Visual Detection With Different Granularities For Minute Change Monitoring Of High Value Targets

Detection is one of the most common problem in the field of computer vision.According to the difference of detection granularities,it can be classified into coarse grain and fine grain.Coarse-grained detection refers to object-level detection,often used in object recognition,image understanding,robot navigation and other fields.Saliency detection and co-saliency detection are the most important problems in coarse-grained visual detection.Fine-grained detection refers to the pixel-level detection,which is often used in noise,blur,mural deterioration and minute change detections.In practical applications,there are many visual detection problems with different granularities,such as in minute change monitoring of mural,we can use coarse-grained saliency detection for monitoring points selection,and use co-saliency detection for camera-pose relocation,use fine-grained blur detection for obtaining high quality images,and use mural deterioration and minute change detections for analysis.In this paper,we study visual detection with different granularities for minute change monitoring of high value targets from the granularities of the detected targets and the number of input images.The main research contents and contributions are as follows:1)Single-image coarse-grained saliency detection: We propose two different coarsegrained saliency detection methods from the aspects of color feature selection and the super-pixel scale.The proposed methods can uniformly highlight the foreground objects and suppress the background regions.2)Multi-image coarse-grained co-saliency detection: We propose two different coarse-grained co-saliency detection methods based on the Gaussian mixture model and color feature reinforcement.The obtained co-saliency detection results have a good inhibitory effect on the non co-salient objects.3)Single-image fine-grained blur detection: We propose a fine-grained multiscale blur detection method to obtain the pixel-level blur probability,and also analyze the functions of the filters in each layer of the proposed network.The proposed model achieves better blur detection results on benchmark dataset.4)Multi-image fine-grained mural deterioration and minute change detections: We present two multi-image fine-grained detection methods,including mural deterioration detection and minute change detection.In the minute change detection,we design a multi-path convolutional neural network to collaboratively detect deterioration of the multi-lighted mural images,which fills the blank of the automatic mural deterioration labeling.In minute change detection,we design a large network which consists two subnetworks(i.e.,pose correction network and change detection network),solving the sensibility of the parameters and low detection precision problems of the traditional minute change detection methods.