Lightweight neural memory ordinary differential equation cascading method based on YOLO V5

Abstract

The invention discloses a lightweight neural memory ordinary differential equation cascading method based on YOLO V5, and the method comprises the steps: outputting feature maps with different resolutions through a Backbone network, carrying out the cascading of NmODE modules, carrying out the repeated fusion of the feature maps between the NmODE modules, and enabling the feature maps with different resolutions to be fully integrated, thereby maintaining the fine local information, and achieving the complete integration of the feature maps with different resolutions. The NmODE module not only receives the output of the previous module as the input, but also fuses the output with the feature map of the Backbone network, so that the feature expression is further enriched in each layer of cascading, and in addition, after each NmODE module, the feature expression is further enriched. And fusing the semantic features output by the NmODE with the current input of the currently cascaded NmODE module again through an enhanced aggregation (EF) module. The dual-path fusion strategy not only enhances the detection capability of a multi-scale target object, but also significantly improves the detection precision through the interaction and fusion of features.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence neural networks, and in particular to a lightweight neural memory ordinary differential equation cascade method based on YOLO V5. Background Art

[0002] Object detection tasks based on deep learning are widely used in many fields such as intelligent security, autonomous driving, robot navigation, and medical diagnosis. Detection networks based on convolutional neural networks (CNNs) have gradually developed the object detection framework from two stages (e.g., Faster RCNN, Mask RCNN, and Cascade RCNN) to a single stage (e.g., YOLO). Among them, YOLO stands out for its small size, fast computing speed, and strong generalization ability. It transforms the object detection problem into a regression problem, thereby achieving fast and efficient object detection, but the detection accuracy still needs to be improved. Subsequently, a large number of YOLO variants have emerged, and these studies have sought a balance between the detection accuracy and efficiency of the balanced model. These variants can be divided into two categories: anchor-based (e.g., YOLOv3, YOLOv4, and YOLOv5) and anchor-free (e.g., YOLOv8 and YOLOX). These variants almost all contain the same parts: backbone, neck, and head.

[0003] The backbone network architecture has been widely studied (e.g., CSPDarknet, Swin Transformer, or ConvNeXt). Replacing the backbone network is a common method to improve the YOLO series models. The neck in the YOLO series usually uses the feature pyramid network (FPN) and its variants to fuse multi-level features from different backbones. The learned features are then fed into an intuitive head architecture to predict position information. However, although FPN and its variants have alleviated the information fusion problem to a certain extent, previous models still have information fusion problems, which to a certain extent hinders the detection accuracy of the YOLO series models. Therefore, a lot of research has focused on improving the neck module. Although these methods effectively improve multi-scale information fusion, they still increase the computational complexity of the original structure and ignore the expressive power of features in datasets in different fields.

[0004] Therefore, the present invention proposes a lightweight neural memory ordinary differential equation cascade method based on YOLO V5. Summary of the invention

[0005] In order to overcome the shortcomings and deficiencies of the prior art, the present invention provides a lightweight neural memory ordinary differential equation cascade method based on YOLO V5.

[0006] A lightweight neural memory ordinary differential equation cascade method based on YOLO V5, the method includes:

[0007] Step S1: After data enhancement, the input image is sent to the Backbone network. The Backbone consists of the Focus structure and CSPDarknet53. The network scales of the Backbone of the YOLOV5 model of different scales are different, but three feature maps of different resolutions of the same scale are output, which come from the shallow, middle and deep networks, respectively, named P3, P4 and P5;

[0008] Step S2: The NmODE module is based on the neural memory ordinary differential equation. It aggregates the deep and shallow features by cascading four NmODE modules. The output of a single NmODE module is used as the h(0) value input of the next NmODE module. The next NmODE module aggregates the features output by the backbone. Each module receives the output of the previous module and fuses it with the deep feature map.

[0009] Step S3: The semantic features of the NmODE output and the current input of NmODE are fused through the enhanced aggregation EF module. The enhanced aggregation EF module adopts a dual-path structure, namely a main path and an auxiliary path. The first main path strengthens the fusion of the semantic features of the NmODE output and the current input of NmODE. The second auxiliary path fuses the current input of NmODE again. The main path and the auxiliary path convert the input into a feature map output containing three dimensions: target category, position (bounding box) and confidence. The prediction result of the auxiliary path is added to the prediction result of the main path to achieve feature fusion and prediction correction.

[0010] Step S4: The three enhanced aggregation EF modules respectively output feature maps containing three dimensions of target category, position (bounding box) and confidence. The three feature maps from left to right correspond to the category, position (bounding box) and confidence prediction of medium, small and large targets, respectively, and perform OSS calculation to update the network parameters.

[0011] Furthermore, in step S2, the neural memory ordinary differential equation is as follows:

[0012]

[0013] in, is the derivative of the output h(k), representing the hidden state of the network, h(k) is the current state of the network, S is the activation function, f is the convolution operation, o(k) is the current input of the network, θ K are the current parameters of the network.

[0014] Furthermore, in step S2, the first NmODE module expression is as follows:

[0015] y1=NmODE(P4,Resize ↓ (P3))(2),

[0016] Among them, P4 is used as the current input. ↓ (·) refers to downsampling achieved through a 3*3 convolution operation with a stride of 2. After this downsampling, the resolution of P3 is reduced from 80*80 to 40*40 and is used as the input of h(0).

[0017] Furthermore, in step S2, the second NmODE module expression is as follows:

[0018] y2=NmODE(Resize ↑ (P5),y1)(3),

[0019] Among them, Resize ↑ (·) indicates the nearest neighbor interpolation layer, that is, P5 is upsampled and the resolution is enlarged from 20*20 to 40*40 as the current input. y1 is used as h(0) input. The output y2 of this module combines the features from three levels (P3, P4 and P5), and then passed to the next NmODE module and the first enhanced aggregation EF module for further enhancement and fusion.

[0020] Furthermore, in step S2, the third NmODE module expression is as follows:

[0021] y3=NmODE(Resize ↑ (P4),Resize ↑ (y2))(4),

[0022] Among them, Resize ↑ (·) indicates the nearest neighbor interpolation layer, that is, P4 and y2 are upsampled, the resolution of P4 is enlarged from 40*40 to 80*80, as the current input, and the resolution of y2 is enlarged from 40*40 to 80*80, as the input of h(0). The output is then passed to the fourth NmODE module and the second enhanced aggregation EF module for further enhancement and fusion.

[0023] Furthermore, in step S2, the fourth NmODE module expression is as follows:

[0024] y4=NmODE(P5,Resize ↓ (y3))(5),

[0025] Among them, P5 is the current input. ↓(·) refers to downsampling through a 3*3 convolution operation with a stride of 4. After this downsampling, the resolution of y3 is reduced from 80*80 to 20*20, which is used as the input of h(0). The output of this module is then passed to the last enhanced aggregation EF module for further enhancement and feature fusion.

[0026] Furthermore, in step S3, the enhanced aggregation EF module fusion method is as follows:

[0027] The enhanced aggregation EF module adopts a dual-path structure, namely the main path and the auxiliary path; the main path strengthens the semantic features of the fusion NmODE output and the current input of NmODE; as shown in formula (6), the two features are extracted through a standard 3*3 convolution module to obtain a rough feature map containing three dimensions: target category, position (bounding box) and confidence; the two feature maps are concatenated (concat) and the fused feature map is reduced in dimension using a 1*1 convolution module to obtain a feature map F containing higher-level semantic information. P ;

[0028] F P =Conv 1×1 {Concat[Conv 3×3 (S),Conv 3×3 (D)]}(6),

[0029] The second auxiliary path, as shown in formula (7), uses a 3*3 convolutional module to transform the deep features of the backbone into a feature map F containing three dimensions: target category, location (bounding box) and confidence a ;

[0030] F a =Conv 3×3 (D)(7),

[0031] The prediction results of the main path are added to the prediction results of the auxiliary path, as shown in formula (8) to achieve feature fusion and prediction correction;

[0032] F final =F P +F a (8).

[0033] Beneficial effects:

[0034] The present invention proposes a lightweight neural memory ordinary differential equation cascade method based on YOLO V5, and realizes the improvement of layer-by-layer feature fusion and detection accuracy by cascading multiple NmODE modules in sequence. In this method, the Backbone network outputs feature maps of different resolutions, which represent semantic information at different levels, covering from shallow detail features to deep abstract features. Through the cascade of NmODE modules, feature maps are repeatedly fused between each NmODE module, and feature maps of different resolutions are fully integrated, which not only retains fine local information, but also enhances the expression ability of global context. The structural design of the NmODE module is simple and contains only one convolution layer, which makes the cascade structure significantly reduce the computational complexity of the network while maintaining efficient detection capabilities. Compared with traditional multi-layer convolutional networks, this lightweight design is suitable for embedded devices or other scenarios with limited computing resources, while still ensuring efficient detection of target objects. During the cascade process, the NmODE module not only receives the output of the previous module as input, but also fuses with the feature map of the Backbone network, thereby further enriching the feature expression in each layer of cascade. In addition, after each NmODE module, the semantic features output by NmODE are fused again with the current input of the current cascaded NmODE module through the enhanced aggregation EF module. This dual-path fusion strategy not only enhances the detection capability of multi-scale target objects, but also significantly improves the detection accuracy through the interaction and fusion of features. The dual-path structure of the enhanced aggregation EF module effectively combines semantic features, spatial information, and confidence information, ensuring excellent detection performance on datasets in different fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a general step flow chart of the present invention. DETAILED DESCRIPTION

[0036] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application may be combined with each other. The present application is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0037] like Figure 1 As shown, a lightweight neural memory ordinary differential equation cascade method based on YOLO V5 includes:

[0038] Step S1: After data enhancement, the input image is sent to the Backbone network. The Backbone consists of the Focus structure and CSPDarknet53. The network scales of the Backbone of the YOLOV5 model of different scales are different, but three feature maps of different resolutions of the same scale are output, which come from the shallow, middle and deep networks, respectively, named P3, P4 and P5;

[0039] Specifically, the input image can be a natural scene image, a medical image, a traffic monitoring image, etc. Therefore, the input image can contain multiple target objects. The size of the input image is adjusted to 640*640 pixels. This is to enable the model to process images of different sizes and maintain a consistent input size in the feature extraction stage. The Backbone network of YOLO V5 consists of the Focus structure and CSPDarknet53. The Focus structure is a pre-module introduced by YOLO V5, which is specifically used to preprocess the input image. It converts the spatial information of the image into richer channel information by slicing the input image. This operation not only reduces the size of the image, but also reduces the amount of calculation. In the specific calculation process, the Focus structure slices each 2*2 area of ​​the input image into four channels (equivalent to compressing the spatial information into the channel dimension), thereby improving the calculation efficiency. For example, an input image of 640*640*3 becomes 320*320*12 after passing through the Focus module (each 2*2 pixel block is converted into a feature channel, with a total of 12 channels). CSPDarknet53 is the core backbone structure of YOLO V5. It is improved based on Darknet53 and combines the design of Cross Stage Partial Network (CSPNet). CSPNet achieves more efficient feature fusion and reduces redundant calculations by segmenting feature maps. CSPDarknet53 consists of multiple residual blocks and extracts features at different levels through layer-by-layer deep convolution. In each residual block, the CSP structure divides the feature map into two parts, one part is directly connected through the residual, and the other part is merged with the former after convolution. This structure effectively alleviates the gradient disappearance problem and enhances the feature representation capability. The backbone network of YOLO V5 extracts three feature maps of different resolutions from the input image, representing the semantic information of the shallow, middle and deep layers respectively. These feature maps are named P3 (shallow), P4 (middle) and P5 (deep). The resolution of P3 is 80*80 and the number of channels is 256. It is mainly responsible for extracting local detail information in the picture, such as edges and textures, which are very important for identifying small objects and objects with rich details. The resolution of P4 is 40*40 and the number of channels is 512. It is responsible for further extracting higher-level semantic information while maintaining certain spatial details. It combines the detail information of the shallow network and the abstract semantic information of the deep network. The resolution of P5 is 20*20 and the number of channels is 1024. It mainly extracts global semantic information in the picture, such as the shape and category of the object. Since the deep network has a larger receptive field, it can capture the global features of the entire object or even the scene, and is suitable for detecting large objects and more abstract targets.

[0040] Step S2: The NmODE module is based on the neural memory ordinary differential equation. It aggregates the deep and shallow features by cascading four NmODE modules. The output of a single NmODE module is used as the h(0) value input of the next NmODE module. The next NmODE module aggregates the features output by the backbone. Each module receives the output of the previous module and fuses it with the deep feature map.

[0041] Specifically, the first NmODE module is responsible for fusing shallow and middle-level features. The shallow feature P3 is used as the initial hidden state h(0). It is first down-sampled by a 3*3 convolution operation with a stride of 2. After the downsampling, the resolution of P3 is reduced from 80*80 to 40*40. After the 3*3 convolution operation, BatchNorm normalization and Si lu activation function processing, it is combined with the middle-level feature P4 through the solution process of the neural memory ordinary differential equation to generate the fused feature y1. y1 is passed to the second NmODE module as its initial state; the second NmODE module is responsible for fusing y1 and the deep feature P5. P5 is up-sampled by the nearest neighbor interpolation method, and the resolution is expanded from 20*20 to 40*40. After the 3*3 convolution operation, BatchNorm normalization and Si lu activation function processing, it is combined with the middle-level feature P4 through the solution process of the neural memory ordinary differential equation to generate the fused feature y1. y1 is passed to the second NmODE module as its initial state; the second NmODE module is responsible for fusing y1 and the deep feature P5. P5 is up-sampled by the nearest neighbor interpolation method, and the resolution is expanded from 20*20 to 40*40. lu activation function, and y1 are processed through the process of solving the neural memory ordinary differential equation to generate the fused feature y2, which is passed to the third NmODE module as its initial state; the third NmODE module is responsible for fusing y2 and the middle-level feature P4. P4 and y2 are upsampled by the nearest neighbor interpolation method, and the resolution is expanded from 40*40 to 80*80, and then processed by 3*3 convolution operation, BatchNorm normalization and Si lu activation function, and the fused feature y3 is generated through the process of solving the neural memory ordinary differential equation. y3 is passed to the fourth NmODE module as its initial state; the fourth NmODE module is responsible for fusing y3 and the deep feature P5. y3 is first downsampled by a 3×3 convolution operation with a stride of 4. After the downsampling, the resolution of y3 is reduced from 80*80 to 20*20, and then the 3*3 convolution operation, BatchNorm normalization and Si After being processed by the lu activation function, it is combined with y3 through the solution process of the neural memory ordinary differential equation to generate the fusion feature y4.

[0042] Step S3: The semantic features of the NmODE output and the current input of NmODE are fused through the enhanced aggregation EF module. The enhanced aggregation EF module adopts a dual-path structure, namely a main path and an auxiliary path. The first main path strengthens the fusion of the semantic features of the NmODE output and the current input of NmODE. The second auxiliary path fuses the current input of NmODE again. The main path and the auxiliary path convert the input into a feature map output containing three dimensions: target category, position (bounding box) and confidence. The prediction result of the auxiliary path is added to the prediction result of the main path to achieve feature fusion and prediction correction.

[0043] Specifically, the number of channels output by the three enhanced aggregation EF modules is consistent. For a dataset with i categories, the output channel dimension is 3*(i+4+1), where 3 is the number of anchors, 4 is the relative coordinates of the bounding box, and 1 is the confidence score of the box. The main path of the first enhanced aggregation EF module fuses the output y2 of the second NmODE and its current input P5, and the auxiliary path fuses the current input of the second NmODE again. The main path and the auxiliary path transform y2 and P5 into feature map outputs containing three dimensions: target category, position (bounding box), and confidence. The prediction results of the auxiliary path are added to the prediction results of the main path to achieve feature fusion and prediction correction. The output resolution of this module is 40*40; the main path of the second enhanced aggregation EF module fuses the output y3 of the third NmODE and its current input P4, and the auxiliary path fuses the current input of the third NmODE again. The main path and the auxiliary path transform y3 and P4 into The feature map output contains three dimensions: target category, position (bounding box) and confidence. The prediction result of the auxiliary path is added to the prediction result of the main path to achieve feature fusion and prediction correction. The output resolution of this module is 80*80; the main path of the third enhanced aggregation EF module fuses the output y4 of the fourth NmODE and its current input P5, and the auxiliary path fuses the current input of the fourth NmODE again. The main path and the auxiliary path convert y4 and P5 into feature map outputs containing three dimensions: target category, position (bounding box) and confidence. The prediction result of the auxiliary path is added to the prediction result of the main path to achieve feature fusion and prediction correction. The output resolution of this module is 20*20.

[0044] Step S4: The three enhanced aggregation EF modules respectively output feature maps containing three dimensions of target category, position (bounding box) and confidence. The three feature maps from left to right correspond to the category, position (bounding box) and confidence prediction of medium, small and large targets, respectively, and perform OSS calculation to update the network parameters.

[0045] Specifically, after the feature map is generated, the network compares the predicted category, position and confidence with the true label and calculates the corresponding loss function, including classification loss, bounding box regression loss and confidence loss. Through back propagation, the loss is used to optimize and update the network parameters, thereby improving the model's detection performance for objects of different scales and ensuring the accuracy of the network in category distinction, position prediction and confidence evaluation.

[0046] Furthermore, in step S2, the neural memory ordinary differential equation expression is:

[0047]

[0048] in, is the derivative of the output h(k), representing the hidden state of the network, h(k) is the current state of the network, S is the activation function, f is the convolution operation, o(k) is the current input of the network, θ K are the current parameters of the network.

[0049] Furthermore, in step S2, the first NmODE module expression is:

[0050] y1=NmODE(P4,Resize ↓ (P3))(2),

[0051] Among them, P4 is the current input, Resize ↓ (·) refers to downsampling achieved through a 3*3 convolution operation with a stride of 2. After this downsampling, the resolution of P3 is reduced from 80*80 to 40*40 and is used as the input of h(0).

[0052] Furthermore, in step S2, the second NmODE module expression is:

[0053] y2=NmODE(Resize ↑ (P5),y1)(3),

[0054] Among them, Resize ↑ (·) indicates the nearest neighbor interpolation layer, that is, P5 is upsampled and the resolution is enlarged from 20*20 to 40*40 as the current input. y1 is used as h(0) input. The output y2 of this module combines the features from three levels (P3, P4 and P5), and then passed to the next NmODE module and the first enhanced aggregation EF module for further enhancement and fusion.

[0055] Furthermore, in step S2, the third NmODE module expression is:

[0056] y3=NmODE(Resize ↑(P4),Resize ↑ (y2))(4),

[0057] Among them, Resize ↑ (·) indicates the nearest neighbor interpolation layer, that is, P4 and y2 are upsampled, and the resolution of P4 is enlarged from 40*40 to 80*80 as the current input. The resolution of y2 is enlarged from 40*40 to 80*80 as the input of h(0), and the output is then passed to the fourth NmODE module and the second enhanced aggregation EF module for further enhancement and fusion.

[0058] Furthermore, in step S2, the fourth NmODE module expression is:

[0059] y4=NmODE(P5,Resize ↓ (y3))(5),

[0060] Among them, P5 is the current input. ↓ (·) refers to downsampling through a 3*3 convolution operation with a stride of 4. After this downsampling, the resolution of y3 is reduced from 80*80 to 20*20, which is used as the input of h(0). The output of this module is then passed to the last enhanced aggregation EF module for further enhancement and feature fusion.

[0061] Furthermore, in step S3, the enhanced aggregation EF module fusion method is as follows:

[0062] The enhanced aggregation EF module adopts a dual-path structure, namely the main path and the auxiliary path; the main path strengthens the semantic features of the fusion NmODE output and the current input of NmODE; as shown in formula (6), the two features are extracted through a standard 3*3 convolution module to obtain a rough feature map containing three dimensions: target category, position (bounding box) and confidence; the two feature maps are concatenated (concat) and the fused feature map is reduced in dimension using a 1*1 convolution module to obtain a feature map F containing higher-level semantic information. P ;

[0063] F P =Conv 1×1 {Concat[Conv 3×3 (S),Conv 3×3 (D)]}(6),

[0064] The second auxiliary path, as shown in formula (7), uses a 3*3 convolutional module to transform the deep features of the backbone into a feature map F containing three dimensions: target category, location (bounding box) and confidence a ;

[0065] F a =Conv 3×3 (D)(7),

[0066] The prediction results of the main path are added to the prediction results of the auxiliary path, as shown in formula (8) to achieve feature fusion and prediction correction;

[0067] F final =F P +F a (8).

[0068] Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A lightweight neural memory ordinary differential equation cascade method based on YOLO V5, characterized by , the method comprises: Step S1: After data enhancement, the input image is sent to the Backbone network. The Backbone consists of the Focus structure and CSPDarknet53. The network scales of the Backbone of the YOLOV5 model of different scales are different, but three feature maps of different resolutions of the same scale are output, which come from the shallow, middle and deep networks, respectively, named P3, P4 and P5; Step S2: The NmODE module is based on the neural memory ordinary differential equation. It aggregates the deep and shallow features by cascading four NmODE modules, and uses the output of the current NmODE module as the h(0) value input of the next NmODE module. The next NmODE module aggregates the features output by the backbone, where each module receives the output of the previous module and fuses it with the deep feature map; Step S3: The semantic features of the NmODE output and the current input of NmODE are fused through the enhanced aggregation EF module. The enhanced aggregation EF module adopts a dual-path structure, namely a main path and an auxiliary path. The first main path strengthens the fusion of the semantic features of the NmODE output and the current input of NmODE. The second auxiliary path fuses the current input of NmODE again. The main path and the auxiliary path convert the input into a feature map output containing three dimensions of target category, position and confidence. The prediction result of the auxiliary path is added to the prediction result of the main path to achieve feature fusion and prediction correction. Step S4: The three enhanced aggregation EF modules respectively output feature maps containing three dimensions of target category, position and confidence. The three feature maps from left to right correspond to the category, position and confidence prediction of medium, small and large targets, respectively, and loss calculation is performed to update the network parameters.

2. The lightweight neural memory ordinary differential equation cascade method based on YOLO V5 as claimed in claim 1, characterized in that: In step S2, the neural memory ordinary differential equation expression is: in, is the derivative of the output h(k), representing the hidden state of the network, h(k) is the current state of the network, S is the activation function, f is the convolution operation, o(k) is the current input of the network, θ K are the current parameters of the network.

3. The lightweight neural memory ordinary differential equation cascade method based on YOLO V5 as claimed in claim 1, characterized in that: In step S2, the first NmODE module expression is: y1=NmODE(P4,Resize ↓ (P3))(2), Among them, P4 is the current input, Resize ↓ (·) refers to downsampling achieved through a 3*3 convolution operation with a stride of 2. After this downsampling, the resolution of P3 is reduced from 80*80 to 40*40 and is used as the input of h(0).

4. The lightweight neural memory ordinary differential equation cascade method based on YOLO V5 as claimed in claim 1, characterized in that: In step S2, the second NmODE module expression is: y2=NmODE(Resize ↑ (P5),y1)(3), Among them, Resize ↑ (·) indicates the nearest neighbor interpolation layer, that is, P5 is upsampled and the resolution is enlarged from 20*20 to 40*40 as the current input, y1 is used as h(0) input, and the output y2 of this module combines the features from three levels (P3, P4 and P5), and then is passed to the next NmODE module and the first enhanced aggregation EF module for further enhancement and fusion.

5. The lightweight neural memory ordinary differential equation cascade method based on YOLO V5 as claimed in claim 1, characterized in that: In step S2, the third NmODE module expression is: y3=NmODE(Resize ↑ (P4),Resize ↑ (y c ))(4), Among them, Resize ↑ (·) represents the nearest neighbor interpolation layer, i.e., P4 and y - The P4 resolution is upsampled from 40*40 to 80*80 as the current input; - The resolution is enlarged from 40*40 to 80*80 as the input of h(0); the output is then passed to the fourth NmODE module and the second enhanced aggregation EF module for further enhancement and fusion.

6. The lightweight neural memory ordinary differential equation cascade method based on YOLO V5 as claimed in claim 1, characterized in that: In step S2, the fourth NmODE module expression is: y4=NmODE(P5,Resize ↓ (y3))(5), Among them, P5 is the current input, Resize ↓ (·) refers to downsampling achieved through a 3*3 convolution operation with a stride of 4. After this downsampling, the resolution of y3 is reduced from 80*80 to 20*20 and used as the input of h(0). The output of this module is then passed to the last enhanced aggregation EF module for further enhancement and feature fusion.

7. The lightweight neural memory ordinary differential equation cascade method based on YOLO V5 as claimed in claim 1, characterized in that: In step S3, the enhanced aggregation EF module fusion method is as follows: The enhanced aggregation EF module adopts a dual-path structure, namely the main path and the auxiliary path; the main path strengthens the semantic features of the fusion NmODE output and the current input of NmODE; Formula (6) specifically extracts the two features through a standard 3*3 convolution module to obtain a rough feature map containing three dimensions of target category, position and confidence; the two feature maps are spliced, and the fused feature map is reduced in dimension using a 1*1 convolution module to obtain a feature map F containing higher-level semantic information. P ; F P =Conv 1×1 {Concat[Conv 3×3 (S),Conv 3×3 (D)]}(6), The second auxiliary path, formula (7), uses a 3*3 convolutional module to transform the deep features of the backbone into a feature map F containing three dimensions: target category, location, and confidence. a ; F a =Conv 3×3 (D)(7), The prediction results of the main path are added to the prediction results of the auxiliary path, and formula (8) realizes feature fusion and prediction correction; F final =F P +F a (8)。

Images