In order to achieve efficiency and accurate identification of objects on high voltage transmission lines, this paper studies the improvement based on YOLO network. Traditional YOLO options a new basic model, called Darknet-19, including 19 voluntary layers and 5 maxpooling layers Darknet-19 is consistent with the VGG16 model design principle, mainly using 3 * 3 solution, After using 2 * 2 maxpooling layer, the feature map dimension is reduced by 2 times, while the channels of the feature map are doubled Similar to NIN, Darknet-19 ultimate uses global avgpooling for prediction, and uses 1 * 1 revolution between 3 * 3 revolution to compress the feature map to reduce the computational amount and parameters of the model In this paper, a two dimensional chaotic attractor of the sinusoidal basis function is generated by iterative operation to be adjusted to the image, and the attractor can be used as a feature input network of the image, replacing the one layer pooling operation of the Darkenet-19 network For objects on high voltage transmission lines, the improved network is used to identify them, improve accuracy, reduce calculation time, and reduce the structural complexity of the network The methods of other literature were applied to the database of this paper, and the indicators were recorded experimentally, and then compared and analyzed with the methods in this paper The mechanism of iterative attractor replacement revolution and pooling is studied, and the learning method of iterative function parameter change is realized
|