Detection of Wheat Unsound Kernels Based on Improved ResNet

In the process of grain acquisition, the unsound kernels of wheat are traditionally detected manually. The determination method based on computer vision typically requires expensive equipment for image acquisition and has disadvantages of low recognition efficiency and difficulties in adhesion segmentation, which strongly limit the application in routine detections. In this paper, six kinds of wheat including sound kernel, broken kernel, sprouted kernel, injured kernel, moldy kernel and spotted kernel are considered as the samples. An image acquisition platform is built with low cost to capture wheat pictures. The designed two-kernels adhesion wheat segmentation algorithm based on concave-mask exhibits high accuracy, with the error rate of 0.93% for total 9988 wheat grains. By comparing the advantages and disadvantages of GoogleNet, DenseNet, IX-ResNet, Res2Net, this paper studies the optimization of depth, width, downsampling mode, convolution order, attention mechanism, receptive field. Finally a wheat unsound kernel detection method is proposed based on Res24_D_CBAM_Atrous. The Macro avg values of Precision, Recall and F1 are respectively 94%, 95% and 94%, which are increased by 3%-4% based on the original Res34. The prediction time is reduced by 220s, which can meet the rapid and accurate evaluation of wheat appearance quality. The method shows important theoretical significance and practical application value for wheat unsound kernel routine detection.