To realize the parallel robot sorting system to recognize and detect the location of fruit stalks of stacked fruit bunches with high accuracy under a large field of view, a small target low-contrast detection method based on CNC-YOLOv5s (improved YOLOv5s model) was proposed. For the small target of fruit stalks under the large field of view, the C3 convolution in the backbone network of YOLOv5s model was improved to C3-MSE convolution with parallel residual structure, and the feature information extracted from different convolution layers was fused by multiplexing to solve the problem of insufficient extracted texture feature information of fruit stalks in YOLOv5s model. When recognizing and detecting the location of low-contrast fruit stalks in stacked fruit bunches, the YOLOv5s model pays insufficient attention to the key features of the target due to the average distribution of weights during learning, and there is a problem of inadequate extraction of edge-level feature information of the stalks, which is solved by embedding CA (coordinate attention) module in the last three C3 convolutional layers of the backbone network and adding a Maximum pooling layer, to dynamically adjust the weights of YOLOv5s model for target feature learning, and enhance the attention to the target edge level feature information. To address the problem of slower convergence of the YOLOv5s model caused by the improvement of the model structure, CIoU (complete intersection over union) regression loss function in the head network of the YOLOv5s model was improved to one that is unaffected by the deviation of the target bounding box NWD to accelerate the convergence of the YOLOv5s model. The experimental results show that the mAP value of the improved CNC-YOLOv5s model reaches 95.64%, which is 4.79% higher than that of the original model. The mAP value is 19.4%, 13.02%, 7.33%, and 3.11% higher than that of the mainstream models, such as SSD, Faster RCNN, YOLOv4, and YOLOvX, etc., which verifies the effectiveness of the proposed method for the superiority of the proposed method to detect the position of fruit stalks in stacked bunches can better meet the sorting requirements of parallel robots, which can provide technical reference for the post-production intelligent sorting of grapes and other fruits.