Object detection is one of the most fundamental and actively researched topics in the field of computer vision, with significant practical value in industrial applications. Based on algorithmic architecture and detection processes, existing object detectors can generally be categorized into three types: two-stage detectors, single-stage anchor-based detectors, and single-stage anchor-free detectors. Two-stage detectors use neural networks to extract candidate regions and then perform precise classification and bounding box regression on each candidate region. Representative algorithms of this class of detectors include R-CNN and its improved version, Fast R-CNN [10]. Among these, Faster R-CNN achieved end-to-end training by introducing a region proposal network(RPN). Building on this foundation, various RPN-based improved algorithms have emerged, including Cascaded R-CNN and Mask R-CNN [11].

Two-stage detectors achieve higher detection accuracy but have slower inference speeds. Compared with two-stage detectors, single-stage detectors can simultaneously perform object classification and regression tasks, significantly improving inference speed but at the cost of reduced detection accuracy. Single-stage anchor-based detectors, such as RetinaNet [12] and YOLO [13] complete detection tasks using predefined anchor boxes, but this introduces additional hyperparameters and requires redundant anchor box pruning mechanisms during the processing. In contrast, single-stage anchor-free detectors, represented by FCOS [14] and GFL [15], abandon the anchor box design and use keypoint regression to directly predict bounding boxes.

Knowledge distillation is an efficient model compression technique that extracts and transfers implicit supervisory knowledge from a high performance teacher model to train a lightweight student model. It enables the student model to achieve performance close to that of the teacher model while maintaining its lightweight nature. Existing methods are primarily divided into two categories: feature-based distillation and response-based distillation.

Feature-based distillation focuses on the multi-scale features output by feature pyramid networks, extracting the implicit knowledge embedded in these intermediate-layer features. As an early knowledge distillation model targeting intermediate features, FitNet [16] demonstrated that the intermediate-layer features of the teacher network contain rich local structural and pattern information. This information can serve as an effective supervisory signal to guide the training of the student model. Lv et al. [17] introduced feature map transformation and alignment mechanisms to optimize the distribution differences between teacher and student features. Yang et al. [8] proposed a feature masking knowledge distillation method that can adaptively select key regions from the teacher’s feature maps. Li et al. [18] proposed CSAKD, which utilized the teacher’s attention information to generate spatial masks and enhanced student features by adding teacher global context. However, CSAKD relied on the teacher’s unilateral attention to determine the masking regions, ignoring the real time learning state and capability gap of the student model.

Response-based distillation focuses on the final network outputs generated by classification and regression tasks. The LD method [19] adopted the bounding box probability distribution proposed by GFL, used the KL divergence to align the outputs of the teacher and student models. It also designed a region-weighting strategy based on the intersection over union (IoU) ratio to optimize knowledge distillation performance. BCKD [20] reconstructed the classification branch output logic into multiple sets of binary classification results, effectively alleviating the mismatch between detection tasks and classification loss functions. Zhang et al. [21] proposed SAMKD, which introduced a spatial pyramid mechanism and used the difference between teacher and student features to weight logit distillation. Although SAMKD achieved hierarchical logit alignment at multiple scales, it uniformly applied KL divergence for logit distillation across all levels, thereby overlooking the fundamental task disparity between classification and regression heads in object detection.

However, feature-based distillation methods fail to fully exploit the capacity differences between teacher and student models, resulting in insufficient alignment between feature supervision and the student model. Response-based distillation methods overlook the task differences between the classification head and the regression head in detection tasks. In contrast, the method proposed in this paper further refines the features of the neck network and the final response outputs of the object detector by introducing feature difference and response difference modules.

The proposed BMGKD framework consists of a feature difference distillation module and a response difference distillation module. During the model training phase, both the teacher and student networks perform forward passes simultaneously. The parameters of the teacher network are fully frozen, while the student network learns by transferring knowledge from the teacher network and utilizing supervision from the detection task, iteratively updating its own parameters through backpropagation. Subsequently, feature maps are extracted from each layer of the teacher-student feature pyramid network and processed through the feature difference module to achieve dual alignment of global and local features. Meanwhile, the response difference module selectively distills response values from the teacher and student network branches, focusing on optimizing regions where the outputs of the teacher and student networks differ significantly in classification and regression. Additionally, we designed a dedicated loss function to bridge the task gap between the classification head and the regression head. The overall network architecture of BMGKD is shown in Fig. 1.

To ensure that the student model can fully learn from the teacher model, we propose a feature difference distillation module. This module uses a Gram matrix-based CKA loss function to distill the structural relationships between features rather than their absolute values. Furthermore, by combining spatial and channel-wise cosine similarity masks, it dynamically assigns weights to each distillation region based on the differences between the teacher and student models. This achieves comprehensive alignment between the global and local features of the teacher and student models, thereby improving the student model’s learning efficiency and its ability to understand features.

Traditional knowledge distillation methods for object detection optimize the response outputs of the classification and regression heads independently. This results in a significant amount of useful information being overlooked in both branches, further degrading the distillation performance. To overcome this limitation, the response difference distillation module enables collaborative optimization and joint distillation of the classification and regression tasks. For the predicted scores output by the classification head, we define the base classification mask Mt=max0<c≤C|pct|, where pct is the prediction score of the teacher model for the c-th target class. This mask is a static mask and cannot be adapted during the model training process. Therefore, based on the design of the feature-differential distillation module, we construct a dynamic classification mask Mt,s=max0<c≤C|Zct−Zcs| that employs joint teacher-student decision-making.

Based on the coordinate values output by the regression head, we use a region selection function to identify high quality prediction boxes and filter out low quality ones during the regression output refinement stage. Bs={brs}r=1R is the set of prediction boxes for a single image in the student model, containing a total of R prediction boxes. Bgt={br′gt}r′=1R′ is the set of ground-truth bounding boxes corresponding to the image, containing a total of R′ ground-truth boxes. The region selection function for the r-th prediction box brs generated by the student model is defined as:(11)f(brs)=I(maxr′∈{1,…,R}IoU(brs,br′gt)>t)where IoU(brs,br′gt) is the formula for calculating the standard merger ratio. I is the indicator function. When the maximum intersection ratio exceeds the set threshold t, the prediction box is considered a positive sample. Otherwise, it is considered a negative sample.

To ensure consistency between distillation and classification tasks, binary cross-entropy loss is used as the basis for designing the classification head distillation loss function. This minimizes potential inconsistencies arising from differences in loss functions, thereby achieving more efficient knowledge distillation. A joint masked distillation operation is performed on the classification head distillation loss function to resolve these discrepancies, as shown in Eq. (12):(12)Lcls=1Nc∑r=1RMt,sf(brs)∑k=1KLBCE(pr,kt,pr,ks)where Lcls represents the classification distillation loss, and LBCE represents the binary cross-entropy loss. Nc is the normalization factor. pr,kt and pr,ks represent the prediction scores for class k at position i in the teacher detector and student detector, respectively.

In addition, for the regression loss of the object detection head, we adopt the GIoU loss function, which takes into account not only the position and size of the bounding box but also its shape. The formula for the GIoU loss is shown in Eq. (13):(13)LGIoU(brt,brs)=1−I(brt,brs)U(brt,brs)+A(brt,brs)−U(brt,brs)A(brt,brs)where brt and brs represent the r-th bounding box from the teacher detector and the student detector, respectively. I(⋅) and U(⋅) denote the intersection area and union area of the two bounding boxes, respectively. A(⋅) represents the area of the smallest bounding box that contains the two rectangles. The regression head distillation loss function is given by Eq. (14):(14)Lreg=1Nr∑r=1RMt,sf(brs)LGIoU(brt,brs)

The overall loss function of the method proposed in this paper consists of five components: Ldet represents the original detection task loss of the student model. Lglobal represents the global feature alignment loss of the feature difference distillation module. Llocal represents the local feature alignment loss of the feature difference distillation module. Lcls represents the classification head distillation loss of the response difference distillation module. Lreg represents the regression head distillation loss of the response difference distillation module. The overall loss function is defined as shown in Eq.(15):(15)Lloss=Ldet+αLglobal+βLlocal+γLcls+ηLregwhere α, β, γ, and η are hyperparameters used to balance the individual loss components. The overall distillation loss is calculated solely based on the feature maps and response outputs, without involving the internal structure of the detector. Consequently, the proposed method exhibits excellent adaptability and can be applied to various mainstream object detection architectures.

To verify the effectiveness and efficiency of the proposed method, We conducted experimental validation using the MS COCO object detection dataset [23] and Pascal VOC 2007 dataset [24]. The training phase utilized the trainval135k subset, comprising 115,000 labeled images. The validation phase employed the minival validation subset, containing 5000 images, to assess model performance during training. Finally, the MS COCO test-dev 2019 dataset, which contains 20,000 images, was used for evaluation. Pascal VOC 2007 dataset contains 4952 images covering 20 object categories, including people, animals, vehicles, and indoor objects, among others.

The model was trained using the stochastic gradient descent (SGD) optimizer throughout the entire process, with a momentum coefficient of 0.9 and a weight decay coefficient of 0.0001. BMGKD is compatible with various mainstream object detection architectures, including the two-stage detector Faster-RCNN, the anchor-based single-stage detectors RetinaNet and GFL, and the anchor-free single-stage detector FCOS. All methods were trained on four NVIDIA A100 GPUs for 12 training epochs. The evaluation adopted the standard COCO metric of average precision (AP), and reporting the mean average precision (mAP) values at intersection-over-union (IoU) thresholds of 0.5 and 0.75, as well as the AP values for small, medium, and large-scale targets. All results are based on three independent replicates and report the average performance. The experiments demonstrate that the improvements made to BMGKD are stable and reliable, with a standard deviation of mAP between runs of less than 0.2%.

As shown in Table 1, we conducted a comprehensive performance comparison between our proposed method and other state-of-the-art methods (TBD [25], SamKD [21], CSAKD [18], FGD [8], BCKD [20], DrKD [26], CrossKD [27], HMKD [28], MLD [29], and PCD [30]) on the COCO dataset. On the two-stage Faster R-CNN, single-stage anchor-free GFL, and single-stage anchor-based RetinaNet detector architectures, we evaluate the performance of BMGKD under three teacher-student configurations: the standard capacity gap ResNet101 to ResNet50, the larger capacity gap ResNet101 to ResNet18, and the cross-architecture ResNet50 to MobileNetV3. In all experiments, BMGKD achieved the highest mAP and outperformed other models across all evaluation metrics, including AP50, AP75, APS, APM, and APL. Specifically, in the ResNet101→ResNet50 configuration, BMGKD achieved mAPs of 41.1%, 44.4%, and 40.4% on the three detectors, respectively, representing improvements of 2.7, 4.2, and 2.8 percentage points over the baseline, significantly outperforming existing methods. When the capacity gap between teachers and students widened further from ResNet101 to ResNet18, BMGKD still achieved mAPs of 37.4%, 40.0%, and 34.6%, respectively, maintaining a clear lead. In the more challenging cross-architecture scenario from ResNet50 to MobileNetV3, BMGKD ranked first with mAP scores of 36.9%, 39.4%, and 34.8%. These results fully demonstrate that BMGKD can consistently handle different detection architectures, varying degrees of capacity differences, and cross-backbone family transfer requirements, showcasing exceptional consistency and generalization capabilities.

As shown in Table 2, on the Pascal VOC 2007 dataset, when using ResNet101 as the teacher network to train the student network ResNet50, BMGKD achieved the best performance across all three detectors. Faster R-CNN, RetinaNet, and GFL achieved mAPs of 56.2%, 58.0%, and 58.3%, respectively. These results not only significantly outperformed the student baselines but also comprehensively surpassed the corresponding teacher models, maintaining a clear lead over existing methods such as AMD, BCKD, and SamKD. These results fully validate BMGKD’s effectiveness across different datasets, its generalization capability, and its ability to elevate student models to a level surpassing that of the teacher.

Table 3 compares the training overhead and accuracy gains of various distillation methods under the GFL ResNet101→ResNet18 setting. None of the methods alter the student architecture, and there are no additional parameters or computational costs during the inference stage. Compared to the response-based method CrossKD, the feature-based method FGD, and the hybrid method SamKD, BMGKD requires only 5.9 GFLOPs of additional computational effort and 1.14 times the training duration, achieving the highest mAP improvement of 4.6% and striking the optimal balance between training efficiency and accuracy.

To determine the optimal loss weights, this section conducted hyperparameter sensitivity experiments using the control variable method based on the GFL detection framework, with ResNet-101 serving as the teacher model and ResNet-50 as the student model. As shown in Table 4, the optimal hyperparameter settings for BMGKD were determined to be α=0.04, β=0.02, and γ=η=1.0. Under this configuration, the student model achieved a mAP of 44.4% on the GFL detection framework. Searches in different directions within the neighborhood of the optimal values did not reveal any performance gains. All deviations led to a consistent decline in detection accuracy.

To validate the effectiveness of each module in the BMGKD method, we used ResNet-101 and ResNet-50 as teacher-student backbone networks and conducted systematic ablation experiments based on the single-stage anchor-free detector GFL. The experimental results are shown in Table 5. When using only the global feature difference alignment module, the model’s detection accuracy improved by 2.8%. When using only the local feature difference alignment module, the model’s detection accuracy improved by 2.6%. After combining the global and local feature difference modules, the accuracy improvement increased to 3.8%. Furthermore, when the feature difference module was integrated with the response output difference module, the model achieved optimal performance, with an mAP improvement of 4.2% compared to the baseline model. The experimental results fully demonstrate that the complementary collaboration among the modules in the proposed method comprehensively improves the model’s final detection accuracy.

Fig. 2 shows the object detection visualization results of the teacher model, student model, and the proposed BMGKD method across different scenarios. The pure student model exhibits significant performance shortcomings. It not only makes redundant errors such as misclassifying a bench as an object but also suffers from issues such as inflated confidence scores and insufficient bounding box accuracy, resulting in a significant gap in detection performance compared to the teacher model. In contrast, the BMGKD method, through its bidirectional mask-guided knowledge distillation mechanism, successfully transfers the teacher model’s target discrimination capability to the student model. This not only eliminates false background detections in the student model but also calibrates the detection confidence to a level highly consistent with that of the teacher model, demonstrating greater robustness and detection reliability in scenarios with multiple targets and complex background interference.

To visually demonstrate the proposed method’s ability to focus on key regions of interest, we selected a representative test image from the MS COCO dataset and visualized the attention heatmap extracted from the GFL-ResNet50 student model after 12 epochs of training. The results are shown in Fig. 3. The intensity of the colors in the heatmap indicates the level of attention weighting, with red areas representing the highest level of attention and blue areas representing the lowest. As shown in the attention map for aircraft class detection, BMGKD can accurately focus on three key feature regions, the aircraft nose, engines, and tail, while filtering out redundant background interference. This fully validates that the collaborative masking design in the feature difference module can effectively enhance the student network’s ability to understand and distinguish target features.

As shown in Fig. 4, we extracted feature maps from seven layers B3 and B4 in the backbone, and P2 through P6 in the neck of the GFL detection network. We calculated the channel averages layer by layer to obtain two-dimensional activation matrices, which were then visualized using a Cividis color map, where higher brightness indicates stronger activation. After BMGKD distillation, the feature activation distributions of the student model at each layer closely align with those of the teacher model. In the deeper layers P3 to P5, the student model successfully reproduces the semantic regions and object contours of interest to the teacher, whereas the activations of the original student model were relatively diffuse and lacked sufficient semantic distinctiveness. These results demonstrate that the global and local alignment mechanisms within the feature difference distillation module effectively transfer the teacher’s tacit knowledge to the student model.