Research On Small Target Detection Based On Feature Enhancement And Convolutional Neural Network

In recent years,Convolutional Neural Network has been widely used in target detection because of its powerful image feature representation ability.These networks do work well on high-resolution images,objects with a clear look and structure.For small objects,however,the accuracy needs to be improved.As the network structure deepens,the scale and resolution of the feature layer will decrease,and the detailed information will inevitably be lost in the process of convolution.Therefore,it is difficult to learn rich feature representations from the appearance and structure with poor quality.Later,there are methods to try to improve the performance of small target detection by data enhancement or increasing feature dimensions.However,simply increasing the proportion of input images usually leads to increased computational load or time for training and testing,which slows down the detection speed of the model.In addition,simple multi-scale fusion is like a black box,which cannot guarantee that the fusion features are effective and interpretable and discriminant.Therefore,how to make target detection in complex scenes to ensure high detection accuracy and real-time requirements is an urgent problem to be solved in this paper.Based on this,this paper starts from feature enhancement,analyzes the deep and shallow features extracted by the existing feature extraction network,and improves the performance of target detection by improving the structure of convolutional neural network.First of all,in the shallow characteristics,due to the feature extraction of shallow features include the target position information and more background noise,we design a feature enhancement model,based on feedback mechanism via the Gaussian high-pass filter to deal with shallow,then a layer back propagation fusion characteristics,a new enhancement characteristics of shallow after Gaussian high-pass filtering processing,then the upper back propagation fusion characteristics,the process of shallow characteristics in the new generation on every time,after n iterations of shallow characteristics,generation can enhance foreground area target and suppress background noise.Second,in the deep features,the use of hollow convolution to expand feeling,but considering the empty convolution pore will lead to more small target information omission,put forward three kinds of deep hole rate combination of stack structure to extract features,through the design experiment to find the best hole rate combination of(1,2,3),close cover figure at the same time reducing the target information missing in the "empty" detection effect is better.Thirdly,a multi-scale fusion module is constructed in this paper to integrate the above enhanced shallow features and deep features to obtain output features with strong representation capability for subsequent detection and classification tasks.Finally,build the model and migrate the usage.The model was designed and built on VGG-16 and Res Net-50 networks,and then migrated to Darknet-19 and Darknet-53,and the validity and stability of the model were verified.The training and testing results on PASAL VOC and COCO data sets show that the algorithm performance after adding the design module in this paper is better than the original benchmark network,and the accuracy m AP is improved by about 2.5%～3%,which has a significant improvement effect in small targets.In addition,the real-time target detection algorithm can still meet the requirements of real-time detection.