A method, device and storage medium for target detection based on fixed point modeling
By modeling the fine-tuning process of the target detector as a fixed-point iteration, and employing a two-layer structure and a fine-tuning perceptual gradient strategy, the problem of parameter redundancy and training instability in existing target detectors is solved, thereby improving detection performance and reducing GPU memory requirements.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV
- Filing Date
- 2023-06-14
- Publication Date
- 2026-06-19
AI Technical Summary
Existing object detectors suffer from parameter redundancy during fine-tuning. Gradient estimation based on simple estimation and RNN expansion cannot effectively utilize fixed-point modeling, leading to unstable training and low efficiency.
The fine-tuning process of the target detector is modeled as a fixed-point iteration. A two-layer target detector, DEQDet, is used, including an initialization layer and a fine-tuning layer. Stable fine-tuning results are constructed by fine-tuning perturbations and gradients, and training is performed using fine-tuned perceptual gradients.
It significantly improves the performance of the object detector, reduces the number of parameters and memory requirements, and achieves higher detection accuracy and fewer training epochs on the COCO dataset.
Smart Images
Figure CN116824234B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer software technology and relates to target detection and fixed point theory. Specifically, it is a method, device and medium for target detection based on fixed point modeling. Background Technology
[0002] Object detection, a widely used traditional computer vision task, generally follows an initialization-fine-tuning strategy in existing object detectors, whether two-stage or query-based. This involves initializing the detection results from image features and then continuously fine-tuning to achieve the final result. However, the fine-tuning process in object detectors uses decoder layers with non-shared parameters. These decoder layers essentially fine-tune the output of the previous layer, and their functions are essentially the same. Therefore, these decoder layers can be compressed into a single layer, forming an iterative RNN paradigm. Furthermore, some research indicates that neural networks with near-infinite depth tend to stabilize. Fixed-point theory states that as long as the state update function is a compression mapping, any initialized hidden state of each node can converge to a fixed point, i.e., a fixed point. Therefore, the fine-tuning process can be modeled as fixed-point iteration to achieve better fine-tuning results. Additionally, due to the characteristics of fixed-point modeling, storing intermediate layers and some parameters during the optimization process can be avoided, significantly reducing GPU memory usage.
[0003] Introducing the fixed-point modeling paradigm into object detection requires constructing the gradient for training the object detector based on the implicit function formula of the fixed point. Previous research has introduced fixed-point modeling into low-level visual tasks, such as optical flow estimation, through simple gradient estimation. However, the high-dimensional semantic understanding and sparse nature of objects in object detection make the simple gradient estimation used in low-level dense tasks unsuitable for detectors. Algorithms based on simple estimation even outperform the original detector. Gradient estimation based on RNN iterative expansion suffers from unstable training and large gradient fluctuations due to the instability of RNN gradients. Gradient estimation based on implicit functions requires even longer iterations and cannot capture the fine-tuning nature of the fine-tuning layer, thus making it unsuitable for training. Summary of the Invention
[0004] The problem this invention aims to solve is that existing target detectors based on initialization-fine-tuning strategies have redundant parameters and cannot fully exploit the performance of the decoder layer in the initialization-fine-tuning paradigm. In particular, simple estimation, RNN expansion, and gradient estimation of implicit functions cannot effectively train target detectors based on fixed-point modeling.
[0005] The technical solution of this invention is as follows: a method for object detection based on fixed-point modeling, which models the fine-tuning process of the object detector as fixed-point iteration to obtain a fixed-point modeling object detector DEQDet. The object detector DEQDet first extracts features from the image through a backbone network to obtain a multi-scale feature pyramid. The feature pyramid and a learnable query are then fed into a fixed-point modeling object detection decoder. The object detection decoder consists of two layers: the first layer is an initialization layer, used to obtain preliminary detection results related to the image content and perform fixed-point initialization; the second layer is a fine-tuning layer, which performs fixed-point iteration on the initialization results to obtain stable fine-tuning results. Finally, the stable fine-tuning results are decoded into bounding boxes and classification confidence scores of the detected objects through an FFN structure. The training strategy for the object detector DEQDet includes fine-tuning the perceptual gradient based on two-step fixed-point expansion and fine-tuning the perceptual perturbation based on multi-step noise.
[0006] The fine-tuning layer perceives perturbations by adding Gaussian noise to each query at random probability during the fixed-point iteration fine-tuning process. The Gaussian noise is then mapped to the latent space of the query through the fine-tuning layer to simulate the noise perturbations during the fixed-point iteration process. This allows the fine-tuning layer to perceive the fine-tuning process and the essence of the fixed-point iteration, thus obtaining a stable point.
[0007] The fine-tuning perception gradient is constructed by expanding the gradient structure through two fixed points, allowing the fine-tuning layer to perceive the fine-tuning process from the backpropagation gradient.
[0008] Furthermore, the inference process of the target detector DEQDet based on fixed-point modeling is as follows:
[0009] 1) Data processing, input image Image∈R H·W Normalization is performed, with the normalized mean being μ and the normalized variance being σ.
[0010]
[0011] 2) Feature extraction: The normalized image... The input is fed into the backbone network, and a multi-scale feature pyramid is extracted:
[0012]
[0013] 3) Initialize the detection results. The initialization layer is represented as G(X,Y). 0- |ω), where ω represents the learnable parameters in the initialization layer. The G function in DEQDet is the same as the decoder layer in the baseline model. Y 0- For image content-independent learnable queries, including position vectors p∈R initialized to global size.4 And the learnable content vector q∈R d That is, Y 0- = (p,q), the initialization layer is obtained by using the feature pyramid X = [X0,X1,...X] of the image. L The interaction of ] makes Y 0- Update the initial Y0 to be related to the image content.
[0014] Y0=G(X,Y 0- |ω)
[0015] 4) Fixed-point iteration: The fine-tuning layer is defined as the implicit function of the fixed-point iteration, θ is the learnable parameter of the fine-tuning layer, and the F function in DEQDet is the same as that in the decoder layer of the baseline model, expressed as:
[0016] Y * =F(X,Y) * |θ)
[0017] The iterative process of the fine-tuning layer starts from Y0, using Y t+1 =F(X,Y) t Iterate using |θ) to form the sequence [Y0, Y1, Y2...Y]. T ], where Y T The stable point Y of the iteration * T is the maximum number of iterations;
[0018] 5) Decode the detection result to obtain Y. * =(p * ,q * After that, q is passed through FFN. * Decode the classification confidence cls for object detection, and then... * Decode the bounding box coordinates (bbox) to complete object detection:
[0019] Y * =(p * ,q * )
[0020] bbox = decode(p * )
[0021] cls = FFN(q * ).
[0022] Furthermore, the training strategy specifically includes: a training loss function comprising Focal Loss for classification and L1 loss and Giou Loss for bounding box regression; a Hungarian algorithm for bipartite graph matching for positive and negative sample allocation during training; and a classification loss only for negative samples. The training employs the AdamW optimizer and a learning rate decay strategy based on the number of training steps. After decoding the initialization result Y0 obtained from the initialization layer, the loss is calculated with the image label, and the gradient is backpropagated. The following configuration is used during training:
[0023] Fine-tuning perturbation: The iteration process of the fine-tuning layer starts from the initial fixed point Y0 and continuously calls the implicit function of the fine-tuning layer for iteration. During the call, Gaussian noise is mapped to the implicit space of the fixed point iteration through the fine-tuning layer, and noise is added randomly. During the iteration process, for the content vector, the magnitude of the vector before and after the addition remains unchanged, and it is linearly mixed with the noise. For the position vector, the position vector is first decoded into a bounding box structure, and then noise is added to the diagonal points of the bounding box. After adding noise, if the diagonal points are flipped, the coordinates of the diagonal points are swapped. Then the bounding box with added noise is re-encoded into a position vector. After adding noise, the iteration is performed again.
[0024] Fine-tuning the perceptual gradient: By expanding the fixed-point formula in two steps, the gradient structure of the perceptual fine-tuning process is constructed:
[0025] Y * =F(X,F(X,Y) * |θ)|θ)
[0026] Taking the implicit derivative of the above formula, let H(X,Y) * |θ)=F(X,F(X,Y * |θ)|θ), we have:
[0027]
[0028] Where I is the identity matrix and its Jacobian inverse. Using a simple approximation: get:
[0029]
[0030] The gradient structure is obtained by simplification:
[0031]
[0032] (·) represents the parameter θ in X or F. Let Y be a fixed point * The gradient of (·) is used for backpropagation of the network to calculate the gradient and update the parameters in the neural network;
[0033] In this process, to supervise the fixed-point iteration, the supervision position Ω supervises the intermediate results every K iterations, for a total of c iterations. Supervision signals are also applied to the first iteration and the final result.
[0034] Ω = [1,K,2K,...cK,T]
[0035] The supervised gradient uses a fine-tuned perceptual gradient. The gradient after backpropagation is not updated immediately, but the computation graph is preserved and gradients are accumulated until the fixed-point iteration process is completed, after which unified optimization is performed.
[0036] The present invention also provides an electronic device, the electronic device including a processor and a memory, the memory storing at least one instruction or at least one program, the at least one instruction or the at least one program being loaded and executed by the processor to realize the target detection method based on fixed-point modeling as described above.
[0037] The present invention also provides a computer-readable storage medium storing at least one instruction or at least one program, wherein when the at least one instruction or at least one program is executed, the target detection method based on fixed-point modeling as described above is implemented to obtain the target detector.
[0038] This invention introduces fixed-point modeling into the field of object detection, constructing the object detector DEQDet. Compared to existing detectors, the decoder of this invention is simpler, with only two layers: one for initialization and the other for fixed-point iterative fine-tuning. Based on the training strategy in this invention, the performance of the object detector in the initialization-fine-tuning paradigm is maximized. The embodiments of this invention use the Adamixer object detector as the experimental benchmark, and experiments are conducted on the COCO object detection dataset. Experimental results show that with ResNet50 as the backbone network and 300 queries, DEQDet can achieve a detection mAP of 49.5 and a small object detection AP of 33.0, while requiring only 24 epochs for training, 12 epochs less than the experimental benchmark. With ResNet50 as the backbone network and 1x (12 epochs) of 100 queries, DEQDet can achieve a detection mAP of 46.0, an improvement of 3.3 mAP compared to the benchmark (42.7 mAP).
[0039] Compared with the prior art, the present invention has the following advantages: The present invention introduces fixed-point modeling for the first time in the field of object detection. Combining the proposed gradient based on fine-tuning perception and the strategy based on fine-tuning perturbation, it trains the object detector for the high-dimensional semantic understanding characteristics and sparse nature of the object detection problem, which greatly improves the performance of the object detector, while reducing the number of parameters and the GPU memory required for training. Attached Figure Description
[0040] Figure 1 This is a schematic diagram of the reasoning process of the method of the present invention.
[0041] Figure 2 This is a schematic diagram of the training process of the method of the present invention. Detailed Implementation
[0042] This invention proposes a target detector based on fixed-point modeling, DEQDet. It models the fine-tuning process of existing target detectors as fixed-point iteration. DEQDet first extracts features from the image through a backbone network, obtaining a multi-scale feature pyramid. The feature pyramid and a learnable query are then fed into a fixed-point modeling-based target detection decoder. The decoder consists of two layers: an initialization layer (Initialization Layer) to obtain preliminary detection results related to the image content and perform fixed-point initialization; and a refinement layer (Refinement Layer) to iterate the initialization results using fixed-point modeling, obtaining stable fine-tuning results. Both the initialization and refinement layers include positional encoding, a self-attention layer (SA) for feature aggregation interaction between queries, and an adaptive sampling layer (AS) for feature aggregation from multi-scale image features. After the feature pyramid passes through the two decoder layers, the final query result—the stable fine-tuning result output by the fixed-point iteration layer—is decoded into bounding boxes and classification confidence scores of the detected objects using an FFN structure. To achieve superior performance, this invention proposes fine-tuning perceptual gradients and perceptual perturbations to train a target detector based on fixed-point modeling. Embodiments of this invention are tested on the COCO target detection dataset using AdaMixer and SparseRCNN as baseline networks, respectively, demonstrating significant performance improvements over the benchmark models.
[0043] The training process of the specific implementation of this invention includes data processing, backbone network feature extraction, detection query initialization, fixed-point iteration for fine-tuning perceptual perturbations, and construction of perceptual gradients. The algorithm inference process includes backbone network feature extraction, detection query initialization, and fixed-point iteration. To further illustrate the technical solution of this invention, the inference process of the object detector DEQDet is described first, followed by the training process of the object detector DEQDet.
[0044] The inference flow of the target detector DEQDet constructed in this invention is as follows: Figure 1 As shown, the explanation is as follows.
[0045] 1) Data processing, input image Image∈R H·W Normalization is performed, with the normalized mean being μ and the normalized variance being σ.
[0046]
[0047] 2) Feature extraction: The normalized image... The input is fed into the backbone network, and a multi-scale feature pyramid is extracted.
[0048]
[0049] 3) Initialization of detection results: The initialization layer is represented as G(X,Y). 0- |ω), where ω is a learnable parameter in the Initialization Layer. The G function in DEQDet is the same as the decoder layer in the baseline model. Y 0- For image content-independent learnable queries, which include a position vector p∈R initialized to global size. 4 And the learnable content vector q∈R d That is, Y 0- = (p, q).
[0050] The initialization layer combines image features X = [x0, x1, ... x...]. L The interaction of ] makes Y 0- Update the initial Y0 to be related to the image content.
[0051] Y0=G(X,Y 0- |ω)
[0052] 4) Fixed-point iteration: The refinement layer is defined as the implicit function of the fixed-point iteration. Let θ be the learnable parameter of the refinement layer. The F function in DEQDet is the same as that in the decoder layer of the baseline model, and can be expressed as:
[0053] Y * =F(X,Y) * |θ)
[0054] The F function is the function for the fine-tuning layer, Y * The above formula represents the stable point, and it reflects the fixed-point iteration to obtain a stable fine-tuning result.
[0055] The iteration process of the Refinement Layer starts from Y0, using Y t+1 =F(X,Y) t Iterate using |θ) to form the sequence [Y0, Y1, Y2...Y]. T ], the final Y T That is, the stable point Y of the iteration. * T is the maximum number of iterations.
[0056] 5) Decoding the detection results, Y * =(p * ,q * ), p * q * Let Y represent the iteratively stable position vector and content vector, respectively, to obtain Y. * Then, q is transferred via FFN. * Decoding into classification confidence for object detection and p * Decoding the coordinates into bounding boxes yields the bounding box and classification confidence score of the detected object, thus completing object detection.
[0057] Y * =(p * ,q * )
[0058] bbox = decode(p * )
[0059] cls = FFN(q * ).
[0060] The training strategy of the present invention includes fine-tuning the perceptual gradient based on two-step fixed-point expansion and fine-tuning the perceptual perturbation based on multi-step noise.
[0061] The Refinement Perceived Perturbation (RAP) simulates the noise perturbation during the fixed-point iteration process by adding Gaussian noise to the Query at each step with random probability during the fixed-point iteration fine-tuning process of the Refinement Layer. The Refinement Layer then maps the Gaussian noise to the latent space of the Query, thereby enabling the Refinement Layer to perceive the fine-tuning process and the essence of the fixed-point iteration.
[0062] The Gradient Awareness Adjustment (RAG) constructs a gradient structure through two-step fixed-point expansion, allowing the Refinement Layer to sense the fine-tuning process from the backpropagation gradient, making the Refinement Layer more suitable for fixed-point iteration.
[0063] The training strategy process of this invention is as follows: Figure 2 As shown.
[0064] The loss functions used during training include Focal Loss for classification and L1 loss and Giou Loss for bounding box regression. Positive and negative sample assignment during training employs bipartite graph matching, commonly used in Query, with only classification loss applied to negative samples. The AdamW optimizer and the StepLRscheduler learning rate decay strategy based on the number of training steps are used during training.
[0065] 1) Data processing, for images Image∈R H·W Data augmentation is performed using corresponding annotations. Under a 1x training setting, only image flipping is performed; otherwise, image scaling is applied. Normalization is then applied, with a mean of μ and a variance of σ.
[0066]
[0067] 2) Feature extraction: The normalized image... The input is fed into the backbone network to extract a multi-scale feature pyramid. In this invention, L=4 is used.
[0068]
[0069] 3) Initialization of detection results: The Initialization Layer is represented as G(X,Y). 0- |ω), where ω is a learnable parameter in the Initialization Layer. The G function in DEQDet is the same as the decoder layer in the baseline model. Y 0- For image content-independent learnable queries, which include a position vector p∈R initialized to global size. 4 And the learnable content vector q∈R d That is, Y 0- = (p, q). The Initialization Layer uses image features X = [x0, x1, ... x...] L The interaction of ] makes Y 0- Update the initial Y0 to be related to the image content.
[0070] Y0=G(X,Y 0- |ω)
[0071] After decoding the initialization result Y0 as described in inference process 5), the loss is calculated with the image label and gradient is backpropagated.
[0072] 4) Fine-tuning the perception fixed-point iteration: The Refinement Layer is defined as the implicit function of the fixed-point iteration, θ is the learnable parameter of the Refinement Layer, and the F function in DEQDet is the same as that in the decoder layer of the baseline model, which can be expressed as:
[0073] Y * =F(X,Y) * |θ)
[0074] The iteration process of the Refinement Layer starts from Y0, using Y t+1 =F(X,Y) t |θ) iteration, during the call process, randomly add noise:
[0075] 4.1) For the content vector q, keep the vector magnitude unchanged before and after the addition, and linearly mix it with Gaussian noise ∈, where σ is the linear mixing coefficient.
[0076]
[0077] 4.2) For the position vector p, first decode the position vector into a bounding box structure, then add noise to the top and bottom right corners of the bounding box. After adding noise, the top left and bottom right corners may be flipped. Therefore, for the flipped case, swap the top left and bottom right coordinates, and then encode the bounding box with added noise into a position vector.
[0078] After adding noise, the fixed-point iteration of the fine-tuning layer yields the sequence [Y0, Y1, Y2...Y]. T The training objective is to obtain the final iteration result of Y. T For the stable point Y * Fine-tuning the gradient by expanding the fixed-point formula in two steps constructs a gradient structure that can sense the fine-tuning process:
[0079] First, by expanding the fixed-point formula in two steps, the gradient structure of the sensing fine-tuning process is constructed:
[0080] Y * =F(X,F(X,Y) * |θ)|θ)
[0081] Taking the implicit derivative of the above formula, let H(X,Y) * |θ)=F(X,F(X,Y * |θ)|θ), we have:
[0082]
[0083] Let I be the identity matrix and its Jacobian inverse. Using a simple approximation:
[0084] get:
[0085]
[0086] Simplifying, we get:
[0087]
[0088] The above formula is the network gradient. (·) can be the parameter θ in X or F. Here, based on the chain rule, when taking the derivative, if the final gradient is with respect to θ, then (·) is θ. If it is with respect to the backbone, then it is necessary to first differentiate to X, and then differentiate X with the chain rule to the backbone parameter. Let Y be a fixed point * The gradient of (·) is used for backpropagation of the network to calculate the gradient and update the parameters in the neural network.
[0089] To fully monitor the fixed-point iteration process, this invention monitors the intermediate results once every K iterations, for a total of c iterations. Simultaneously, monitoring signals are also applied to the first iteration and the final result. The monitoring position Ω is:
[0090] Ω = [1,K,2K,...cK,T]
[0091] If the iteration number t is in Ω, then Y t The decoding process described in step 5) is performed, and the loss is calculated with the image annotations. The fine-tuned perceptual gradient is constructed and the gradient is backpropagated.
[0092] The supervised gradient employs a fine-tuned perceptual gradient. The gradient after backpropagation is not updated immediately; instead, the computational graph is preserved, and gradients are accumulated until the fixed-point iteration process is complete, after which unified optimization is performed. In this invention, the gradients of the network parameters are not updated immediately. Instead, the gradients generated by the losses at fixed points during training, including Focal Loss for classification and L1 and Gieu Loss for bounding box regression, are accumulated, and unified optimization is performed after iteration is complete.
[0093] In the COCO dataset object detection task, performance was evaluated on val. Using AdaMixer as the baseline model, the fixed-point hidden function F and initialization function G were constructed using the decoder layer in AdaMixer. With ResNet50 as the backbone and 300 queries, DEQDet achieved a detection mAP of 49.5 and an AP for small objects of 33.0, while requiring only 24 epochs for training, 12 epochs less than the experimental baseline. With ResNet50 as the backbone and 1x (12 epochs) of 100 queries, DEQDet achieved a detection mAP of 46.0, a 3.3 mAP improvement over the baseline (42.7 mAP). Using SparseRCNN as the baseline model, the DEQDet proposed in this invention constructs the fixed-point hidden function F and the initialization function G using the decoder layer in SparseRCNN. With ResNet50 as the backbone network and 300 queries, the DEQDet proposed in this invention can achieve an mAP detection result of 47.0, while only using 24 epochs for training, which is 12 epochs less than the experimental baseline.
Claims
1. A method for object detection based on fixed point modeling, characterized by The fine-tuning process of the baseline model of the object detector is modeled as fixed-point iteration, resulting in the fixed-point modeling-based object detector DEQDet. DEQDet first extracts features from the image through a backbone network, obtaining a multi-scale feature pyramid. The feature pyramid and a learnable query are then fed into the fixed-point modeling-based object detection decoder. The decoder consists of two layers: an initialization layer to obtain preliminary detection results related to the image content and perform fixed-point initialization; and a fine-tuning layer to iterate the initialization results using fixed-point iteration to obtain stable fine-tuning results. Finally, the stable fine-tuning results are decoded into bounding boxes and classification confidence scores of the detected objects through an FFN structure. The training strategies for DEQDet include fine-tuning the perceptual gradient based on two-step fixed-point expansion and fine-tuning the perceptual perturbation based on multi-step noise. The fine-tuning layer perceives perturbations by adding Gaussian noise to each query at random probability during the fixed-point iteration fine-tuning process. The Gaussian noise is then mapped to the latent space of the query through the fine-tuning layer to simulate the noise perturbations during the fixed-point iteration process. This allows the fine-tuning layer to perceive the fine-tuning process and the essence of the fixed-point iteration, thus obtaining a stable point. The fine-tuning sensing gradient is achieved by constructing a gradient structure through two fixed-point expansions, allowing the fine-tuning layer to perceive the fine-tuning process from the backpropagation gradient. The training strategy specifically includes: a training loss function comprising Focal Loss for classification, and L1 loss and Giou Loss for bounding box regression; bipartite graph matching for positive and negative sample assignment during training; classification loss used only for negative samples; an AdamW optimizer and a learning rate decay strategy based on the number of training steps; and initialization of the initialization layer. After decoding, calculate the image's tag. And backpropagate gradients, using the following configuration during training: Fine-tuning the perturbation: The iterative process of the fine-tuning layer starts from the initial fixed point. Starting from this point, the process iterates by continuously calling the implicit function of the fine-tuning layer. During the call, Gaussian noise is mapped to the latent space of the fixed-point iteration through the fine-tuning layer, and noise is added randomly. During the iteration, for the content vector, the magnitude of the vector before and after the addition remains unchanged, and it is linearly mixed with the noise. For the position vector, the position vector is first decoded into a bounding box structure, and then noise is added to the diagonal points of the bounding box. After adding noise, if the diagonal points are flipped, the coordinates of the diagonal points are swapped. Then the bounding box with added noise is re-encoded into a position vector. After adding noise, the iteration is performed again. Fine-tuning the perceptual gradient: By expanding the fixed-point formula in two steps, the gradient structure of the perceptual fine-tuning process is constructed: Taking the implicit differentiation of the above equation, we have and in Let be the identity matrix, and let Jacobi inverse be the inverse of the matrix. Using a simple approximation: ,get: , The gradient structure is obtained by simplification: for Or the parameters in F , Fixed point right The gradient is used for backpropagation of the network to calculate the gradient and update the parameters in the neural network. In this process, to supervise the fixed-point iteration, intermediate results are supervised once every K iterations, for a total of c iterations. Supervision signals are also applied to the first iteration and the final result. The supervision position Ω is: The supervised gradient uses a fine-tuned perceptual gradient. The gradient after backpropagation is not updated immediately, but the computation graph is preserved and gradients are accumulated until the fixed-point iteration process is completed, after which unified optimization is performed.
2. The target detection method based on fixed-point modeling according to claim 1, characterized in that: The inference process of the target detector DEQDet based on fixed-point modeling is as follows: 1) Data processing, input image Normalization is performed, and the normalized mean is... The normalized variance is : 2) Feature extraction, the normalized picture is input into the backbone network to extract a multi-scale feature pyramid: 3) Initialize the detection results; the initialization layer is represented as follows: , To initialize the learnable parameters in the layer, the G function in DEQDet is the same as that in the decoder layer of the baseline model. For image content-independent learnable queries, including position vectors initialized to global dimensions. and learnable content vectors ,Right now The initialization layer is achieved through the feature pyramid of the image. The interaction makes Updated to initial image content related , 4) Fixed-point iteration: The fine-tuning layer is defined as an implicit function of the fixed-point iteration. For the learnable parameters of the fine-tuning layer, in DEQDet The function is the same as the decoder layer in the baseline model, and is represented as: The iterative process of the fine-tuning layer from Departure, using Iteration to form a sequence ,in For the stable point of iteration T is the maximum number of iterations; 5) Decode the detection results to obtain... = Then, via FFN Decoding into classification confidence for object detection ,Will Decoded into bounding box coordinates Complete target detection: 。 3. An electronic device, characterized by The electronic device includes a processor and a memory, wherein the memory stores at least one instruction or at least one program, and the at least one instruction or at least one program is loaded and executed by the processor to implement the target detection method based on fixed-point modeling as described in claim 1 or 2.
4. A computer readable storage medium characterized by The computer-readable storage medium stores at least one instruction or at least one program, which, when executed, implements the target detection method based on fixed-point modeling as described in claim 1 or 2, thereby obtaining the target detector.