Research On Object Detection Based On Convolutional Neural Networks

With the continuous development of deep learning,the target detection algorithm based on convolutional neural network has been significantly developed,and the algorithm has been greatly improved in both detection accuracy and speed.However,the current target detection algorithm based on convolutional neural network relies on the traditional Non-Maximum Suppression(NMS)algorithm for post processing.However,the traditional NMS algorithm is a greedy algorithm.When the detection frame outputted by the target detection model is post-processed,the redundant detection frame is simply suppressed according to a fixed threshold,so that the algorithm itself has certain limitations.Get the global optimal solution.To solve this problem,this paper designs a learning method to remove the redundant detection frame algorithm LSTM-NMS algorithm,and output a unique detection frame for each target in the image.The LSTM-NMS algorithm first fuses the coordinates,area,category scores,and feature vectors of the detection frame,so that the information is used to characterize the suggestion box,and then uses the attention mechanism to reconcile the feature information of the detection frame with the features of the local suggestion frame.The information is merged,and finally the merged feature information is input into the LSTM network for training,and a unique detection frame is output to the target in the image during the test.Experiments show that the Faster RCNN model based on LSTM-NMS algorithm has improved the detection performance on the PASCAL VOC dataset and the insulator dataset collected by the research group.In addition,due to the presence of small target insulators and large variations in rotation in transmission line insulator image.At the same time,the aspect ratio of the detection frame generated in the Faster RCNN model cannot match the shape of the insulator target in the image.Aiming at these problems,the improved Faster RCNN model is used to detect the insulator target in the image.The low-level feature map prediction layer is added based on the Faster RCNN model,and the spatial transformation module is embedded in the layer feature map,so that the model can effectively detect the rotation.Larger and smaller target insulators.In addition,the aspect ratio of the default detection frame is improved so that the generated detection frame aspect ratio can effectively match the insulator target in the image.Experiments show that the improved Faster RCNN model has a significant improvement in the insulator dataset collected by the research team.