Search

Aiming at the community security and other fields which has strict requirements on the scale of infrared / visible image fusion calculation, this paper proposes an infrared / visible image feature domain fusion algorithm. The algorithm first uses the principal component analysis (PCA) model to transform the visible light image into the corresponding feature domain, and then projects the infrared light image onto the feature domain of the visible light image, and then performs segmented fusion of its principal components in the feature domain, and finally the fusion image is reconstructed. Experimental verification: when only 95% of the principal components are used for fusion, the performance of this algorithm in parameters such as STD and MI is better than some multi-scale transformation image fusion algorithms, and it is more excellent in the presentation of feature information.

Because of the complex environment of coal mines, there are many accidents like communication link fault caused by mobile equipment and sliding stones. Therefore, fast rebuilding of remaining nodes and dynamic adding of nodes toted by the rescuer are required after the disaster. Furthermore, the sensor nodes used for localization of the miner and his equipment can only be deployed redundantly along the straight laneway to form a linear wireless sensor network, which proposes a new requirement of topology control strategy. Combining with the self-organizing learning theory of the artificial neural networks, LCACL is presented, a linear clustering algorithm by competitive learning, in which sensor nodes are divided into mobile nodes which add or leave the network dynamically and fixed nodes which could act as sleep nodes, followers, cluster-bridges, or cluster-heads self-adaptively. On the comprehensive consideration of the remaining energy of the sensor node and the number of neighboring nodes, the competition vector, the base for competition, and evaluation vector, the evaluation standard, are formed, which correspond to the input vector and weight vector of artificial neural networks respectively. A similarity comparison using the Euclidean distance is made between them after normalization, and the competitive selection of roles and conversion are achieved using the Euclidean distance. Only the winner node will be entitled to make a learning adjustment, and then broadcasts the adjusted competition vector, which is the evaluation vector in the next competition. According to the rule of cluster-head selection from odd hops and cluster-bridge selection from even hops, the network is initialized as that nodes are divided into clusters of scale-uniform and low overlap ratio. The cluster-head initiates the competitive selection of the sleep node if there are too many followers, and its followers compete to sleep. Along with the energy consumption of the node, the cluster-head or cluster-bridge initiates a new election if it detects its energy consumption exceeding a certain threshold, and all the nodes in the cluster participate in competition by means of competitive learning to achieve role conversion. The simulation results show that the network organized by LCACL has balanced energy consumption, stronger ability to resist destruction, better expansibility, and longer lifetime of network, and the ratio of overlap could be 14% if the communication radius is 70 meters.