A multi-scale linear feature generation system and method

Abstract

The application discloses a multi-scale linear feature generation system and method, wherein a dimension reduction processing module is used for point-by-point convolution dimension reduction processing on an input input feature map, and a first internal feature map obtained is transmitted to a coordinate attention enhancement module; the coordinate attention enhancement module is used for sequentially sampling, feature extraction and mapping processing on the first internal feature map in the X-axis direction and the Y-axis direction respectively, then product processing is performed on the first internal feature map and first and second weight factors obtained, and a second internal feature map obtained is transmitted to a multi-scale feature generation module; the multi-scale feature generation module is used for sequentially performing grouping dilated convolution and channel splicing processing on the second internal feature map to obtain an output feature map. After the multi-scale feature generation module performs grouping dilated convolution on the second internal feature map by using different expansion rates, channel splicing processing is performed to obtain the output feature map, so that rich and diversified linear features are generated.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a multi-scale linear feature generation system and method. Background Technology

[0002] Existing technologies and their disadvantages:

[0003] The core of GhostNet (edge-side neural network), a lightweight convolutional neural network built for mobile or embedded platforms, is the Ghost module. The Ghost module first uses ordinary convolutions (e.g., 1×1 point-wise convolution) to generate intrinsic feature maps, and then uses 3×3 depthwise convolutions as inexpensive linear operations (e.g., 3×3 depthwise convolution) to generate more redundant linear features. Therefore, half of the features in the feature map generated by the Ghost module capture spatial information through inexpensive linear operations, while the other half are generated only through 1×1 pointwise convolutions and do not interact with other pixels. This results in limited spatial information capture, weakening feature representation capabilities and potentially hindering further performance improvements.

[0004] As can be seen, the ability of the Ghost module to generate linear features usually depends on cheap linear operations and varies with the number of input feature channels, resulting in extremely unstable performance of GhostNet in visual tasks (image classification, image segmentation, and object detection, etc.).

[0005] Therefore, the existing technology still needs to be improved and enhanced. Summary of the Invention

[0006] The main objective of this invention is to provide a multi-scale linear feature generation system and method, which aims to solve the problem of limited performance and instability of GhostNet in the prior art when processing visual tasks.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A multi-scale linear feature generation system includes: a dimensionality reduction module, a coordinate attention enhancement module, and a multi-scale feature generation module; the dimensionality reduction module, the coordinate attention enhancement module, and the multi-scale feature generation module are connected sequentially.

[0009] The dimensionality reduction processing module performs pointwise convolutional dimensionality reduction processing on the input feature map to obtain a first intrinsic feature map, and transmits the first intrinsic feature map to the coordinate attention enhancement module. The coordinate attention enhancement module performs sampling, feature extraction, and mapping processing on the first intrinsic feature map in the X-axis and Y-axis directions respectively, and then multiplies the obtained first weight factor and second weight factor with the first intrinsic feature map to obtain a second intrinsic feature map, and transmits the second intrinsic feature map to the multi-scale feature generation module. The multi-scale feature generation module performs grouped dilated convolution and channel concatenation processing on the second intrinsic feature map in sequence to obtain an output feature map.

[0010] In the multi-scale linear feature generation system, the coordinate attention enhancement module includes: a first pooling unit, a second pooling unit, a first feature extraction unit, a second feature extraction unit, a first mapping unit, a second mapping unit, and a pointwise multiplication unit;

[0011] The dimensionality reduction processing module, the first pooling unit, the first feature extraction unit, the first mapping unit, and the point-by-point multiplication unit are connected in sequence; the dimensionality reduction processing module, the second pooling unit, the second feature extraction unit, the second mapping unit, and the point-by-point multiplication unit are also connected in sequence.

[0012] The first pooling unit is used to perform a first average pooling process on the first intrinsic feature map to obtain a first one-dimensional vector, and transmit the first one-dimensional vector to the first feature extraction unit; the first feature extraction unit is used to perform a one-dimensional convolution process on the first one-dimensional vector to obtain a second one-dimensional vector, and transmit the second one-dimensional vector to the first mapping unit; the first mapping unit is used to perform a mapping process on the second one-dimensional vector using an activation function to obtain a third one-dimensional vector, and transmit the third one-dimensional vector to the pointwise multiplication unit.

[0013] The second pooling unit is used to perform a second average pooling process on the first intrinsic feature map to obtain a fourth one-dimensional vector, and transmits the fourth one-dimensional vector to the second feature extraction unit; the second feature extraction unit is used to perform a one-dimensional convolution process on the fourth one-dimensional vector to obtain a fifth one-dimensional vector, and transmits the fifth one-dimensional vector to the second mapping unit; the second mapping unit is used to perform a mapping process on the fifth one-dimensional vector using the activation function to obtain a sixth one-dimensional vector, and transmits the sixth one-dimensional vector to the pointwise multiplication unit;

[0014] The point-by-point multiplication unit is used to perform point-by-point multiplication on the third one-dimensional vector, the sixth one-dimensional vector, and the first intrinsic feature map to obtain the second intrinsic feature map, and then transmit the second intrinsic feature map to the multi-scale feature generation module.

[0015] In the multi-scale linear feature generation system, the multi-scale feature generation module includes: a grouped dilated convolution unit and a channel splicing unit; the pointwise multiplication unit, the grouped dilated convolution unit, and the channel splicing unit are connected in sequence.

[0016] The grouped dilated convolution unit is used to perform grouped dilated convolution processing on the second intrinsic feature map to obtain a first multi-scale linear feature map, a second multi-scale linear feature map, and a third multi-scale linear feature map, which are then transmitted to the channel splicing unit. The channel splicing unit is used to splice the second intrinsic feature map, the first multi-scale linear feature map, the second multi-scale linear feature map, and the third multi-scale linear feature map to obtain the output feature map.

[0017] In the multi-scale linear feature generation system, the first pooling unit includes an X-axis average pooling layer; the second pooling unit includes a Y-axis average pooling layer; the X-axis average pooling layer is connected to the dimensionality reduction processing module and the first feature extraction unit respectively; the Y-axis average pooling layer is connected to the dimensionality reduction processing module and the second feature extraction unit respectively.

[0018] The X-axis average pooling layer is used to perform X-axis average pooling processing on the first intrinsic feature map to obtain the first one-dimensional vector, and transmit the first one-dimensional vector to the first feature extraction unit; the Y-axis average pooling layer is used to perform Y-axis average pooling processing on the first intrinsic feature map to obtain the fourth one-dimensional vector, and transmit the fourth one-dimensional vector to the second feature extraction unit.

[0019] In the multi-scale linear feature generation system, the first weighting factor includes the third one-dimensional vector; the second weighting factor includes the sixth one-dimensional vector; and the activation function includes the Sigmoid function.

[0020] In the multi-scale linear feature generation system, the grouped dilated convolution unit includes: a first grouped dilated convolution layer, a second grouped dilated convolution layer, and a third grouped dilated convolution layer; the first grouped dilated convolution layer is connected to the pointwise multiplication unit and the channel concatenation unit, respectively; the second grouped dilated convolution layer is connected to the pointwise multiplication unit and the channel concatenation unit, respectively; and the third grouped dilated convolution layer is connected to the pointwise multiplication unit and the channel concatenation unit, respectively.

[0021] The first grouped dilated convolutional layer is used to perform a first grouped dilated convolutional process on the second intrinsic feature map to obtain the first multi-scale linear feature map, which is then transmitted to the channel stitching unit; the second grouped dilated convolutional layer is used to perform a second grouped dilated convolutional process on the second intrinsic feature map to obtain the second multi-scale linear feature map, which is then transmitted to the channel stitching unit; the third grouped dilated convolutional layer is used to perform a third grouped dilated convolutional process on the second intrinsic feature map to obtain the third multi-scale linear feature map, which is then transmitted to the channel stitching unit.

[0022] A multi-scale linear feature generation method based on the multi-scale linear feature generation system described above, the multi-scale linear feature generation method comprising the following steps:

[0023] The dimensionality reduction processing module performs pointwise convolution dimensionality reduction processing on the input feature map and transmits the resulting first intrinsic feature map to the coordinate attention enhancement module.

[0024] The coordinate attention enhancement module samples, extracts features, and maps the first intrinsic feature map sequentially in the X-axis and Y-axis directions, respectively. Then, it multiplies the obtained first weight factor and second weight factor with the first intrinsic feature map and transmits the obtained second intrinsic feature map to the multi-scale feature generation module.

[0025] The multi-scale feature generation module sequentially performs grouped dilated convolution and channel concatenation on the second intrinsic feature map to obtain the output feature map.

[0026] In the multi-scale linear feature generation method, the dimensionality reduction module performs pointwise convolutional dimensionality reduction on the input feature map and transmits the resulting first intrinsic feature map to the coordinate attention enhancement module, specifically including:

[0027] After receiving the input feature map, the dimensionality reduction processing module performs point-by-point convolution dimensionality reduction processing on the input feature map to obtain the first intrinsic feature map;

[0028] The dimensionality reduction processing module transmits the first intrinsic feature map to the first pooling unit, the second pooling unit, and the pointwise multiplication unit.

[0029] In the multi-scale linear feature generation method, the coordinate attention enhancement module sequentially samples, extracts features, and maps the first intrinsic feature map along the X-axis and Y-axis directions, respectively. The resulting first and second weighting factors are then multiplied with the first intrinsic feature map, and the resulting second intrinsic feature map is transmitted to the multi-scale feature generation module. Specifically, this includes:

[0030] After performing X-axis average pooling on the first intrinsic feature map, the first pooling unit transmits the resulting first one-dimensional vector to the first feature extraction unit; wherein, the first pooling unit includes: an X-axis average pooling layer;

[0031] After performing a one-dimensional convolution on the first one-dimensional vector, the first feature extraction unit transmits the resulting second one-dimensional vector to the first mapping unit.

[0032] After the first mapping unit performs mapping processing on the second one-dimensional vector using an activation function, the resulting third one-dimensional vector is transmitted to the pointwise multiplication unit; wherein, the first weight factor includes: the third one-dimensional vector;

[0033] After performing Y-axis average pooling on the first intrinsic feature map, the second pooling unit transmits the resulting fourth one-dimensional vector to the second feature extraction unit; wherein, the second pooling unit includes: a Y-axis average pooling layer;

[0034] After performing a one-dimensional convolution on the fourth one-dimensional vector, the second feature extraction unit transmits the resulting fifth one-dimensional vector to the second mapping unit.

[0035] After the second mapping unit performs mapping processing on the fifth one-dimensional vector using the activation function, the resulting sixth one-dimensional vector is transmitted to the pointwise multiplication unit; wherein, the second weight factor includes: the sixth one-dimensional vector; the activation function includes: the Sigmoid function;

[0036] The point-by-point multiplication unit expands the third one-dimensional vector and the sixth one-dimensional vector to obtain the attention weight matrix, then performs point-by-point multiplication on the attention weight matrix and the first intrinsic feature map, and transmits the resulting second intrinsic feature map to the grouped dilated convolution unit and the channel splicing unit.

[0037] In the multi-scale linear feature generation method, the multi-scale feature generation module sequentially performs grouped dilated convolution and channel concatenation on the second intrinsic feature map to obtain an output feature map, specifically including:

[0038] The grouped dilated convolution unit performs a first grouped dilated convolution on the second intrinsic feature map using a first dilation rate, and then transmits the resulting first multi-scale linear feature map to the channel stitching unit; wherein, the grouped dilated convolution unit includes: a first grouped dilated convolution layer, a second grouped dilated convolution layer and a third grouped dilated convolution layer;

[0039] After the grouped dilatation convolution unit performs the second grouped dilatation convolution on the second intrinsic feature map using the second dilation rate, the resulting second multi-scale linear feature map is transmitted to the channel stitching unit.

[0040] The grouped dilatation convolution unit performs a third grouped dilatation convolution on the second intrinsic feature map using a third dilation rate, and then transmits the resulting third multi-scale linear feature map to the channel stitching unit.

[0041] The channel stitching unit stitches the second intrinsic feature map, the first multi-scale linear feature map, the second multi-scale linear feature map, and the third multi-scale linear feature map to obtain the output feature map.

[0042] Compared to existing technologies, this invention provides a multi-scale linear feature generation system and method. In this system, a dimensionality reduction module performs pointwise convolutional dimensionality reduction on the input feature map and transmits the resulting first intrinsic feature map to a coordinate attention enhancement module. The coordinate attention enhancement module sequentially samples, extracts features, and maps the first intrinsic feature map along the X-axis and Y-axis, respectively. It then multiplies the resulting first and second weighting factors with the first intrinsic feature map and transmits the resulting second intrinsic feature map to a multi-scale feature generation module. The multi-scale feature generation module sequentially performs grouped dilated convolution and channel concatenation on the second intrinsic feature map to obtain an output feature map. By using different dilation rates to perform grouped dilated convolution on the second intrinsic feature map and then performing channel concatenation to obtain the output feature map, rich and diverse linear features are generated. Attached Figure Description

[0043] Figure 1 A schematic diagram of the original Ghost module in the prior art provided by the present invention;

[0044] Figure 2 A structural block diagram of a preferred embodiment of the multi-scale linear feature generation system provided by the present invention;

[0045] Figure 3 A schematic diagram of a preferred embodiment of the multi-scale linear feature generation system provided by the present invention;

[0046] Figure 4 A flowchart illustrating a preferred embodiment of the multi-scale linear feature generation method provided by the present invention;

[0047] Figure 5 A flowchart of step S100 of a preferred embodiment of the multi-scale linear feature generation method provided by the present invention;

[0048] Figure 6 A flowchart of step S200 of a preferred embodiment of the multi-scale linear feature generation method provided by the present invention;

[0049] Figure 7 This is a flowchart of step S300 of a preferred embodiment of the multi-scale linear feature generation method provided by the present invention.

[0050] Figure labels: 10: Dimensionality reduction module; 20: Coordinate attention enhancement module; 21: First pooling unit; 22: Second pooling unit; 23: First feature extraction unit; 24: Second feature extraction unit; 25: First mapping unit; 26: Second mapping unit; 27: Pointwise multiplication unit; 30: Multi-scale feature generation module; 31: Grouped dilated convolution unit; 32: Channel concatenation unit. Detailed Implementation

[0051] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0052] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0053] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0054] To facilitate understanding of the embodiments of this application, the relevant elements involved in the embodiments of the present invention will first be introduced here.

[0055] like Figure 2 The process of obtaining the output feature map by using the original Ghost module to produce linear features from the input feature map in the prior art is as follows:

[0056] The original Ghost module first performs pointwise convolution on the input feature map H×W×C to obtain the intrinsic feature map H×W×C / 2. Then, it performs channel-wise convolution to obtain the linear feature map H×W×C / 2. Finally, it concatenates the intrinsic feature map H×W×C / 2 and the linear feature map H×W×C / 2 to obtain the output feature map H×W×C.

[0057] This invention provides a multi-scale linear feature generation system and method. In this invention, a dimensionality reduction module first performs pointwise convolutional dimensionality reduction on the input feature map. A coordinate attention enhancement module then sequentially samples, extracts, and maps the first intrinsic feature map along the X and Y axes, respectively. The resulting first and second weighting factors are then multiplied with the first intrinsic feature map. Finally, a multi-scale feature generation module sequentially performs grouped dilated convolution and channel concatenation on the second intrinsic feature map to obtain the output feature map. In other words, by introducing a coordinate attention enhancement module to incorporate enhanced intrinsic feature maps with positional and global information in the X and Y directions, and by introducing grouped dilated convolution with different dilation rates to generate multi-scale linear features, the Ghost module produces richer and more diverse linear enhanced features, effectively enhancing feature discrimination capabilities and significantly improving the performance of GhostNet models in visual tasks.

[0058] The design scheme of the multi-scale linear feature generation system is described below through specific exemplary embodiments. It should be noted that the following embodiments are only used to explain the technical solution of the invention and are not intended to limit it specifically:

[0059] Please see Figure 1 The present invention provides a multi-scale linear feature generation system, which includes: a dimensionality reduction processing module 10, a coordinate attention enhancement module 20, and a multi-scale feature generation module 30; the dimensionality reduction processing module 10, the coordinate attention enhancement module 20, and the multi-scale feature generation module 30 are connected in sequence.

[0060] The dimensionality reduction processing module 10 performs pointwise convolutional dimensionality reduction processing on the input feature map to obtain a first intrinsic feature map, and transmits the first intrinsic feature map to the coordinate attention enhancement module 20. The coordinate attention enhancement module 20 performs sampling, feature extraction, and mapping processing on the first intrinsic feature map in the X-axis and Y-axis directions respectively, and then multiplies the obtained first weight factor and second weight factor with the first intrinsic feature map to obtain a second intrinsic feature map, which is then transmitted to the multi-scale feature generation module 30. The multi-scale feature generation module 30 performs grouped dilated convolution and channel concatenation processing on the second intrinsic feature map in sequence to obtain an output feature map. The dimensionality reduction processing module 10 includes a 1×1 pointwise convolution.

[0061] Specifically, this application improves the existing Ghost module to obtain an improved Ghost module or a lightweight Ghost module. The improved Ghost module consists of the coordinate attention enhancement module 20 and the multi-scale feature generation module 30.

[0062] See also Figure 3 First, the input feature map is processed using the dimensionality reduction module 10 (1×1 pointwise convolution). Perform pointwise convolution dimensionality reduction to obtain the first intrinsic feature map. (correspond Figure 3 The inherent characteristics of Figure 1 The input feature map is then transmitted to the coordinate attention enhancement module 20, whereby the input feature map undergoes pointwise convolution operations with C / 16 kernels of size 1×1×C, resulting in a first intrinsic feature map with C / 16 channels. Here, H, W, and C represent the height, width, and number of channels of the feature map, respectively.

[0063] Then, the coordinate attention enhancement module 20 sequentially samples, extracts features, and maps the first intrinsic feature map along the X-axis to obtain the first weight factor, and sequentially samples, extracts features, and maps the first intrinsic feature map along the Y-axis to obtain the second weight factor. The coordinate attention enhancement module 20 then multiplies the first and second weight factors with the first intrinsic feature map to obtain the second intrinsic feature map. (correspond Figure 3 The inherent characteristics of Figure 2 ), and the second intrinsic feature map The first intrinsic feature map is transmitted to the multi-scale feature generation module 30. After processing by the coordinate attention enhancement module 20, the second intrinsic feature map is obtained by pointwise convolution to increase its dimensionality to 1 / 4 of the number of channels in the input feature map.

[0064] Finally, the multi-scale feature generation module 30 sequentially performs grouped dilated convolution and channel concatenation on the second intrinsic feature map to obtain an output feature map with C channels. By redesigning the Ghost module, not only are more diverse linear augmented features generated, but the feature discrimination ability is also enhanced, which will ultimately greatly improve the model performance of GhostNet in processing vision tasks.

[0065] Furthermore, the coordinate attention enhancement module 20 includes: a first pooling unit 21, a second pooling unit 22, a first feature extraction unit 23, a second feature extraction unit 24, a first mapping unit 25, a second mapping unit 26, and a pointwise multiplication unit 27; the dimensionality reduction processing module 10, the first pooling unit 21, the first feature extraction unit 23, the first mapping unit 25, and the pointwise multiplication unit 27 are connected in sequence; the dimensionality reduction processing module 10, the second pooling unit 22, the second feature extraction unit 24, the second mapping unit 26, and the pointwise multiplication unit 27 are connected in sequence.

[0066] The first pooling unit 21 is used to perform a first average pooling process on the first intrinsic feature map to obtain a first one-dimensional vector, and transmit the first one-dimensional vector to the first feature extraction unit 23; the first feature extraction unit 23 is used to perform a one-dimensional convolution process on the first one-dimensional vector to obtain a second one-dimensional vector, and transmit the second one-dimensional vector to the first mapping unit 25; the first mapping unit 25 is used to perform a mapping process on the second one-dimensional vector using an activation function to obtain a third one-dimensional vector, and transmit the third one-dimensional vector to the pointwise multiplication unit 27.

[0067] The second pooling unit 22 is used to perform a second average pooling process on the first intrinsic feature map to obtain a fourth one-dimensional vector, and transmit the fourth one-dimensional vector to the second feature extraction unit 24; the second feature extraction unit 24 is used to perform a one-dimensional convolution process on the fourth one-dimensional vector to obtain a fifth one-dimensional vector, and transmit the fifth one-dimensional vector to the second mapping unit 26; the second mapping unit 26 is used to perform a mapping process on the fifth one-dimensional vector using the activation function to obtain a sixth one-dimensional vector, and transmit the sixth one-dimensional vector to the pointwise multiplication unit 27.

[0068] The first pooling unit 21 includes an X-axis average pooling layer; the second pooling unit 22 includes a Y-axis average pooling layer; the X-axis average pooling layer is connected to the dimensionality reduction processing module 10 and the first feature extraction unit 23 respectively; the Y-axis average pooling layer is connected to the dimensionality reduction processing module 10 and the second feature extraction unit 24 respectively; the X-axis average pooling layer is used to perform X-axis average pooling processing on the first intrinsic feature map to obtain the first one-dimensional vector, and transmit the first one-dimensional vector to the first feature extraction unit 23; the Y-axis average pooling layer is used to perform Y-axis average pooling processing on the first intrinsic feature map to obtain the fourth one-dimensional vector, and transmit the fourth one-dimensional vector to the second feature extraction unit 24.

[0069] The pointwise multiplication unit 27 is used to perform pointwise multiplication on the third one-dimensional vector, the sixth one-dimensional vector, and the first intrinsic feature map to obtain a second intrinsic feature map, and then transmits the second intrinsic feature map to the multi-scale feature generation module 30. The first weighting factor includes the third one-dimensional vector; the second weighting factor includes the sixth one-dimensional vector; and the activation function includes the Sigmoid function.

[0070] Specifically, firstly, the input feature map is processed using the dimensionality reduction module 10. Perform pointwise convolution dimensionality reduction to obtain the first intrinsic feature map. Then, the dimensionality reduction processing module 10 transmits the first intrinsic feature map along the upper, middle, and lower branches of the attention enhancement module to the first pooling unit 21 (X-axis average pooling layer), the point-by-point multiplication unit 27, and the second pooling unit 22 (Y-axis average pooling layer), respectively; then, the first pooling unit 21 performs a first average pooling process on the first intrinsic feature map to obtain a first one-dimensional vector (corresponding to...). Figure 3 The first feature extraction unit 23 (a 1D convolutional layer in this embodiment) performs a one-dimensional convolution on the first one-dimensional vector to obtain a second one-dimensional vector (corresponding to the first one-dimensional vector 1); the first feature extraction unit 23 (a 1D convolutional layer in this embodiment) then performs a one-dimensional convolution on the first one-dimensional vector to obtain a second one-dimensional vector (corresponding to the first one-dimensional vector 1). Figure 3 The first mapping unit 25 uses an activation function (Sigmoid function) to map the second one-dimensional vector to obtain a third one-dimensional vector (the first weight factor, corresponding to the first one-dimensional vector). Figure 3 1D vector 3 in the middle.

[0071] Secondly, similarly, the second pooling unit 22 performs a second average pooling process on the first intrinsic feature map to obtain a fourth one-dimensional vector (corresponding to...). Figure 3The second feature extraction unit 24 (in this embodiment, 1D convolution) then performs a one-dimensional convolution on the fourth one-dimensional vector to obtain the fifth one-dimensional vector (corresponding to the 1D vector 4). Figure 3 The second mapping unit 26 uses the activation function to map the fifth one-dimensional vector (5), and obtains the sixth one-dimensional vector (the second weight factor, corresponding to the second weight factor) by mapping the fifth one-dimensional vector. Figure 3 1D vectors in 6).

[0072] Finally, the point-by-point multiplication unit 27 performs point-by-point multiplication on the third one-dimensional vector, the sixth one-dimensional vector, and the first intrinsic feature map to obtain the second intrinsic feature map, so that the multi-scale feature generation module 30 can process the second intrinsic feature map.

[0073] The attention enhancement module is divided into three branches: upper, middle, and lower. The upper branch represents intrinsic features. Figure 1 After X-axis average pooling, a 1D vector 1 (first one-dimensional vector) is obtained. Then, it is processed by a 1D convolutional layer (the first feature extraction unit 23, such as using a convolutional kernel size of 3×1×C / 16, a stride of 1, and padding of 1) to obtain a 1D vector 2 (second one-dimensional vector). Finally, it is processed by the Sigmoid function (the activation function) to obtain a 1D vector 3 (third one-dimensional vector) representing the weight factors. By performing average pooling along the X-axis, positional information in the X-direction and global information in the Y-direction can be incorporated.

[0074] Similarly, the lower branch represents intrinsic characteristics. Figure 1 After Y-axis average pooling, a 1D vector 4 (the fourth one-dimensional vector) is obtained. Then, it is processed by a 1D convolution (using the second feature extraction unit 24, such as a convolution kernel size of 1×3×C / 16, a stride of 1, and padding of 1) to obtain a 1D vector 5 (the fifth one-dimensional vector). Finally, after processing with the Sigmoid function, a 1D vector 6 (the sixth one-dimensional vector) representing the weight factors is obtained. By performing average pooling along the Y-axis, positional information in the Y-direction and global information in the X-direction can be incorporated. This achieves the integration of positional information in the X and Y directions and global information into the enhanced intrinsic feature map by introducing the coordinate attention enhancement module 20.

[0075] Finally, the two 1D weight factors (the first weight factor and the second weight factor) obtained from the upper and lower branches are expanded into a 2D attention weight matrix (the size of the 2D attention weight matrix is ​​related to its intrinsic features). Figure 1 The intrinsic characteristics of identity mapping between the same size and intermediate branches Figure 1 By multiplying channel by channel and point by point, we can obtain the intrinsic features of the coordinates after attention enhancement, which incorporate global and positional information. Figure 2 (Second intrinsic feature map).

[0076] Furthermore, the multi-scale feature generation module 30 includes: a grouped dilated convolution unit 31 and a channel splicing unit 32; the pointwise multiplication unit 27, the grouped dilated convolution unit 31, and the channel splicing unit 32 are connected in sequence.

[0077] The grouped dilated convolution unit 31 is used to perform grouped dilated convolution processing on the second intrinsic feature map to obtain a first multi-scale linear feature map, a second multi-scale linear feature map, and a third multi-scale linear feature map, which are then transmitted to the channel splicing unit 32. The channel splicing unit 32 is used to splice the second intrinsic feature map, the first multi-scale linear feature map, the second multi-scale linear feature map, and the third multi-scale linear feature map to obtain the output feature map.

[0078] The grouped dilated convolution unit 31 includes: a first grouped dilated convolution layer, a second grouped dilated convolution layer, and a third grouped dilated convolution layer; the first grouped dilated convolution layer is connected to the point-by-point multiplication unit 27 and the channel splicing unit 32, respectively; the second grouped dilated convolution layer is connected to the point-by-point multiplication unit 27 and the channel splicing unit 32, respectively; and the third grouped dilated convolution layer is connected to the point-by-point multiplication unit 27 and the channel splicing unit 32, respectively.

[0079] The first grouped dilated convolutional layer is used to perform a first grouped dilated convolutional process on the second intrinsic feature map to obtain the first multi-scale linear feature map, which is then transmitted to the channel stitching unit 32; the second grouped dilated convolutional layer is used to perform a second grouped dilated convolutional process on the second intrinsic feature map to obtain the second multi-scale linear feature map, which is then transmitted to the channel stitching unit 32; the third grouped dilated convolutional layer is used to perform a third grouped dilated convolutional process on the second intrinsic feature map to obtain the third multi-scale linear feature map, which is then transmitted to the channel stitching unit 32.

[0080] Among them, grouped dilated convolution refers to dilated convolution with the number of groups equal to the number of channels in the input feature map, which can complete the channel-by-channel processing of the input feature map. Based on the principle of expanding the receptive field by dilated convolution, dilated convolution with different dilation rates is used to produce multi-scale effects.

[0081] Specifically, after the pointwise multiplication unit 27 performs pointwise multiplication on the third one-dimensional vector, the sixth one-dimensional vector, and the first intrinsic feature map to obtain the second intrinsic feature map, the pointwise multiplication unit 27 transmits the second intrinsic feature map to the grouped dilation convolution unit 31 and the channel splicing unit 32, respectively. Then, the first grouped dilation convolution layer, the first grouped dilation convolution layer, and the first grouped dilation convolution layer in the grouped dilation convolution unit 31 respectively perform grouped dilation convolution on the second intrinsic feature map to obtain the first multi-scale linear feature map. Second multi-scale linear feature map and the third multi-scale linear feature map Imminent Intrinsic Features Figure 2 After processing by the multi-scale feature generation module 30, which uses grouped dilated convolutions with dilation rates of 1, 2, and 4 to simultaneously process the intrinsic feature X2 along the three branches, multi-scale linear features are obtained. and

[0082] After processing each channel individually through grouped dilated convolutional layers, the size of the input feature map (the second intrinsic feature map) and the output feature map (multi-scale linear features) remain unchanged, and the number of channels in the input feature map equals the number of channels in the output feature map. The inexpensive linear operation of the Ghost module is redesigned by introducing grouped dilated convolutional layers with different dilation rates to generate multi-scale linear features, making the linear features produced by the lightweight Ghost module richer and more diverse. Specifically, the dilation rate in the dilation convolution parameter settings is equal to the padding to maintain the input and output sizes constant.

[0083] Finally, the channel stitching unit 32 stitches the second intrinsic feature map, the first multi-scale linear feature map, the second multi-scale linear feature map, and the third multi-scale linear feature map to obtain the output feature map. In other words, intrinsic feature X2 and multi-scale linear feature X 21 X 22 X 23 The feature maps of the four branches, each with a channel count of C / 4, are concatenated along the channel dimension to obtain the output feature map with a channel count of C.

[0084] In another embodiment of the present invention, different convolutional neural network models adopt an inverted bottleneck design. Generally, two Ghost modules are used to construct a Ghost bottleneck, and then different convolutional neural network structures are constructed using the Ghost bottleneck for visual tasks such as image classification, image segmentation, and object detection.

[0085] Further, please refer to Figure 4 The present invention provides a multi-scale linear feature generation method based on the multi-scale linear feature generation system described above, the multi-scale linear feature generation method comprising the following steps:

[0086] S100, the dimensionality reduction processing module 10 performs pointwise convolution dimensionality reduction processing on the input feature map and transmits the resulting first intrinsic feature map to the coordinate attention enhancement module 20.

[0087] Specifically, the multi-scale linear feature generation method described in this invention is implemented based on the multi-scale linear feature generation system, and the multi-scale linear feature generation method is implemented as follows:

[0088] First, the dimensionality reduction processing module 10 (1×1 pointwise convolution) processes the input feature map. Dimensionality reduction is achieved by performing pointwise convolution (C / 16 convolutional kernels of size 1×1×C), resulting in a first intrinsic feature map with C / 16 channels. (correspond Figure 3 The inherent characteristics of Figure 1 The obtained first intrinsic feature map is transmitted to the coordinate attention enhancement module 20 so that the coordinate attention enhancement module 20 can perform sampling, feature extraction and mapping processing.

[0089] In order to make the model lightweight, when generating intrinsic features, pointwise convolution is first used to reduce the dimensionality to 1 / 16 of the number of channels in the input feature map. Then, after processing by the coordinate attention enhancement module 20, pointwise convolution is used to increase the dimensionality to 1 / 4 of the number of channels in the input feature map.

[0090] Furthermore, please refer to Figure 5 S100, the dimensionality reduction processing module 10 performs pointwise convolutional dimensionality reduction processing on the input feature map and transmits the resulting first intrinsic feature map to the coordinate attention enhancement module 20, specifically including:

[0091] S110. After receiving the input feature map, the dimensionality reduction processing module 10 performs pointwise convolution dimensionality reduction processing on the input feature map to obtain the first intrinsic feature map.

[0092] S120, the dimensionality reduction processing module 10 transmits the first intrinsic feature map to the first pooling unit 21, the second pooling unit 22 and the point-by-point multiplication unit 27.

[0093] Specifically, the dimensionality reduction processing module 10 (1×1 pointwise convolution) processes the input feature map. Dimensionality reduction is achieved by performing pointwise convolution (C / 16 convolutional kernels of size 1×1×C), resulting in a first intrinsic feature map with C / 16 channels. (correspond Figure 3 The inherent characteristics of Figure 1 Then, the dimensionality reduction processing module 10 transmits the first intrinsic feature map to the first pooling unit 21, the second pooling unit 22 and the point-by-point multiplication unit 27 respectively.

[0094] For further information, please refer to [link / reference]. Figure 4 S200, the coordinate attention enhancement module 20 samples, extracts features and maps the first intrinsic feature map in the X-axis direction and the Y-axis direction respectively, and then multiplies the obtained first weight factor and second weight factor with the first intrinsic feature map, and transmits the obtained second intrinsic feature map to the multi-scale feature generation module.

[0095] Specifically, after the dimensionality reduction processing module 10 transmits the obtained first intrinsic feature map to the coordinate attention enhancement module 20, the coordinate attention enhancement module 20 sequentially performs sampling, feature extraction, and mapping processing on the first intrinsic feature map in the X-axis direction to obtain the first weight factor, and sequentially performs sampling, feature extraction, and mapping processing on the first intrinsic feature map in the Y-axis direction to obtain the second weight factor. Then, the coordinate attention enhancement module 20 multiplies the first weight factor and the second weight factor with the first intrinsic feature map to obtain the second intrinsic feature map. (correspond Figure 3 The inherent characteristics of Figure 2 ), and the second intrinsic feature map The data is transmitted to the multi-scale feature generation module 30 so that the multi-scale feature generation module 30 can perform the next operation.

[0096] Furthermore, please refer to Figure 6 S200, the coordinate attention enhancement module 20 sequentially samples, extracts features, and maps the first intrinsic feature map in the X-axis and Y-axis directions, respectively. Then, it multiplies the obtained first weight factor and second weight factor with the first intrinsic feature map and transmits the resulting second intrinsic feature map to the multi-scale feature generation module 30. Specifically, this includes:

[0097] S210. After the first pooling unit 21 performs X-axis average pooling on the first intrinsic feature map, the resulting first one-dimensional vector is transmitted to the first feature extraction unit 23; wherein, the first pooling unit 21 includes: an X-axis average pooling layer.

[0098] S220. After the first feature extraction unit 23 performs one-dimensional convolution processing on the first one-dimensional vector, the resulting second one-dimensional vector is transmitted to the first mapping unit 25.

[0099] S230, after the first mapping unit 25 performs mapping processing on the second one-dimensional vector using an activation function, the resulting third one-dimensional vector is transmitted to the point-by-point multiplication unit 27; wherein, the first weighting factor includes: the third one-dimensional vector;

[0100] S240, after the second pooling unit 22 performs Y-axis average pooling on the first intrinsic feature map, the resulting fourth one-dimensional vector is transmitted to the second feature extraction unit 24; wherein, the second pooling unit 22 includes: a Y-axis average pooling layer;

[0101] S250, after the second feature extraction unit 24 performs one-dimensional convolution on the fourth one-dimensional vector, the resulting fifth one-dimensional vector is transmitted to the second mapping unit 26.

[0102] S260, the second mapping unit 26 uses the activation function to map the fifth one-dimensional vector and then transmits the resulting sixth one-dimensional vector to the pointwise multiplication unit 27; wherein, the second weighting factor includes: the sixth one-dimensional vector; the activation function includes: the Sigmoid function;

[0103] S270, after expanding the third one-dimensional vector and the sixth one-dimensional vector to obtain the attention weight matrix, the point-by-point multiplication unit 27 performs point-by-point multiplication on the attention weight matrix and the first intrinsic feature map, and transmits the obtained second intrinsic feature map to the grouped dilated convolution unit 31 and the channel splicing unit 32.

[0104] Specifically, in the dimensionality reduction processing module 10, the first intrinsic feature map is transmitted along the upper, middle, and lower branches of the attention enhancement module to the first pooling unit 21 (X-axis average pooling layer), the point-by-point multiplication unit 27, and the second pooling unit 22 (Y-axis average pooling layer), respectively; then, the first pooling unit 21 performs a first average pooling process on the first intrinsic feature map to obtain a first one-dimensional vector (corresponding to...). Figure 3The first feature extraction unit 23 (a 1D convolutional layer in this embodiment) performs a one-dimensional convolution on the first one-dimensional vector to obtain a second one-dimensional vector (corresponding to the first one-dimensional vector 1); the first feature extraction unit 23 (a 1D convolutional layer in this embodiment) then performs a one-dimensional convolution on the first one-dimensional vector to obtain a second one-dimensional vector (corresponding to the first one-dimensional vector 1). Figure 3 The first mapping unit 25 uses an activation function (Sigmoid function) to map the second one-dimensional vector to obtain a third one-dimensional vector (the first weight factor, corresponding to the first one-dimensional vector). Figure 3 1D vector 3 in the middle.

[0105] Secondly, similarly, the second pooling unit 22 performs a second average pooling process on the first intrinsic feature map to obtain a fourth one-dimensional vector (corresponding to...). Figure 3 The second feature extraction unit 24 (in this embodiment, 1D convolution) then performs a one-dimensional convolution on the fourth one-dimensional vector to obtain the fifth one-dimensional vector (corresponding to the 1D vector 4). Figure 3 The second mapping unit 26 uses the activation function to map the fifth one-dimensional vector (5), and obtains the sixth one-dimensional vector (the second weight factor, corresponding to the second weight factor) by mapping the fifth one-dimensional vector. Figure 3 1D vectors in 6).

[0106] Finally, the point-by-point multiplication unit 27 performs point-by-point multiplication on the third one-dimensional vector, the sixth one-dimensional vector, and the first intrinsic feature map to obtain the second intrinsic feature map, so that the multi-scale feature generation module 30 can process the second intrinsic feature map.

[0107] For further information, please refer to [link / reference]. Figure 4 S300, the multi-scale feature generation module 30 sequentially performs grouped dilated convolution and channel concatenation on the second intrinsic feature map to obtain the output feature map.

[0108] Specifically, after the coordinate attention enhancement module 20 transmits the second intrinsic feature map to the multi-scale feature generation module 30, the multi-scale feature generation module 30 performs grouped dilated convolution and channel concatenation on the second intrinsic feature map sequentially to obtain an output feature map with C channels.

[0109] Furthermore, please refer to Figure 7 S300, the multi-scale feature generation module 30 sequentially performs grouped dilated convolution and channel concatenation on the second intrinsic feature map to obtain an output feature map, specifically including:

[0110] S310. After the grouped dilated convolution unit 31 performs a first grouped dilated convolution on the second intrinsic feature map using a first dilation rate, the resulting first multi-scale linear feature map is transmitted to the channel stitching unit 32; wherein, the grouped dilated convolution unit 31 includes: a first grouped dilated convolution layer, a second grouped dilated convolution layer and a third grouped dilated convolution layer.

[0111] S320. After the grouped dilated convolution unit 31 performs the second grouped dilated convolution process on the second intrinsic feature map using the second dilation rate, the resulting second multi-scale linear feature map is transmitted to the channel splicing unit 32.

[0112] S330. After the grouped dilation convolution unit 31 performs third grouped dilation convolution processing on the second intrinsic feature map using the third dilation rate, the resulting third multi-scale linear feature map is transmitted to the channel splicing unit 32.

[0113] S340, the channel splicing unit 32 splices the second intrinsic feature map, the first multi-scale linear feature map, the second multi-scale linear feature map and the third multi-scale linear feature map to obtain the output feature map.

[0114] Specifically, after the point-by-point multiplication unit 27 transmits the second intrinsic feature map to the grouped dilation convolution unit 31 and the channel splicing unit 32 respectively, the grouped dilation convolution unit 31 performs a first grouped dilation convolution process on the second intrinsic feature map using the first dilation rate (1 in this embodiment) to obtain the first multi-scale linear feature map. Similarly, the grouped dilated convolution unit 31 uses the second dilation rate (2 in this embodiment) to perform a second grouped dilated convolution process on the second intrinsic feature map to obtain the second multi-scale linear feature map. The grouped dilated convolution unit 31 uses the third dilation rate (4 in this embodiment) to perform a third grouped dilated convolution process on the second intrinsic feature map to obtain the third multi-scale linear feature map.

[0115] Then, the grouped dilated convolution unit 31 divides the first multi-scale linear feature map Second multi-scale linear feature map and the third multi-scale linear feature map The data is transmitted to the channel stitching unit 32, which then stitches the second intrinsic feature map, the first multi-scale linear feature map, the second multi-scale linear feature map, and the third multi-scale linear feature map—these four branches—with a channel count of C / 4, along the channel dimension to obtain an output feature map with a channel count of C.

[0116] This invention introduces grouped dilated convolutions with different dilation rates to generate multi-scale linear features, which greatly enriches the diversity of linear features generated by the Ghost module, thereby greatly improving the model performance of GhostNet in processing visual tasks.

[0117] In summary, this invention provides a multi-scale linear feature generation system and method. The system includes a dimensionality reduction module that performs pointwise convolutional dimensionality reduction on the input feature map and transmits the resulting first intrinsic feature map to a coordinate attention enhancement module. The coordinate attention enhancement module sequentially samples, extracts features, and maps the first intrinsic feature map along the X-axis and Y-axis, respectively. It then multiplies the resulting first and second weighting factors with the first intrinsic feature map and transmits the resulting second intrinsic feature map to a multi-scale feature generation module. The multi-scale feature generation module sequentially performs grouped dilated convolution and channel concatenation on the second intrinsic feature map to obtain an output feature map. By using different dilation rates to perform grouped dilated convolution on the second intrinsic feature map and then performing channel concatenation to obtain the output feature map, rich and diverse linear features are generated.

[0118] It is understood that those skilled in the art can make equivalent substitutions or modifications to the technical solution and inventive concept of the present invention, and all such substitutions or modifications should fall within the protection scope of the appended claims.

Claims

1. A multi-scale linear feature generation system, characterized in that, The multi-scale linear feature generation system includes: a dimensionality reduction processing module, a coordinate attention enhancement module, and a multi-scale feature generation module; the dimensionality reduction processing module, the coordinate attention enhancement module, and the multi-scale feature generation module are connected in sequence. The dimensionality reduction processing module is used to perform pointwise convolution dimensionality reduction processing on the input feature map to obtain a first intrinsic feature map, and then transmit the first intrinsic feature map to the coordinate attention enhancement module; The coordinate attention enhancement module is used to sample, extract features and map the first intrinsic feature map in the X-axis direction and the Y-axis direction respectively, and then multiply the obtained first weight factor and second weight factor with the first intrinsic feature map to obtain the second intrinsic feature map, and then transmit the second intrinsic feature map to the multi-scale feature generation module. The multi-scale feature generation module is used to sequentially perform grouped dilated convolution and channel concatenation on the second intrinsic feature map to obtain the output feature map; The coordinate attention enhancement module includes: a first pooling unit, a second pooling unit, a first feature extraction unit, a second feature extraction unit, a first mapping unit, a second mapping unit, and a pointwise multiplication unit; The dimensionality reduction processing module, the first pooling unit, the first feature extraction unit, the first mapping unit, and the point-by-point multiplication unit are connected in sequence; the dimensionality reduction processing module, the second pooling unit, the second feature extraction unit, the second mapping unit, and the point-by-point multiplication unit are also connected in sequence. The first pooling unit is used to perform a first average pooling process on the first intrinsic feature map to obtain a first one-dimensional vector, and transmit the first one-dimensional vector to the first feature extraction unit; the first feature extraction unit is used to perform a one-dimensional convolution process on the first one-dimensional vector to obtain a second one-dimensional vector, and transmit the second one-dimensional vector to the first mapping unit; the first mapping unit is used to perform a mapping process on the second one-dimensional vector using an activation function to obtain a third one-dimensional vector, and transmit the third one-dimensional vector to the pointwise multiplication unit. The second pooling unit is used to perform a second average pooling process on the first intrinsic feature map to obtain a fourth one-dimensional vector, and transmits the fourth one-dimensional vector to the second feature extraction unit; the second feature extraction unit is used to perform a one-dimensional convolution process on the fourth one-dimensional vector to obtain a fifth one-dimensional vector, and transmits the fifth one-dimensional vector to the second mapping unit; the second mapping unit is used to perform a mapping process on the fifth one-dimensional vector using the activation function to obtain a sixth one-dimensional vector, and transmits the sixth one-dimensional vector to the pointwise multiplication unit; The point-by-point multiplication unit is used to perform point-by-point multiplication on the third one-dimensional vector, the sixth one-dimensional vector, and the first intrinsic feature map to obtain the second intrinsic feature map, and then transmit the second intrinsic feature map to the multi-scale feature generation module.

2. The multi-scale linear feature generation system according to claim 1, characterized in that, The multi-scale feature generation module includes: a grouped dilated convolution unit and a channel splicing unit; the pointwise multiplication unit, the grouped dilated convolution unit, and the channel splicing unit are connected in sequence; The grouped dilated convolution unit is used to perform grouped dilated convolution processing on the second intrinsic feature map to obtain a first multi-scale linear feature map, a second multi-scale linear feature map, and a third multi-scale linear feature map, which are then transmitted to the channel splicing unit. The channel splicing unit is used to splice the second intrinsic feature map, the first multi-scale linear feature map, the second multi-scale linear feature map, and the third multi-scale linear feature map to obtain the output feature map.

3. The multi-scale linear feature generation system according to claim 1, characterized in that, The first pooling unit includes an X-axis average pooling layer; the second pooling unit includes a Y-axis average pooling layer; the X-axis average pooling layer is connected to the dimensionality reduction processing module and the first feature extraction unit respectively; the Y-axis average pooling layer is connected to the dimensionality reduction processing module and the second feature extraction unit respectively. The X-axis average pooling layer is used to perform X-axis average pooling processing on the first intrinsic feature map to obtain the first one-dimensional vector, and transmit the first one-dimensional vector to the first feature extraction unit; the Y-axis average pooling layer is used to perform Y-axis average pooling processing on the first intrinsic feature map to obtain the fourth one-dimensional vector, and transmit the fourth one-dimensional vector to the second feature extraction unit.

4. The multi-scale linear feature generation system according to claim 1, characterized in that, The first weighting factor includes the third one-dimensional vector; the second weighting factor includes the sixth one-dimensional vector; and the activation function includes the Sigmoid function.

5. The multi-scale linear feature generation system according to claim 2, characterized in that, The grouped dilated convolution unit includes: a first grouped dilated convolution layer, a second grouped dilated convolution layer, and a third grouped dilated convolution layer; the first grouped dilated convolution layer is connected to the pointwise multiplication unit and the channel splicing unit, respectively; the second grouped dilated convolution layer is connected to the pointwise multiplication unit and the channel splicing unit, respectively; and the third grouped dilated convolution layer is connected to the pointwise multiplication unit and the channel splicing unit, respectively. The first grouped dilated convolutional layer is used to perform a first grouped dilated convolutional process on the second intrinsic feature map to obtain the first multi-scale linear feature map, which is then transmitted to the channel stitching unit; the second grouped dilated convolutional layer is used to perform a second grouped dilated convolutional process on the second intrinsic feature map to obtain the second multi-scale linear feature map, which is then transmitted to the channel stitching unit; the third grouped dilated convolutional layer is used to perform a third grouped dilated convolutional process on the second intrinsic feature map to obtain the third multi-scale linear feature map, which is then transmitted to the channel stitching unit.

6. A method for generating multi-scale linear features based on the multi-scale linear feature generation system according to any one of claims 1-5, characterized in that, The multi-scale linear feature generation method includes the following steps: The dimensionality reduction processing module performs point-by-point convolution dimensionality reduction processing on the input feature map and transmits the resulting first intrinsic feature map to the coordinate attention enhancement module. The coordinate attention enhancement module samples, extracts features, and maps the first intrinsic feature map sequentially in the X-axis and Y-axis directions, respectively. Then, it multiplies the obtained first weight factor and second weight factor with the first intrinsic feature map and transmits the obtained second intrinsic feature map to the multi-scale feature generation module. The multi-scale feature generation module sequentially performs grouped dilated convolution and channel concatenation on the second intrinsic feature map to obtain the output feature map.

7. The multi-scale linear feature generation method according to claim 6, characterized in that, The dimensionality reduction processing module performs pointwise convolutional dimensionality reduction on the input feature map and transmits the resulting first intrinsic feature map to the coordinate attention enhancement module, specifically including: After receiving the input feature map, the dimensionality reduction processing module performs pointwise convolution dimensionality reduction processing on the input feature map to obtain the first intrinsic feature map; The dimensionality reduction processing module transmits the first intrinsic feature map to the first pooling unit, the second pooling unit, and the pointwise multiplication unit.

8. The multi-scale linear feature generation method according to claim 7, characterized in that, The coordinate attention enhancement module sequentially samples, extracts features, and maps the first intrinsic feature map along the X-axis and Y-axis, respectively. It then multiplies the resulting first and second weighting factors with the first intrinsic feature map and transmits the resulting second intrinsic feature map to the multi-scale feature generation module. Specifically, this includes: After performing X-axis average pooling on the first intrinsic feature map, the first pooling unit transmits the resulting first one-dimensional vector to the first feature extraction unit; wherein, the first pooling unit includes: an X-axis average pooling layer; After performing a one-dimensional convolution on the first one-dimensional vector, the first feature extraction unit transmits the resulting second one-dimensional vector to the first mapping unit. After the first mapping unit performs mapping processing on the second one-dimensional vector using an activation function, the resulting third one-dimensional vector is transmitted to the pointwise multiplication unit; wherein, the first weight factor includes: the third one-dimensional vector; After performing Y-axis average pooling on the first intrinsic feature map, the second pooling unit transmits the resulting fourth one-dimensional vector to the second feature extraction unit; wherein, the second pooling unit includes: a Y-axis average pooling layer; After performing a one-dimensional convolution on the fourth one-dimensional vector, the second feature extraction unit transmits the resulting fifth one-dimensional vector to the second mapping unit. After the second mapping unit performs mapping processing on the fifth one-dimensional vector using the activation function, it transmits the resulting sixth one-dimensional vector to the pointwise multiplication unit; wherein, the second weighting factor includes: the sixth one-dimensional vector; The activation function includes: the Sigmoid function; The point-by-point multiplication unit expands the third one-dimensional vector and the sixth one-dimensional vector to obtain the attention weight matrix, then performs point-by-point multiplication on the attention weight matrix and the first intrinsic feature map, and transmits the resulting second intrinsic feature map to the grouped dilated convolution unit and the channel splicing unit.

9. The multi-scale linear feature generation method according to claim 8, characterized in that, The multi-scale feature generation module sequentially performs grouped dilated convolution and channel concatenation on the second intrinsic feature map to obtain an output feature map, specifically including: The grouped dilated convolution unit performs a first grouped dilated convolution on the second intrinsic feature map using a first dilation rate, and then transmits the resulting first multi-scale linear feature map to the channel stitching unit; wherein, the grouped dilated convolution unit includes: a first grouped dilated convolution layer, a second grouped dilated convolution layer and a third grouped dilated convolution layer; After the grouped dilatation convolution unit performs the second grouped dilatation convolution on the second intrinsic feature map using the second dilation rate, the resulting second multi-scale linear feature map is transmitted to the channel stitching unit. The grouped dilatation convolution unit performs a third grouped dilatation convolution on the second intrinsic feature map using a third dilation rate, and then transmits the resulting third multi-scale linear feature map to the channel stitching unit. The channel stitching unit stitches the second intrinsic feature map, the first multi-scale linear feature map, the second multi-scale linear feature map, and the third multi-scale linear feature map to obtain the output feature map.

Images