A Small Target Recognition Algorithm Based on Improved SSD
Yasong Wang, Wenpeng Cui, Hongbo Yang · 2019
In the power industry, there are a lot of image and video data need to be analyzed to find out equipment abnormalities such as bolt looseness and out of stock. Using artificial intelligence technology for image analysis can reduce the artificial dependence, but the general object detection algorithms, such as SSD, can not get a good detection effect for small targets. In this paper, we proposed an improved SSD object recognition algorithm to improve the recognition accuracy of small targets, based on the characteristics of power image recognition requirements. The method can improve the ability of SSD framework to understand small objects and enhance the location information and semantic information of the small targets. The method introduce the deconvolution structure to act on the shallow feature layer Conv4_3 and then use pooling layer to get the fixed size. Finally the method made Conv14 upsampled to the same size, and make full use of the information of small targets. Experimental results show that the proposed method can achieve the good recognition accuracy.