Salient Object Detection Method And Application Based On Visual Saliency

Visual saliency is an important research content in visual perception and sceneunderstanding, which involves the neuroscience, cognitive psychology, computer vision andinformation processing and multiple disciplines. The research content is very extensiveconcerning the scene cognition, target recognition and image retrieval. Visual saliency is thequality subjectively perceived on different areas in the scene image, thus making certain targetin the scene stand out from the adjacent area and attract our attention immediately. The causeof visual saliency is an attractive and fancy stimulation from a specific contrast between theforeground target and the scene background. In the paper, the key technologies in computervision and cognitive psychology were adopted to deeply research the visual saliency in thescene image, and the research achievement was applied in the visual saliency detection insmoke and flame, thus realizing the automatic detection of visual saliency on the salient targetin scene image and video frame.Through the research on two kinds of processing modes for visual perception bottom-upand top-down, we can find that the underlying visual stimuli attracts to note the resourceallocation, while the top visual perceived task and prior knowledge can better guide thedetection of visual saliency, and the detection efficiency can be improved through thecombination. In a scene image, the image local feature influenced the detection of visualsaliency, while the global feature also plays a key role. Parallel processing two features withperception hybrid theory can rapidly realize the detection of visual saliency. In the paper, thebottom-up and top-down combined visual information processing mode was adopted, thelocal feature and global feature in the scene image were conducted the parallel processing,and the detection of visual target in the scene image was realized through the feature fusion.The influence of rapid scene cognition and salient feature on the target recognition wereresearched with the method of cognitive psychology, the simulation on the recognition ofsalient feature target was realized with biased-ART neural network. In the previous studies,we can find that the rapid scene recognition is influenced by the image size and color andother features. In the paper, the recognition of four targets of “bird”,“ball”,“butterfly” and“flower” in color image, grayscale image, edge image and low-pass blurred image wascompleted through psychology experiments, and the experimental result can show that thetarget recognition is occurred after the rapid scene recognition; In the target recognition, thecolor, shape and texture are visual features with important cognitive function, and thedependence of different objects on different features are various, thus influencing therecognition speed and accuracy. Meanwhile, bART neural network model was adopted to simulate the cognitive psychology experiment, and the recognition on four targets wasrealized through the extraction of target featured color, shape and texture. Through the studyand comparison at two stages, bART model’s self-organizing and self-learning abilities werereflected, and a higher target recognition accuracy was achieved at the critical point amongthe detection of seven targets.The fire target detection technology based on visual saliency can automatically processand analyze the collected image without human intervention, detect and match according tothe fire visual target feature, locate, recognize and track the fire in dynamic scene. Based onthe research of SR algorithm proposed by Hou Xiaodi and others, the salient target can bedetected with multi-video frame-frame spectrum difference and single-video-frame windowspectrum difference. The salient object can be determined with the color visual feature of thefire through CIELab color space clustering method, and the visual salient point can be lockedthrough plot inverse probability difference. The fire target can be locked according to RGBand CIELab color feature value at salient point to realize the fire alarm. The smoke generallyoccurs at the early of the fire, especially presenting the spreading and diffusion feature at theoutdoor complex scene, and it is featured with the visual feature accumulated over time. Thedetection of salient movement area can be realized with the multi-step long frame differenceaccumulated algorithm in the paper. Based on the foundation of LRMR theory, theultra-complete dictionary can be constructed through multi-video frames to find out theforeground sparse matrix, and realize the separation of smoke foreground object from thecomplex background. Meanwhile, the translucent visual feature of the smoke color was alsoresearched in the paper, and the detection of smoke target can be realized through the growthof motion area and HSV color space saliency. The experiment can show that it has achieved agood result on the detection speed and accuracy with the method in the paper, and the methodcan be applicable to different outdoor scenes.