Research On Small Target Detection Based On Adaptive Deep Feature

With the continuous development of human industrial technology and the deepening of artificial intelligence research,the application of object detection methods based on deep learning to actual scenes has become an important research topic.The task content of object detection is to find out the preset target coordinates from the image or video data and classify the positioning of the target.The task of detecting a target whose target area occupies a small proportion of the entire image area or the target has a small number of pixels is called small object detection.Because the small object itself contains less information,it is continuously convolved with the surrounding background information during the deep convolution process,and its own characteristics are weakened.And in the actual application scenarios of industrial vision,due to the changeable scenes,unobvious object features and lack of samples,the accuracy of object detection methods is difficult to guarantee.Therefore,small object detection has always been the main goal of object detection algorithms from experimental theory to engineering implementation obstacle.To solve the above situation,the main work of this paper is as follows:(1)This paper proposes a small object detection method based on a multi-scale crossattention model.This method is aimed at the problem of small object detection,starting from the characteristics of the convolutional neural network,and using the YOLO series of algorithms as the basic framework,a multi-scale cross-attention module is designed.In the attention model,this method uses the method of adding one to the global attention element point to ensure that the module does not have problems such as feature disappearance.The attention mask is cross-multiplied,and the convolutional network shallow feature map is used to generate spatial attention mask,and add the attention mask to the deeper feature map to supplement the target edge information;use the convolutional network deep feature map to generate the channel attention mask,and add the attention mask to the shallower layer,the feature map highlights the semantic information of the target;it improves the representation ability of the feature map.Experimental results show that this method can significantly improve the detection accuracy of small objects by the basic network,and it also can improve detection accuracy for normal object detection tasks.(2)This paper proposes a small object detection method based on adaptive spatialfrequency domain feature fusion.This method first uses the image wavelet transform to obtain the original image wavelet features,and then uses the wavelet feature network to project and scale the wavelet features to obtain the multi-resolution wavelet feature close to the convolution feature,and finally the wavelet feature and the feature extraction network convolution feature are characterized.Fusion is used to supplement the convolution feature information and perform subsequent detection operations.In this method,an adaptive channel attention module is designed,and 12-dimensional channel selection is performed on the image wavelet features.With a small amount of calculation and no additional supervision,the network detection accuracy is improved.Through experimental comparison,using different feature fusion methods to fuse wavelet features and convolution features,and based on selecting features at different positions in the process of obtaining wavelet features,the method finally achieved more than 1% on multiple small object data sets.The accuracy is improved,and a higher overall detection accuracy is ensured,which proves that the method is competent for engineering applications.(3)This paper designs a small object detection system based on the attention model.Starting from the actual needs of user engineering,this system has designed a user-friendly graphical interface and an efficient batch detection model;and realized three modules: user management module,data management module and algorithm execution module.The user interface of the system designed in this paper is concise and easy to understand,the operation logic is efficient,the system is easy to transplant,and it is easy to maintain.