OBP-HRNetV2: An attention-enhanced network for robust safflower filament picking point localization in field environments

Le résumé fourni par la source

The core prerequisite for an intelligent safflower harvesting robot is accurate picking-point localization. Owing to the filament's diminutive dimensions and the indistinct boundary contours at the intended picking loci, the harvesting mechanism is precluded from attaining sub-pixel localization accuracy. This study instantiates an augmented HRNetV2 architecture, designated OBP-HRNetV2, dedicated to the simultaneous detection and metric localization of safflower filament abscission points. First, a self-attention module was integrated into the original HRNetV2 basic residual block to enhance the safflower contour feature detection capabilities. A network attention downsampling comprising three groups of convolutional layers incorporating channel attention and max pooling was employed to extract safflower features and strengthen the function mapping of interchannel dependencies. Additionally, an OCR module combined with a coordinate attention mechanism enhanced the recognition of different feature-region contours. Finally, key points such as the filament centroid, fruit ball centroid, and key contour vertices were selected using the Region of Interest extracted by OBP-HRNetV2. The pose of the safflower was determined using the centroid connection line slope and the key points of the fruit ball. The picking-point coordinates were calculated through image binarization processing. Experiments indicated that the algorithm achieved precise localization of safflower filaments. The OBP-HRNetV2 algorithm achieved an average Intersection over Union and pixel accuracy of 89.88% and 93.99%, respectively, showing improvements compared to other algorithms. In the safflower filament localization and harvesting experiments, the picking success rate was 86.66%. These methods demonstrated stronger robustness in field environments, providing technical support for the precise localization of safflower filament picking points in intelligent harvesting devices under field conditions.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.