Research on contour extraction and matching method of stump based on deep neural network
In order to find stolen trees timely and accurate,the public security organs need to compare the upper and lower surfaces of the stumps by searching for their common ground and confirming whether the two stumps belong to the same tree.However,due to different storage environments,there are significant differences in the color and texture of the upper and lower surfaces of the stumps.Due to the influence of felling method,there are also significant differences in the contour of the upper and lower of the stumps.The lower surface of the stump is also easily affected by wood chips and difficult to clean.In response to these difficulties,based on the previous research,this study used the chessboard as a feature and placed it on the surface of the stump,and used the PPYOLO_MobileNetV3 convolu-tional neural network to detect the chessboard in the images.After the corners in the chessboard were detected and sor-ted,a perspective transformation was performed to restore the original area and contour of the stump,and then used PP-LiteSeg network to extract the stump contour under the complex background,and then used CAE_ViT_base net-work to implement the stump contour matching.CAE_ViT_base network divided the input images into uniformly sized image blocks,and the training process of the decoder needed to pay attention to the characteristics of each block.The difficulty in matching the contour of the stump lies in the presence of local missing contours,and large local gradient errors.CAE_ViT_base network used the self supervised pre-training mechanism,which can effectively compensate for the above shortcomings.At the same time,a method of randomly rotating the sample images from multiple angles was adopted,so that the feature extraction of the images can maintain rotation invariance.The features extracted by this mechanism were higher in scale than gradient based features and local point set topology-based features,but lower than global features.The final matching degree given by the CAE_ViT_base network was relatively high,which was consistent with the proposed method of manually comparing with whether the contours of two logging piles match.Ex-periments were conducted on 344 stumps in this study,and the experimental results showed that the algorithm's suc-cess detection rate for the entire test set was 100％,the rate of success for checkerboard detection was 100％,the ac-curacy of contour extraction was 98.8％,and the accuracy of contour matching was 100％.In the process of extracting the contour of the logging pile images,errors are inevitable,and the upper and lower surface contours of the same logging pile often have one or even multiple segments that do not match.Therefore,matching methods based on local gradients,matching methods based on topological features of local point sets,and matching methods based on global features of contours are all infeasible.The results of this study have no false detections or mismatches.The entire process of contour matching required less than 30 s,which can meet the practical application requirements and has good application value.
deep neural networkchessboard detectioncontour detectioncontour matchingstump rectification
万方数据
维普
