IOCSegFormer: Enhancing Wheat Ears Counting in Field Conditions Through Augmented Local Features

The number of wheat ears under field conditions is pivotal for production assessment and yield forecasting. Due to the occlusion, overlap, and influence of different weather conditions, it is challenging to accurately count wheat ears with existing object counting methods, which overlook local features surrounding wheat ears. To address this issue, we propose a new deep neural network, named IOCSegFormer, to augment local features around wheat ears. Specifically, IOCSegFormer extends the existing IOCFormer object counting network by incorporating a segmentation branch responsible for generating a segmentation map delineating wheat ears. Within this branch, we introduce a local feature extraction module to enhance counting accuracy by capturing contextual information pertinent to the local surroundings. Experimental evaluations conducted on both a proprietary dataset and the Wheat Ears Detection Dataset demonstrate effectiveness of IOCSegFormer, yielding mean absolute errors of 8.39 and 12.42, and root mean square errors of 12.73 and 17.35, respectively. Comparative analysis against previous state-of-the-art methods underscores the superior performance of our model.