Dual Channel Attention Mechanism for Apple Localization Detection and Segmentation

Apples in orchards are surrounded by complex environments with varying light, weather, and overlapping shade, making accurate apple identification difficult. In order to improve the accuracy of apple localization detection and segmentation in complex environments, the Mask R-CNN model is adopted and enhanced in this paper. Firstly, a dual channel SE module is introduced and embedded into the ResNet-50 network for feature extraction to emphasise valid information and suppress irrelevant information, and to calculate the acquisition of different feature channel weights to enhance the expressiveness of the model; secondly, improving the anchor box generation mechanism in the regional suggestion network to reduce misidentification while reducing the number of anchor boxes, based on the actual situation of apple exposure ratio and area; finally, the boundary weighted loss function is then added to the segmentation error loss function to improve the accuracy of apple boundary segmentation and increase the precision of segmentation results. The experiment with apple images in a real orchard yielded the following results: 98.5% for Pixel Acc, 80.4% for Mean IoU, 61.6% for Class IoU, and 85.5% for Class Mean Acc. The results demonstrate that the performance of the improved model proposed in this paper is superior to that of the original model as well as some other widely used models, thereby increasing the precision of apple detection and segmentation in complex orchard environments.