Research Of Weed Seeds Classification Based On Compressive Sensing Theory

Agriculture is the crucial industry of national economy, therefore, improving the qualityand output of agricultural products can contribute significantly to the economic development.Weeds can seriously affect the growth and development of crops, the propagation of weeds ismainly through the seeds. Therefore, there is a longing for highly efficient and reliable weedseeds classification methods. Fast implementation of the existing methods is of greateconomical and technical importance. This study incorporates the achievements ofcompressive sensing theory into the agricultural industry and proposes the robust algorithmframework for weed seeds registration and classification. After the registration of weed seedsimages, different dimension reduction methods and classifiers are taken into account in orderto compare the advantages of the classification method based on compressive sensing. Through the discussion about the validation of the test images and the error correction modelof weed seeds classification, a feasible idea could be provided in the design of the weed seedsrecognition system.The main contributions are as follows:(1) Weed seeds registration based on matrix reconstructionThis study uses the extended problem of compressive sensing theory, matrixreconstruction, representing the weed seeds registration as seeking an optimal set of imagedomain Euclidean transformation so as to get the low-rank matrix of recovered alignedimages. The registration problem can be equivalently solved by convex optimization.Therefore, weed seeds could be classified after registration even without extracting thetranslation, rotation, scale invariant features. Experimental results indicate that the registrationis satisfactory and robust to some corruption and occlusion. In addition, the contrastclassification tests verify the importance of registration for the classification algorithm.(2) Sparse representation for weed seeds classificationThis study is carried out on80categories of well-aligned weed seeds images to do theclassification work. The classification algorithm is based on compressive sensing theory,which has been applied to the field of machine learning. And the test image is expressed as alinear combination of all the training samples after solving the convex optimization problemofl1minimization. Four dimension reduction methods and three classifiers are used to compare the results which are done by k-fold cross validation method. The results show thatthe classification based on compressive sensing has certain advantages, the highestrecognition rate achieves98.75%. In addition, this paper briefly discusses the validation of thetest images which is the issues not covered in previous studies.(3) Research on error correction model of weed seeds classificationWeed seeds are prone to have mouldy blocks and disease spots due to the damp andmicrobial infections. Besides, the acquisition of weed seeds images could introduce errors invarious degree. This study adds contiguous occlusion or pixel corruption to simulate thecondition in which the seeds may occur in the real world. In order to deal with theseconditions, this paper solves a compound optimization problem which is overdeterminedlinear system. The classification results reveal the correctness of error correction model whichis based on compressive sensing theory.