first initialisation of k ‐means clustering for saliency detection

Abstract

According to the nature of saliency map generation with colour contrast, a non-attention region first initialisation (NARFI) k-means clustering for saliency detection is proposed. The NAR is obtained by multiwavelet reconstruction based on the cutoff low-frequency. The initial seeds of the k-means are chosen from the NAR. This way, the NAR is clustered in a fine manner, whereas the attention region is clustered in a coarse manner. As a result, the saliency values of the attention region with the NARFI k-means clustering are more conspicuous than those with the k-means++ clustering.

Visual saliency detection: a Kalman filter based approach

“…In this paper we have presented a Kalman filter based saliency detection method which generates a visually expected scene and based on that builds a saliency map. We have developed our model around the notion of visual surprise and it can be extended easily for video data, where instead Model AUC-Judd AUC-Borji CC SIM NSS CNN-VLM [20] 0.79 0.79 0.44 0.43 1.18 MKL [21] 0.78 0.78 0.42 0.42 1.08 CAS [13] 0.74 0.73 0.36 0.43 0.95 LGS [23] 0.76 0.76 0.39 0.42 1.02 GNM [14] 0.74 0.67 0.34 0.42 0.97 NARFI [15] 0.73 0.61 0.31 0.38 0.83 STC [22] 0.79 0.78 0.40 0.39 0.97 RCSS [12] 0.75 0.74 0.38 0.44 0.95 CIWM [17] 0 of traversing the spatial domain, we will progress through the time domain. Our proposed model also provides a great deal of flexibility as anybody can use their own definition of the function, M k , combining multiple features.…”

“…As the ground fixation maps for MIT-300 images are not publicly available, we compared our model only quantitatively with the other approaches on this dataset. In addition to the 5 models used for comparison on Toronto dataset, we have assessed our models performance on MIT-300 against 7 other state of the art methods which are: CNN-VLM [20], Multiple Kernel Learning model (MKL) [21], Context Aware Saliency (CAS) [13], Generalized Nonlocal Mean Saliency (GNMS) [14], NARFI saliency (NARFI) [15], Sampled Template Collation (STC) [22] and LGS model [23]. In table 3, we have presented quantitative performance of various models on MIT-300 data set and these results from MIT-300 clearly demonstrates the superiority of our kalman based method which outperformed all other approaches against AUC-Judd, AUC-Borji and CC metric.…”

WEPSAM: Weakly Pre-Learnt Saliency Model

“…It is to be noted that WEPSAM was fine-tuned only on images of iSUN and SALI-CON datasets and thus the test images are substantially different than training images. We compare our model with recent state-of-the-art methods such as multi resolution CNN (MR-CNN) [19], CNN-VLM [20], multiple kernel based learning (MKL) [12], RARE-2012 [21], Context Aware Saliency Model(CAS) [9], Local+Global Saliency Model (LGS) [22], Generalized Nonlocal Mean Saliency (GNMS) [23], NARFI saliency (NARFI) [24], Sampled Template Collation (STC) [8] and Chromatic Induction Wavelet Model (CIW) [7]. The first three models are essentially learning based.…”