Adaptive cascade single-shot detector on wireless sensor networks

The target detection model based on convolutional neural networks has recently achieved a series of exciting results in the target detection tasks of the PASCAL VOC and MS COCO data sets. However, limited by the data set for a particular scenario, some techniques or models applied to the actual environment are often not satisfactory. Based on cluster analysis and deep neural network, this paper proposed a new Statistic Experience-based Adaptive One-shot Network (SENet). The whole model solved the following practical problems. (1) By clustering the existing image classification dataset ImageNet, a common set of target detection datasets is formed, and a data set named ImageNet iLOC is formed to solve the object detection. The problem of single and insufficient quantities in the task. (2) We use cluster analysis on the size and shape of objects in each sample, which solves the problem of inaccurate manual selection of suggested areas during object detection. (3) In the multi-resolution training and prediction process, we reasonably allocate the size and shape of the suggested frame at different resolutions, greatly improve the utilization rate of the proposed frame, reduce the calculation amount of the model, and further improve the real-time performance of the model. The experimental results show that the model has a breakthrough in accuracy and speed (FPS reaches 54 in the case of a 3.4% increase in mAP).