Research On Small Target Detection Method Based On Convolutional Neural Networ

Target detection is widely used in military reconnaissance,intelligent driving,medical diagnosis and other fields.Because of small target imaging size and little pixel information,traditional target detection methods based on artificial features are difficult to achieve efficient and robust detection of small targets.With the rapid development of deep learning theory,the target detection algorithm based on convolutional neural network has the advantages of high accuracy,fast speed and strong generalization ability.Therefore,this thesis takes convolutional neural networks as the core means to study the detection of small targets in two typical detection tasks,namely remote sensing ground detection and infrared military warning strike,as follows.For the detection of small targets(about 30*30 pixels)in the complex background of remote sensing images,an SSD algorithm based on multi-scale feature fusion is designed in this paper using SSD(Single Shot Multi Box Detector)as the base network.First,a feature map fusion mechanism is constructed to fuse the shallow feature maps with high resolution and the deep feature maps with rich semantic information,and a feature pyramid is constructed between the feature maps to enhance the small target features.Then,the channel attention module is introduced to reduce the background interference by constructing a weight parameter space to focus more attention on the channels focusing on the target region.Finally,the scale of the a priori frame relative to the original image is adjusted so that it can be better adapted to the remote sensing small target scale.The method performance is tested qualitatively and quantitatively using the collected remote sensing aircraft image dataset,and the results show that the detection accuracy of the improved method is improved by 4.3% compared to SSD.For small targets with even smaller scales(below 15*15 pixels)in infrared military warning strikes,this thesis proposes a target detection algorithm based on local contrast attention mechanism by fusing the traditional infrared small target detection algorithm with convolutional neural network using Res Net as the backbone network.The algorithm embeds the traditional multi-scale local contrast algorithm into the Res Net network after modularization by expanding convolution,and constructs a cyclic displacement acceleration scheme to accelerate the network,and at the same time achieves the enhancement of small target features by constructing a top-down attention mechanism to fuse the features at different scales.Experimental validation on the publicly available infrared SIRST dataset demonstrates that the algorithm achieves an n IOU value of 0.755,an improvement of more than 9.4% compared to the conventional algorithm.