Design of a High Efficiency Soybean Segmentation Model via Bias Field Estimation with Fuzzy Clustering
Segmentation of Soybean chunks from varying illumination imagery requires design of pre-processing, filtering & adaptive threshold estimation models. These models are designed to incorporate various image types via use of machine learning techniques for detection of dynamic threshold levels. But they are highly context-sensitive, and cannot be applied to general purpose Soybean images that are captured from camera & other low-resolution sources. To overcome this limitation, a novel Bias Field Estimation with Fuzzy Clustering segmentation model is discussed in this text. The proposed model initially estimates multiple bias fields (MBF) for input images, and then clusters these fields via Fuzzy C Means (FCM) grouping model, which assists in application independent segmentation operations. The MBF model extracts multiple entropy levels from input images, which assists in representing these images in multispectral scales, thereby reducing its dependency on image quality. Lower quality scales are discarded via use of Naturalness Image Quality Evaluator (NIQE) levels, while higher quality scales are used for clustering process. The clustered images are examined via region-level structural parameters including maximum diameter, minimum diameter, eccentricity, and shape indices. Regions with abnormal shapes are discarded, while normal shaped regions are used for post-processing purposes. The model was evaluated on a large dataset of standard and camera-based Soybean images, and it was observed that the proposed model was capable of improving segmentation efficiency by 7.6% when compared with other state-of-the-art methods. It was also observed that the proposed model showcased higher accuracy, lower delay, and better precision when compared with other models, thus making it useful for a wide variety of on-field segmentation applications.