Object Detection Based On Visual Attention Models

Visual attention mechanism is an important part of human visionsystem, which could direct the attention to region of interest (ROI) rapidlyand exploit the limited resources to process the ROI. Traditional objectdetection approach is based on local features, which uses a sliding windowacross the image and apply a binary classifier at each window to detect thepresence or absence of the desired target object. The approach is slow andcomputationally expensive as each classifier is run independently at eachwindow within the image. Applying the visual attention mechanism toobject detection can process the salient region and un-salient region indifferent ways as the salient region is assigned high priority whileprocessing. In this way, visual attention mechanism could accelerate thedetection speed.This paper focuses on the visual attention mechanism and itsapplication in object detection. This paper proposes a visual saliencycomputational model based on the coding-based classification framework.The model could discriminate the stimuli of target and background well. Itconsists of four successive steps: feature extraction, feature coding, localpooling and saliency detection. The experimental results demonstrate thatthe model could highlight the region with target and inhibit the regionwithout the target at the same time.Based on the model mentioned earlier, the paper also proposes amulti-scale contextual pooling method as well as the automatic scaleselection algorithm. The automatic scale selection algorithm could selectthe optimal scale to represent the local features. The experimental resultsprove that multi-scale context facilitate the saliency detection process and our model could effectively improve the performance of goal-drivensaliency detection as well as related object detection. Compared with othermodels, the saliency maps by our model are more similar to theground-truth maps and get a smaller average KLD value. Furthermore, theP-R curves calculated by our model reveal that our model could ensurehigh precision and recall at the same time. Thus, our model is moresuitable for object detection.