Target Detection Technology Based on Object Model Optimization Neural Network Learning

Remote sensing images often contain large ground range. A large number of targets, irregular distribution rules and large scale transformation are included in an image, so the target detection is difficult and it takes a lot of time to calculate. The traditional target detection can locate the target in the image by multi-scale sliding window, but the selection speed, quantity and quality of the candidate frame not only affect the time efficiency of the target detection, but also affect the accuracy of the target detection. The sliding window method thinks that the possibility of each position in the image is the same. Therefore, it traverses every position in the image into a candidate frame window, and exhaustion of the search images with violent exhaustion, resulting in a large number of redundant and low quality redundant windows.

WEI [1] proposes a target intention recognition model based on radial basis function neural network; YU [2] proposes a joint supervised recognition method based on dense convolution neural network, which combines local and global features, and obtains image features based on dense convolution neural network. ZHANG [3] proposed a fine target recognition method for color image under complex background, and used Bayesian model to distinguish skin color and background color in color image; ZHANG [4] aiming at the problems of tedious process and difficult feature extraction in traditional image recognition algorithm, an image adaptive target recognition algorithm based on depth feature learning is proposed; WANG [5] simply processes the original data and inputs it directly as input data into the convolution neural network. The convolution neural network is used to analyze the local features.

The calculation process is very time-consuming and violates the human visual mechanism. In order to solve the limitation of the selection speed, quantity and quality of the target candidate frame in the traditional sliding window detection technology, the detection efficiency and accuracy of the target detection are improved. In this paper, a candidate frame screening preprocessing algorithm is proposed to tell the detection network which areas should be paid attention to, and can be combined with the actual features of remote sensing images to target specific targets. The initial candidate box is selected to reduce the false alarm rate. On this basis, the object feature sharing can reduce the calculation cost of the target area discovery based on the object character, avoid the full graph search, reduce the time consuming and improve the correct rate of the candidate region discovery. This detection technology greatly improves the recognition rate of remote sensing image targets, reduces the computation time and has practical promotion and application.