Research On The Application Of Deep Network Image Classification Methods In Wood Sorting And Cell Segmentation

Image classification is an image processing method that distinguishes different classes of objects.At present,image classification method based on deep learning is often used to classify different objects.Compared with traditional image classification methods,this method does not require artificial design or selection of feature descriptors.It is an end-to-end method of high robustness and accuracy.In this paper,the deep classification networks are applied to the real-time classification of rubber wood and the segmentation(pixel-level classification)of dense cells.And the deep classification networks have achieved good results.The research content of this paper is as follows:(1)The classification of wood types plays an important role in the construction industry and furniture manufacturing industry.In order to manufacture rubber wood furniture with highly uniform color and texture,rubber wood boards of different types should be sorted accurately.This paper uses the Res Net-based classification network to improve the classification performance of rubber wood images.Specifically,based on the residual structure of Res Net,we design an SSR(split-shuffle-residual)module by combining operations such as channel split,channel shuffle,and group convolution.In each SSR module,the input is split into two low-dimensional branches,and the channel shuffle operation is used to enable the information communication between the input and the two separated branches,which enlarges network capacity without increasing complexity.In addition,we present an SSR-based CNN that can learn features automatically from wood images for real-time classification of rubber wood boards.In order to evaluate the performance of the proposed classification network,we conduct experiments on 17,501 rubber wood images.The comprehensive experiments demonstrate that our algorithm outperforms other traditional machine learning classification methods and the common deep classification networks,yielding an accuracy of 94.86%.In addition,it only takes 26.55 ms to handle a single image,which can be employed for wood classification in real time.(2)In the time-lapse live imaging of plant cell microscopic images,automatic tracking of dense cells is an important task.However,the false cell segmentation result will cause serious errors in subsequent tracking procedure,and the existing cell segmentation methods cannot correctly segment the dense cells in the noise region.In this paper,VGG-based network is used to improve the accuracy of cell classification,further to improve the cell segmentation effect in noisy areas.And a two-stage cell segmentation algorithm is proposed to segment dense cells in the noise region.In the first stage,the cells are preliminarily segmented by the existing segmentation approach such as the watershed algorithm.In the second stage,the VGG-based classification network is used to divide the candidate cells of the first stage into two classes,and the candidate cells with clear boundaries(positive class)are retained.With the proposed two-stage cell segmentation algorithm,an accurate cell segmentation result can be used as the input of the next tracking process.Experimental results show that the VGGbased classification network has obtained the best candidate cell classification performance on three plant cell datasets,and the recall and precision are both above95%.In addition,by improving the accuracy of cell classification,the proposed twostage cell segmentation algorithm reduces the false segmentation of cells in the noise region,and further improves the performance of cell segmentation.