To address the low accuracy and efficiency of apple maturity detection under large-scale, lighting, and occlusion conditions, an improved FPB-DETR detection model based on RT-DETR was proposed. Firstly, a frequency-adaptive dilated convolution (FADC) module was introduced into the backbone network to precisely focus on subtle color gradients, immature spots, and texture stripes on apple surfaces by resolving the conflict between effective receptive field and feature bandwidth, as well as overcoming the limitations of fixed dilation rates. Secondly, a polaformer-attention-based intra-scale feature interaction（Pola-AIFI) module was designed to mitigate the issues of negative value neglect and excessive information entropy, suppressing interference from target apples under varying environmental conditions. Finally, a bi-directional feature pyramid network（BIFPN) structure was introduced during the multi-scale fusion stage to optimize feature fusion efficiency and key information focusing capability, reducing ambiguity interference in maturity feature transmission. The results show that the precision, recall rate and average accuracy of the FPB-DETR model proposed in this study are 92.5%, 92.7% and 96.8%, respectively, which increases by 2.0%, 1.7% and 1.8%, respectively compared with the original model, and are superior to those of Faster R-CNN, YOLOv5m, YOLOv8m, YOLOv11m and YOLOv12m object detection models, significantly enhancing the detection capability of the model; The average detection time of the model is 31 ms, which meets the real-time detection requirements for apple maturity. This study realizes better detection effect by combining feature extraction, attention mechanism and multi-scale fusion, providing reference for the optimization design of intelligent harvesting robots.