Optical Remote Sensing Image Object Detection Based On Multiple-scale Features And Model Compression Acceleration

With the advancement of technology,remote sensing image technology is becoming more and more important in various industries.Remote sensing technology has the characteristics of high real-time,periodic collection of information,highly adjustable electromagnetic wave information and wide coverage of information,thus the image information of ground object obtained by remote sensing technology is more comprehensive and accurate.Optical remote sensing images is an important branch of remote sensing images.Object detection of optical remote sensing image is of great significance.In the military,it can detect the change information of aircraft ships on enemy airports,aircraft carriers and important docks in real time,so as to achieve the opportunity to quickly gain combat and gain the initiative in the information age.Commercially,it is possible to monitor naval vessels,quickly combat illegal fishing,maritime smuggling,illegal smuggling and maritime rescue.Although deep learning has achieved certain breakthroughs in object detection tasks,object detection network based on deep learning is slow in complex operation,high false alarm rate,low accuracy,and poor detection of small targets.Therefore the thesis applies full convolution semantic segmentation,neural network compression and expanded convolution to the optical remote sensing image object detection task to improve the detection accuracy and reduce the false positives.The main research contents of the thesis are as follows:1.In the detection of aircraft and ships for optical remote sensing images,we propose a new object detection algorithm based on full convolution semantic segmentation to sloving the problem of the misdetecting ships on land.The algorithm uses an end-to-end semantic segmentation network to segment the land and ocean in optical remote sensing images.Because we need two-classification of segmentation,the original complex segmentation network is simplified and the segmentation is combined with the object detection network.The semantic segmentation network shares the first seven convolution layers with the object detection network.Finally,the results of semantic segmentation are used to assist the correction of the detection results to reduce the false positives and improve the accuracy of the detection results.We validated our algorithm in the optical remote sensing image dataset collected by Quick Bird satellite and applied for a national invention patent(patent application number: 201010112969.6).At present,the algorithm needs to further optimize the network structure to reduce the complexity.2.Aiming at the problem that the target detection network has deep layers,many weight parameters and high computational complexity,we propose a object detection network base on multi-state neural network.The multi-state neural network quantizes the weight and activation values of the neural network and encodes the multi-bit(1~8-bit){-1,1} of the network parameters and activation values according to the degree of sparsity of the quantization to achieve network model compression.And The original convolution layer is decomposed into multiple binary neural network convolution layers to speed up the network operation.Applying a multi-state neural network to the object detection network can greatly increase the speed of the network while saving computing resources.The object detection network base on multistate neural network can achieve 3 to 4 times compression under the condition of the same detection result as the full-precision network.We performed algorithm verification on the optical remote sensing image dataset collected by Quick Bird satellite and the natural image VOC2007/2012 object detection dataset.And we have published a paper named “MultiPrecision Quantized Neural Networks via Encoding Decomposition of {-1,+1} at the international top conference(AAAI-2019),and applied for a national invention patent(patent application number: 201108866365.0).At present,the problem that the algorithm needs to solve is how to realize the model acceleration of the convolution layer which is not decomposed into a binary neural network in the network.3.Because the convolutional neural network framework is more suitable for image classification tasks and the problem of small target miss detection rate in object detection tasks is high,we propose a object detection network base on dilated convolution.The object detection network base on dilated convolution replaces the convolutional layer and the pooling layer used in extracting deep features with its own dilated convolution module,and increases the characteristic receptive field by dilated convolution.Using the dilated convolution module,deeper and more abstract feature extraction can be performed on the image without changing the size of the feature image,and the features of the small target are not lost.Finally,the fusion of the shallow feature image and the deep feature image is obtained.The fusion feature contain abstract features required for classification and specific features required for regression.We validated our algorithm in the optical remote sensing image dataset collected by Quick Bird satellite and applied for a national invention patent(patent application number:201110907184.8).At present,the problem of the algorithm is how to not use the weights obtained from the classified data set to train the object detection network to improve the accuracy of the target detection network.