Research And Implementation Of Model Compression Technology For Aircraft Target Detection In Remote Sensing Image

Target detection and recognition based on remote sensing image has great application prospects in many fields.Especially in the military field,we can get a lot of information that is difficult to get by analyzing the remte sensing images acquired by airborne and satellite,and using the target detection technology on the remote sensing images can make some information more intuitive.With the coming of the era of artificial intelligence,deep learning technology is far superior to traditional methods in the task of target recognition and detection.However,for airborne,satellite and other application environments,the huge amount of computation and large space storage environment requirements for supporting deep learning network have become an urgent problem,so this paper proposes a model compression technology based on convolutional neural network to solve this problem.The research mainly includes convolutional neural network storage space compression and compputational complexity compression,aiming to reduce the convolutional neural network weight storage space,memory space needed for operation,and improve the processing speed of convolutional neural network operation.In order to achieve the above purpose,this paper proposes a lightweight target detection network based on channel rearrangement,which reduces the storage space and speeds up the running speed while maintaining the accuracy of model detection;at the same time,a new structural model pruning strategy is proposed,the rules used to measure the importance of convolution kernel are re formulated,and a new pruning strategy based on convolution kernel is constructed The basic unit can compress the volume of the vgg16 network by more than 16 times and the calculation amount by more than 3 times under the condition of less precision loss.Based on the imp rovement of algorithm this paper proposes the design and implementation of aircraft target detection software system,including the overall archiecture,design ideas and deep learning training reasoning architecture.The detection task of convolutional neural network is realized by hardware and software cooperation.Through multi-scale sensitive area screening and other methods,the detection accuracy of the system is improved,and the multi-language standard network interface is provided.