ResNet Small Target Detection Algorithm Based on Deep Separable Sonvolution

Abstract: Because of the poor effect of traditional small target detection, by analyzing the characteristics of small targets, taking RESNET of deep separable convolution as the feature extraction network, a small target detection algorithm based on RESNET of deep separable convolution is proposed. In order to ensure that the network has good adaptability and strong small target feature extraction ability, the algorithm adopts the topology of the multi convolution kernel. The convolution form of the network is improved by packet convolution to reduce the number of network parameters and computation, and the improved channel shuffling is used to enhance the exchange of feature information between different packets and splice the output features. Finally, combined with the residual connection form, a deep separable packet convolution RESNET network (mower) with multiple convolution cores is formed. The experimental results of the DOTA data set show that the Top1 error rate and top5 error rate of the RESNET network based on deep separable convolution are 30.68% and 8.75%, respectively, which is 3.34% and 1.56% lower than that of traditional RESNET network. The complexity of the model is also reduced, which has obvious advantages over other network models.