An improved image processing scheme for automatic detection of harvested soybean seeds

Physical quality assessment of harvested soybeans is one of the basic requirements in the commercial trade market. The image-processing-based automated system helps to achieve reliability in quality assessment over the manual inspection method. Image acquisition and object detection are the crucial blocks of such automated systems. In this paper, we propose an improved image acquisition system to acquire two-sided images of harvested soybean samples. During the object detection process, the watershed algorithm is commonly used to separate connected objects. However, the performance of the watershed algorithm degrades for harvested soybeans due to the loss of object information. We propose a linear staircase path approximation scheme (LSPA) to disconnect objects with higher accuracy. The LSPA scheme utilizes the topological features of connected objects, which is less sensitive to the local variation of object features. The LSPA scheme identifies the objects of harvested soybean samples with marginal variation in object count w.r.t the actual object count, which improves the object classification accuracy. The multi-class and binary classification jointly performed to improve the classification accuracy of sound quality soybean. Experimental results show that the detection accuracy of sound quality soybean is almost 90% for the proposed scheme, which is substantially higher than the existing scheme based on the watershed algorithm.