An efficient target detection method, device, medium and system

Abstract

The present invention provides an efficient target detection method, including: obtaining a target detection training set; using the training set for pre-training a target detection network to obtain a full-precision target detection network; automatically adjusting the accuracy and parameters of the target detection network to obtain a mixed-precision target detection network under a specified compression rate; using the mixed-precision target detection network to process an image to be detected to obtain an image detection result. The present invention can automatically adjust the parameters of the designed target detection network based on given image data, reduce the model size, and does not require complicated manual design. It can adaptively adjust the model size under different computing and storage resources, while achieving a higher image detection accuracy, thereby ensuring the stability and practicality of the image detection method.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and image processing technology, and in particular to an efficient target detection method, device, medium and system Background Art

[0002] With the improvement of computing power and the development of big data, artificial intelligence has achieved rapid development in recent years, and the demand for image processing applications such as image classification, target tracking, and instance segmentation has also increased. Object detection plays an extremely important role in practical applications. It first searches for the position and shape of the objects in the image, and then determines which category these objects belong to. Unlike image classification tasks, object detection has to solve both positioning and classification problems, which is more complicated.

[0003] The development of object detection algorithms is mainly divided into one-stage and two-stage. The two-stage refers to the detection algorithm is completed in two steps, first obtaining the candidate area, and then classifying. The R-CNN series is the representative. In order to improve the performance and speed of the network, by improving the creation method of the candidate area, R-CNN, FastR-CNN and Faster R-CNN have appeared successively.

[0004] Among the two-stage algorithms, Faster-RCNN is the most commonly used network structure with the best overall performance. Although region-based detectors can guarantee relatively high accuracy, the extraction of candidate regions requires a cost in computing time. In contrast to two-stage detection, one-stage detection does not require a separate search for candidate regions. It can be said that detection and classification are done in one step, which greatly improves the processing speed. However, its accuracy is relatively low. Therefore, typical works include SSD, YOLO, and RetinaNet, all of which aim to balance accuracy and real-time processing speed.

[0005] In the first-stage detection work, RetinaNet is the network structure with excellent comprehensive performance and the focus of researchers. However, the deep neural network structure used for target detection tasks is very deep, requiring huge computing and storage overhead, which hinders the application of this technology in resource-constrained environments and scenarios with low latency requirements, such as mobile phones, surveillance cameras, and autonomous driving. In order to obtain an efficient target detection method, the original first-stage and second-stage structures need to be lightweight.

[0006] In the traditional target detection network structure, convolution based on 32-bit floating-point number calculation generally has a large redundancy. Reducing network parameters and precision can not only compress the model, save bandwidth and computing resources, but also speed up the inference speed, and significantly reduce the memory usage during operation. At the same time, many research works have pointed out that low-precision network models can also have good performance, which provides the possibility for model quantization. However, model quantization usually leads to information loss. Typically, in image classification tasks, the classification accuracy of the network after low-bit quantization will be significantly reduced. Mixed precision quantization can solve this problem to a certain extent. Mixed precision quantization selects appropriate bit widths for different parameters. The network weights with small impact on the classification results and large parameter volume are assigned low bits, which ensures both compression rate and accuracy. Adaptive bit width allocation calculates the importance of parameters according to the set criteria, and automatically gives the bit width of the corresponding parameters.

[0007] In summary, the existing efficient target detection methods still have the following shortcomings: (1) The network performance decreases significantly after compression; (2) The existing mixed precision quantization method for target detection requires manual specification of the bit width, resulting in a huge search space, computational and storage overhead, and there is still much room for performance improvement. Summary of the invention

[0008] In view of the defects in the prior art, an object of the present invention is to provide an efficient target detection method, device, medium and system.

[0009] According to one aspect of the present invention, there is provided an efficient target detection method, comprising:

[0010] Get the target detection training set;

[0011] Using the training images for pre-training of an object detection network to obtain a full-precision object detection network;

[0012] Automatically adjusting the accuracy and parameters of the target detection network to obtain a mixed-precision target detection network at a specified compression rate;

[0013] The mixed precision object detection network is used to process the image to be detected to obtain an image detection result.

[0014] Preferably, the using the training images for pre-training the target detection network to obtain a full-precision target detection network; and automatically adjusting the accuracy and parameters of the target detection network to obtain a mixed-precision target detection network at a specified compression rate, comprises:

[0015] Compute the quantitative sensitivity of network layers to assess the importance of network parameters;

[0016] Select the network layer with the lowest quantization sensitivity, and reduce the accuracy of the parameters of this layer by using the method of decreasing the accuracy step by step. Dynamically update the quantization sensitivity of the network layer with reduced parameter accuracy by using the L-BFGS algorithm;

[0017] Repeatedly select the network layer with the lowest sensitivity and iteratively adjust the accuracy of the layer parameters until the model compression rate of the entire network reaches the specified threshold, thus obtaining a mixed precision object detection network;

[0018] All training images are input into the mixed precision object detection network, and fine-tuned and trained by direct estimation until convergence for object detection.

[0019] Preferably, the quantization sensitivity can measure the importance of different network layer parameters and evaluate the impact of reducing the accuracy of network layer parameters on network performance, which is:

[0020] in:

[0021] n i represents the number of network parameters of the i-th layer, is the network parameter matrix of the i-th layer, Q(·) is the quantization function used to reduce the parameter accuracy, is the quantization perturbation, is the loss function, is the loss function About the parameter matrix of the i-th layer The average value of the trace of the Hessian array.

[0022] Preferably, the average value of the trace of the Hessian matrix is ​​the sum of the diagonal elements of the Hessian matrix corresponding to each parameter calculated using the Hutchinson method.

[0023] Preferably, the method of gradually decreasing the accuracy of network layer parameters refers to selecting a network layer with the lowest quantization sensitivity, and adjusting its parameter accuracy from high precision to low precision according to the mapping method of 32 bits → 8 bits, 8 bits → 6 bits, 6 bits → 4 bits, 4 bits → 3 bits, 3 bits → 2 bits, and 2 bits → 1 bit.

[0024] Preferably, the L-BFGS algorithm calculates the Hessian matrix of the current round according to the Hessian matrix of the previous round of the network layer and the weight difference and gradient difference of the parameters before and after the current round of training during training, and reduces the storage and calculation overhead by storing the weight difference and gradient difference vectors instead of the complete Hessian matrix:

[0025]

[0026] in: is the gradient difference vector of the network layer parameters before and after training, ΔW k =W k+1 -W k is the difference vector of network layer parameters before and after training,

[0027] Preferably, the fine-tuning training of the direct estimation includes:

[0028] Copy the obtained low-precision network with the specified compression rate, use one as the teacher network and the other as the student network; input all training images into the student network, and the student network will pass the gradients obtained during training back to the teacher network. The teacher network updates the network parameters according to the obtained gradients, and quantizes the network parameters according to the accuracy of the corresponding layer, and passes them back to the student network until convergence, and the final student network is obtained as the mixed precision target detection network.

[0029] According to a second aspect of the present invention, there is provided a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the efficient target detection method described in any one of the above items is implemented.

[0030] According to a third aspect of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-mentioned efficient target detection methods.

[0031] According to a fourth aspect of the present invention, a chip system is provided, comprising a processor, the processor being coupled to a memory, the memory storing program instructions, and implementing any of the above-mentioned efficient target detection methods when the program instructions stored in the memory are executed by the processor.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] The present invention can automatically adjust the parameters of the designed target detection network based on given image data, reduce the model size, and does not require complicated manual design. It can adaptively adjust the model size under different computing and storage resources, while achieving a higher image detection accuracy, thus ensuring the stability and practicality of the image detection method. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0035] Figure 1 A flowchart of an efficient target detection method according to an embodiment of the present invention;

[0036] Figure 2Schematic diagram of the structure of an efficient target detection method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several variations and improvements may be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0038] The present invention provides an embodiment, an efficient target detection method, comprising:

[0039] S100, obtaining a target detection training set;

[0040] S200, using the training image for pre-training of the target detection network to obtain a full-precision target detection network;

[0041] S300, automatically adjusting the accuracy and parameters of the target detection network to obtain a mixed-precision target detection network under a specified compression rate;

[0042] S400, using the target detection network to process the image to be detected to obtain an image detection result.

[0043] Based on the above embodiment, further optimization is performed, such as Figure 1 As shown, a flowchart of an efficient target detection method according to a preferred embodiment of the present invention is provided, including:

[0044] S0, the collected images are used for pre-training of the target detection network to obtain a full-precision network model;

[0045] S1, using the average value of the trace of the Hessian matrix of the Hutchinson initialized network layer, calculate the quantization perturbation of each layer parameter from 32 bits to 8 bits, and complete the initialization of the quantization sensitivity;

[0046] S2, using the method of gradually decreasing the bit width precision, each time select the layer with the lowest quantization sensitivity, and reduce its bit width precision according to the bit width mapping method of (32-8, 8-6, 6-4, 4-3, 3-2, 2-1);

[0047] S3, using the L-BFGS algorithm, dynamically updates the Hessian array and quantization sensitivity of the convolution layer that reduces the bit width accuracy each time to make it more accurate;

[0048] S4, determine whether the network has reached the specified compression rate, if it has reached it, enter S5; if not, return to S1;

[0049] S5, all training images are input into the target network that achieves the specified compression rate, and fine-tuned training is performed through direct estimation until convergence for target detection.

[0050] In order to better obtain the quantization sensitivity of the network layer, the present invention provides an embodiment to execute S1 to calculate the quantization sensitivity of the network layer for evaluating the importance of network parameters.

[0051] Quantization sensitivity refers to the sensitivity of a specified quantized layer to network performance. It estimates the impact of quantizing network layer parameters from high precision to low precision on network performance. The greater the impact on network performance after quantization, the more sensitive the network layer is. Quantization sensitivity is an indicator that estimates the importance of network layer parameters.

[0052] After theoretical derivation, the relationship between the loss function and the average value of the trace of the Hessian array is as follows:

[0053] when hour,

[0054] where n i represents the number of network parameters of the i-th layer, is the network parameter matrix of the i-th layer, Q(·) is the quantization function used to reduce the parameter accuracy, is the quantization perturbation, is the loss function, is the loss function About the parameter matrix of the i-th layer The average value of the trace of the Hessian array.

[0055] The product of the average value of the trace of the Hessian matrix and the quantization perturbation can be used as the quantization sensitivity. In fact, due to the different parameter amounts of different network layers, the network layer with more quantization parameters will compress the network more. Considering network compression and network performance comprehensively, a new quantization sensitivity is designed:

[0056]

[0057] First, the collected training images are used for pre-training of a one-stage target detection network to obtain a full-precision target detection network of RetinaNet, in which the network parameters are 32-bit floating point numbers.

[0058] According to the obtained full-precision target detection network, the Hutchinson method is used to calculate the average value of the trace of the Hessian array of each layer of network parameters. Since the Hessian array can be stabilized by using 200 training images, in order to save time and computing overhead, 200 input images are used and iterated 50 times to obtain the average value of the trace of the initialized Hessian array of each layer.

[0059] The precision of each layer parameter is reduced from 32 bits to 8 bits using a quantization function, and the quantization error before and after the weight quantization is obtained. Combined with the average value of the trace of the Hessian matrix, the initial quantization sensitivity of each network layer is obtained. In this embodiment, linear symmetric quantization is selected as the quantization function, and the second norm of the weight difference before and after quantization is used as the quantization disturbance.

[0060] In order to reduce the bit width precision, the present invention provides a preferred embodiment to perform S2, using a method of decreasing precision step by step, each time selecting a layer with the lowest quantization sensitivity to reduce its bit width precision;

[0061] like Figure 2 As shown in the figure, it is a schematic diagram of the structure of an efficient target detection method. In the figure, y i represents the i-th layer network, n i represents the number of network parameters corresponding to the i-th layer, b i is the bit width precision of the i-th network layer, s i is the lowest quantization sensitivity of the i-th network layer. As can be seen from the figure, the quantization sensitivity of all network layers obtained according to the initialization is selected each time, and its precision is adjusted from high precision to low precision according to the mapping method of 32 bits → 8 bits, 8 bits → 6 bits, 6 bits → 4 bits, 4 bits → 3 bits, 3 bits → 2 bits, 2 bits → 1 bit. Specifically, if the precision of the selected network layer before the current round of iteration is 6 bits, its precision is adjusted from 6 bits to 4 bits in the current round of iteration. Since its precision has changed, the quantization sensitivity of this layer also needs to be recalculated. When the precision of the network parameters reaches 1 bit, the layer is no longer adjusted.

[0062] When recalculating the quantization sensitivity, the parameters of the network layer change in value because they are adjusted from high precision to low precision, and the average value of the trace of its Hessian matrix and the quantization perturbation also change accordingly and need to be recalculated.

[0063] The present invention provides a preferred embodiment to execute S3, using the L-BFGS algorithm to dynamically update the Hessian matrix and quantization sensitivity of the convolution layer that reduces the bit width accuracy each time to make it more accurate.

[0064] like Figure 2As shown in the figure, after adjusting the parameters of the selected network layer from high precision to low precision, the change in parameters will cause the actual Hessian matrix to change. If the average value of the trace of the initialized Hessian matrix is ​​still used, the quantization sensitivity will be inaccurate. If the Hutchinson method is used to recalculate the Hessian matrix of this layer, the time and computational overhead are too large. Using the L-BFGS algorithm to dynamically update the Hessian matrix of the selected layer can fully reduce the storage and computational complexity.

[0065] The L-BFGS algorithm is the most commonly used method for solving unconstrained nonlinear programming problems. It has the advantages of fast convergence speed and low memory overhead. Its essence is a quasi-Newton method that reduces computational cost and storage space by approximating the Hessian matrix and storing two sequences instead of the entire approximate matrix. This method is used in deep learning to update gradients during network training.

[0066] In fact, the pre-trained full-precision network in this embodiment can be regarded as the optimal solution for network quantization. The network parameters will gradually deviate from the optimal value after each reduction in precision. On the contrary, the small batch fine-tuning after each reduction in precision will make the parameters gradually approach the new optimal value under the current precision constraint. This process conforms to the application scenario of L-BFGS. Therefore, the L-BFGS algorithm can be used to update the Hessian matrix of the selected network layer. Since only the trace of the Hessian matrix of the network layer is considered, only its diagonal elements are stored and updated in this embodiment to further reduce the computational and storage costs.

[0067] Specifically, this embodiment calculates the Hessian matrix of the current round approximately according to the Hessian matrix of the previous round of the network layer and the weight difference and gradient difference of the parameters before and after the current round of training during small batch fine-tuning training, and reduces the storage and calculation overhead by storing the weight difference and gradient difference vector instead of the complete Hessian matrix:

[0068]

[0069] in: is the gradient difference vector of the network layer parameters before and after training, ΔW k =W k+1 -W k is the difference vector of network layer parameters before and after training,

[0070] In this embodiment, in order to reduce the time overhead of fine-tuning training, only 50 images are used for fine-tuning training each time, and 20 images are used to calculate the gradient difference of the layer parameters before and after training. For different network layers, these images are used for calculation. In theory, increasing the number of images can further improve the accuracy of the trace of the Hessian array, thereby making the quantization sensitivity more accurate.

[0071] S4, repeat S2 to S3 to continuously select the network layer with the lowest quantization sensitivity, iteratively adjust the accuracy of the parameters of the network layer until the model compression rate of the entire network reaches the specified threshold, and obtain the target network;

[0072] The present invention provides a preferred embodiment to perform S5, input all training images into the target network that achieves the specified compression rate, and perform fine-tuning training through direct estimation until convergence for target detection.

[0073] This embodiment copies the obtained low-precision network with a specified compression rate, uses one as the teacher network and the other as the student network; inputs all training images into the student network, and the student network passes the gradient obtained during training back to the teacher network. The teacher network updates the network parameters according to the obtained gradients, and quantizes the network parameters according to the accuracy of the corresponding layer, and passes them back to the student network until convergence, and obtains the final student network as the target network. In fine-tuning training, for the obtained efficient RetinaNet target network, 10 -5 is the initial learning rate. After two rounds of training, the learning rate of the network is reduced to 1 / 10 of the original.

[0074] This embodiment automatically determines the bit width accuracy of each layer parameter of the target detection network based on given image data, without complicated manual design, for target detection, greatly reducing bit width allocation time and human resource consumption. While reducing the size of the network model, it can achieve good target detection accuracy.

[0075] In one embodiment of the present invention, for the first-stage target detection network RetinaNet, ResNet-50 is used as the skeleton, and the accuracy of the network parameters is adjusted to an average of 4 bits. This embodiment is evaluated on the Microsoft COCO 2017 dataset, which is currently recognized as the most commonly used and advanced dataset for evaluating target detection networks. The target network achieves a mAP value of 37.1% on COCO2017 while the model size is 1 / 8 of the original.

[0076] The specific comparison results can be seen in the following Table 1, which is a result evaluation comparison of an embodiment of the present invention on a one-stage target detection network RetinaNet, with an average parameter accuracy of 4 bits and a model size of 1 / 8 of the original. In the table, a higher mAP represents a better performance, and a smaller bit width accuracy represents a smaller model size.

[0077] Table 1

[0078]

[0079] The above experiments show that the one-stage efficient target detection method proposed in the embodiment of the present invention can ensure the performance of the target detection method while adaptively reducing the model size under limited computing resources, achieve optimal performance under the same compression rate, and can automatically design and adjust an efficient target detection method based on a neural network in a relatively short time, thereby having a wider range of application scenarios.

[0080] In another embodiment of the present invention, an efficient target detection method includes: obtaining a target detection training set; minimizing the classification loss and regression loss of the training image, and automatically obtaining the parameters of the two-stage target detection network under the specified compression rate by training; using the obtained target detection network to process the image to be detected to obtain the image detection result. In fine-tuning training, for the obtained efficient Faster R-CNN target network, 0.002 is used as the initial learning rate. After two rounds of training, the learning rate of the network is reduced to 1 / 10 of the original. For the two-stage target detection network Faster R-CNN, ResNet-50 is used as the skeleton, and the accuracy of the network parameters is adjusted to an average of 4 bits. This embodiment is evaluated on the Microsoft COCO2017 dataset. The target network achieves a mAP value of 37.1% on COCO2017 while the model size is 1 / 8 of the original. The specific comparison results can be seen in the following Table 2, which is a result evaluation comparison of an embodiment of the present invention on a two-stage target detection network Faster R-CNN, with an average parameter accuracy of 4 bits and a model size of 1 / 8 of the original. In the table, a higher mAP represents a better performance, and a smaller bit width accuracy represents a smaller model size.

[0081] Table 2

[0082]

[0083] The above experiments show that the two-stage efficient target detection method proposed in another embodiment of the present invention can ensure the performance of the target detection method while adaptively reducing the model size under limited computing resources, achieve optimal performance under the same compression rate, and can automatically design and adjust an efficient target detection method based on a neural network in a relatively short time, thereby having a wider range of application scenarios.

[0084] Based on the same concept as the above embodiment, in another embodiment of the present invention, a computer-readable storage medium is further provided, on which a computer program is stored, and when the program is executed by a processor, any of the following methods is implemented: the efficient target detection method of the above embodiment.

[0085] Based on the same concept as the above embodiments, in another embodiment of the present invention, a chip system is also provided, including a processor, a coupling between the processor and a memory, the memory storing program instructions, and when the program instructions stored in the memory are executed by the processor, any one of the following methods is implemented: the efficient target detection method of the above embodiments.

[0086] Optionally, the memory is used to store programs; the memory may include volatile memory (English: volatile memory), such as random-access memory (English: random-access memory, abbreviated: RAM), such as static random-access memory (English: static random-access memory, abbreviated: SRAM), double data rate synchronous dynamic random access memory (English: Double Data Rate Synchronous Dynamic Random Access Memory, abbreviated: DDR SDRAM), etc.; the memory may also include non-volatile memory (English: non-volatile memory), such as flash memory (English: flash memory). The memory 62 is used to store computer programs (such as applications, functional modules, etc. that implement the above method), computer instructions, etc., and the above computer programs, computer instructions, etc. can be partitioned and stored in one or more memories. And the above computer programs, computer instructions, data, etc. can be called by the processor.

[0087] The above-mentioned computer programs, computer instructions, etc. may be stored in one or more memories in partitions, and the above-mentioned computer programs, computer instructions, data, etc. may be called by a processor.

[0088] The processor is used to execute the computer program stored in the memory to implement the various steps in the method involved in the above embodiment. For details, please refer to the relevant description in the above method embodiment.

[0089] The processor and the memory may be independent structures or integrated structures. When the processor and the memory are independent structures, the memory and the processor may be coupled and connected via a bus.

[0090] Computer-readable media includes both computer storage media and communication media, including any media that facilitates transfer of a computer program from one place to another. Storage media may be any available media that can be accessed by a general or special purpose computer.

[0091] In summary, the method and device in the embodiment of the present invention adopts an efficient target detection method, automatically adjusts the accuracy of the target detection network parameters according to the given target detection image data, obtains the target detection network under the specified compression rate, and is used for the target detection task, saving the time and resource consumption of manual allocation, so that the target detection method can be efficiently deployed in industrial applications. Furthermore, the efficient target detection method can achieve performance close to that of a full-precision network while reducing the model size under limited computing resources. This efficient, automated and high-performance target detection method can be deployed on various industrial devices to perform target detection tasks.

[0092] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various modifications or variations within the scope of the claims, which does not affect the essence of the present invention. The above preferred features can be used in any combination without conflicting with each other.

Claims

1. An efficient target detection method, characterized in that: include: Get the target detection training set; Using the training set for pre-training of an object detection network to obtain a full-precision object detection network; Automatically adjusting the accuracy and parameters of the target detection network to obtain a mixed-precision target detection network at a specified compression rate; The mixed precision target detection network is used to process the image to be detected to obtain an image detection result; The training set is used for pre-training of the target detection network to obtain a full-precision target detection network; Automatically adjust the accuracy and parameters of the target detection network to obtain a mixed-precision target detection network at a specified compression rate, including: Compute the quantitative sensitivity of network layers to assess the importance of network parameters; Select the network layer with the lowest quantization sensitivity and use the method of decreasing precision step by step to reduce the precision of the parameters of this layer; Dynamically updating the quantization sensitivity of the network layer with reduced parameter precision using the L-BFGS algorithm; Repeatedly select the network layer with the lowest quantization sensitivity and iteratively adjust the accuracy of the layer parameters until the model compression rate of the entire network reaches the specified threshold, thus obtaining a mixed precision object detection network; Inputting all training images into the mixed precision object detection network, performing fine-tuning training through direct estimation until convergence, and using it for object detection; The quantitative sensitivity can measure the importance of different network layer parameters and evaluate the impact of reducing the accuracy of network layer parameters on network performance, which is: Among them, n i represents the number of network parameters of the i-th layer, is the network parameter matrix of the i-th layer, Q(·) is the quantization function used to reduce the parameter accuracy, is the quantization perturbation, is the loss function, is the loss function About the parameter matrix of the i-th layer The average value of the trace of the Hessian array.

2. The efficient target detection method according to claim 1, characterized in that: The average value of the trace of the Hessian matrix is ​​the sum of the diagonal elements of the Hessian matrix corresponding to each parameter calculated using the Hutchinson method.

3. The efficient target detection method according to claim 1, characterized in that: The method of gradually decreasing the network layer parameter precision refers to selecting the network layer with the lowest quantization sensitivity, and adjusting its parameter precision from high precision to low precision according to the mapping mode of 32 bits→8 bits, 8 bits→6 bits, 6 bits→4 bits, 4 bits→3 bits, 3 bits→2 bits, and 2 bits→1 bit.

4. The efficient target detection method according to claim 1, characterized in that: The L-BFGS algorithm, during training, calculates the Hessian matrix of the current round based on the Hessian matrix of the previous round of the network layer and the weight difference and gradient difference of the parameters before and after the current round of training. in: is the gradient difference vector of the network layer parameters before and after training, ΔW k =W k+1 -W k is the weight difference vector of the network layer parameters before and after training, 5. The efficient target detection method according to claim 1, characterized in that: The fine-tuning training of the direct estimation comprises: Copy the obtained mixed precision network with the specified compression rate, and use one as the teacher network and the other as the student network; All training images are input into the student network. The student network passes the gradients obtained during training back to the teacher network. The teacher network updates the network parameters according to the obtained gradients and quantizes the network parameters according to the accuracy of the corresponding layer and passes them back to the student network until convergence. The final student network is obtained as the mixed precision target detection network.

6. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the efficient target detection method according to any one of claims 1 to 5 is implemented.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the efficient target detection method described in any one of claims 1 to 5 is implemented.

8. A chip system, comprising a processor, the processor coupled to a memory, the memory storing program instructions, characterized in that: When the program instructions stored in the memory are executed by the processor, the efficient target detection method according to any one of claims 1 to 5 is implemented.

Images