Wheat Head Detection In The Wild By Improved CenterNet++

The detection and identification of the wheat ear during its growth process is not only a crucial factor that affects the planting decision-making of farmers,but also plays an important role in crop yield estimation.Although the wheat ear detection methods proposed at present have achieved good detection results,they also face enormous challenges such as high density,scale changes,and lighting differences.Moreover,since it requires setting a reference box manually,these methods suffer from reduced accuracy in detecting small targets and partially obscured targets.In this article,we improve the Center Net++object detection model to detect wheat ears from aerial views in the field.Firstly,contrast adjustment and Gaussian blur are applied to images containing wheat ears for data augmentation to improve the generalization ability of the model.Secondly,the Slicing Aided Hyper Inference is used to divide the input image into several sub-images,which solves the problem of small target omission in the model.Finally,the Augmented Feature Pyramid Networks is used to reduce information loss of the highest level pyramid feature map,which further solves the problem of missing information in wheat ear detection.Meanwhile,this article abandons the idea of tiling reference boxes and introduces the Rep Points technology that represents targets at multiple points to achieve the purpose of ancho-free detection of wheat ears.Moreover,in terms of positive and negative sample distribution strategies,this article adopts an adaptive training sample selection method to ensure the quality of positive samples and eliminate the influence of hyperparameters.The experimental results on the test set images indicate that the method predicts with an average precision(AP₅₀)of 90.57%,and an average precision（AP_s）of58.77%for small targets.The accuracy in detecting tiny targets is significantly better than that of other traditional detection models.These results indicate that this method can meet the requirements of wheat ear counting in outdoor environments,providing reliable reference data for wheat yield estimation.The main work of this article is as follows:（1）In response to the wheat image detection under outdoor conditions with varying lighting and diverse types,this article improves the Center Net++wheat image detection model.The improved model has three typical characteristics:a)It is a anchor-free image detection model;b)It has strong robustness to complex environments;c)It has good detection ability for small targets in images.（2）This article proposes Slicing Aided Hyper Inference as a preprocessing step for convolutional neural networks to improve the information of obscured or small targets in images,which effectively improves the results of object detection by convolutional neural networks.（3）This article adopts the feature fusion method of Augmented Feature Pyramid Networks to enrich the extracted feature information,making the network model more sensitive to the target.（4）This article uses the Rep Points technique that represents the object with multiple points,which can be used for localization and identification.Rep Points can automatically learn the spatial position of the object box and the significant semantic features locally.They do not require reference boxes and belong to a anchor-free object detection method,which can be as effective as the reference anchor-based methods.（5）This article adopts the adaptive training sample selection（ATSS）method to select positive samples,which requires almost no hyperparameters and ensures the quality of selected positive samples.It is very suitable for box-free detection models.