CN-122024284-A - Sichuan monkey lightweight detection method based on edge-aware double-fusion network

CN122024284ACN 122024284 ACN122024284 ACN 122024284ACN-122024284-A

Abstract

The invention discloses a method for detecting light-weight of a golden monkey based on an edge-aware double-fusion network, and relates to the technical field of image detection. The invention provides a lightweight detection method aiming at the problems of complex background, serious shielding, difficult small target identification, limited calculation force of edge equipment and the like in wild monitoring of a golden monkey. The method comprises the steps of firstly collecting images and constructing a dataset, secondly constructing a EADFNet network based on YOLOv n improvement, introducing an edge perception Stem module (EAS) at the front end of the network to strengthen boundaries and inhibit backgrounds, using a partial dynamic weighted residual block (PDW) to stabilize feature learning in a feature extraction stage, using a high-low frequency correlation guide double fusion module (HCF) to realize efficient fusion without splicing in a feature fusion stage, and introducing a small target detail head (SOD) in a detection head. The invention greatly reduces the parameter quantity and the calculation cost and simultaneously remarkably improves the detection precision and recall rate of the golden monkey in a complex field environment.

Inventors

ZHANG JUNGUO
LIU YANG
XIE SHANSHAN

Assignees

北京林业大学

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (7)

1. EADFNet-based light-weight detection method for golden monkey is characterized by comprising the following steps: s1, acquiring image data of a golden monkey in a field environment, constructing an experimental data set and performing pretreatment operation; S2, constructing EADFNet a detection network model, wherein the EADFNet detection network model is based on a YOLOv n architecture and comprises an edge-aware backbone network, a high-low frequency related guide dual-fusion neck network and a detection head; s3, constructing an edge perception Stem module (EAS) at the front end of a backbone network, extracting an edge prior by using a fixed Sobel operator, and generating a space gating signal to inhibit background textures; S4, constructing a partial dynamic weighted residual block (PDW) at a feature extraction layer of the backbone network, generating channel weight by using global statistical information and dynamically re-weighting the features; S5, constructing a high-low frequency related guide double fusion module (HCF) in the neck network, and fusing high-level semantic features and low-level detail features by utilizing the related guide and edge enhancement under the condition of not performing channel splicing; s6, constructing a small target detail head (SOD) at the detection head part, aggregating multi-scale depth convolution characteristics and injecting channel attention to enhance the detection capability of a long-distance small target; and S7, training a EADFNet detection network model by using the constructed experimental data set to obtain optimal weight, and detecting the golden monkey in real time by using the trained model.
2. The EADFNet-based light-weight detection method for golden monkey according to claim 1, wherein the edge-aware Stem module (EAS) in S3 specifically operates as follows: s3.1 for input features Respectively carrying out main path convolution extraction and edge branch extraction; s3.2, in the edge branches, a fixed horizontal Sobel operator is utilized And a vertical Sobel operator For input features Performing depth convolution to generate gradient response horizontal direction And in the vertical direction In which, in the process, Representing a depth convolution; S3.3, calculating an edge amplitude chart And pass through Generating spatial gating by convolution projection and Sigmoid activation function In which, in the process, Representing the Sigmoid activation function, Representation of Convolving; The main path is characterized by the following formula And space gating Soft-gated fusion and output features : In which, in the process, Representing element-by-element multiplication.
3. The EADFNet-based golden monkey lightweight detection method as claimed in claim 1, wherein the S4 middle partial motion weighted residual block (PDW) specifically operates as follows: S4.1 by Convolution will input features The generation of the characteristic expression is as follows: In which, in the process, Representation of The function is activated and the function is activated, For the purpose of batch normalization, Representation of Convolving; input features expressed as partially dynamically weighted residual blocks, using The expression of the depth convolution extraction local spatial feature is as follows: In which, in the process, The method comprises the steps of representing depth convolution, namely, independently convolving each channel of input features by using a convolution kernel to extract spatial features in a single channel, wherein Z is represented as features compressed by the channels; S4.2, for characteristics of Global averaging pooling GlobalAveragingPooling, GAP to get descriptors And generating a channel weight expression by grouping convolution and Sigmoid functions as follows: In which, in the process, The grouping convolution operation is represented, wherein the grouping convolution is to divide channels of input features into a plurality of groups, and convolution operation is independently carried out in each group, so that the parameter number and the calculation cost can be remarkably reduced compared with the common convolution; representing Sigmoid activation functions for mapping output values to The interval to generate normalized weight coefficients; Represented as global feature descriptors pooled by global averaging; Represented as channel weights, for dynamically adjusting the importance of the feature channels; s4.3, utilizing the channel weight For local spatial features And carrying out channel-level dynamic scaling to obtain a weighted characteristic formula as follows: In which, in the process, Representing element-by-element multiplication; S4.4 by Convolving the spread channels and outputting final features via residual connections The calculation formula is as follows: wherein Representation of The function is activated and the function is activated, For the purpose of batch normalization, Representation of And (5) convolution.
4. The EADFNet-based light-weight detection method for golden monkey according to claim 1, wherein the high-low frequency correlation-guided double fusion module (HCF) in S5 specifically operates as follows: s5.1, high-level features And low-level features Projected to the same width, the resulting formula is: And ; In the formula, The representation SiLU activates the function, For the purpose of batch normalization, Representation of Convolution, and carrying out space normalization to obtain the formula: And In the formula, The mean value is represented as such, The standard deviation is indicated as such, And (3) with Are all calculated from the dimensions of the space, Is a very small positive constant set to prevent the denominator from becoming zero; S5.2, calculating a position correlation diagram of the high-layer and low-layer characteristics The formula is: In which, in the process, Representing an element-by-element multiplication, Representing averaging along a channel, computing residuals of feature differences To restore the details of the blur, the specific expression is: In which, in the process, Representing a depth convolution; S5.3, combining edge guide information extracted from high-level features The fusion characteristics are calculated by the following formula The specific formula is as follows: , ; Wherein, the A spatial multi-layer perceptron is represented, Representing the Sigmoid activation function, In order to enable the gating of the fusion, As the weight coefficient of the light-emitting diode, Representing element-by-element multiplication.
5. The method for detecting the light-weight of the golden monkey based on EADFNet as claimed in claim 1, wherein the small target detail head (SOD) in S6 is specifically operated by inputting characteristics Respectively parallel application And And adding the two results to obtain The specific formula is as follows: In which, in the process, Representation of Is a deep convolution of (1); By passing through Convolving the mixed channel to obtain features In which, in the process, Representation of The function is activated and the function is activated, For the purpose of batch normalization, Representation of Convolution, computing features Is pooled by global averaging of two layers Convolution generated channel attention weights Outputting final characteristics through the following soft gating mechanism : , wherein, Representing element-by-element multiplication.
6. The EADFNet-based light-weight detection method for golden monkey according to claim 1, wherein the S1 specifically comprises: S1.1, collecting golden monkey images under different illumination, different distances and shielding conditions; s1.2, screening the acquired images, removing blurred samples, and constructing an experimental data set; S1.3, labeling the golden monkey targets in the experimental data set by using a labeling tool to generate a label file in a YOLO format; s1.4, dividing the experimental data set into a training set, a verification set and a test set.
7. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the method of any of claims 1-6.

Description

Sichuan monkey lightweight detection method based on edge-aware double-fusion network Technical Field The invention relates to the technical field of image detection, in particular to a Sichuan golden monkey lightweight detection method based on an edge-aware double-fusion network. Background The golden monkey is taken as a special rare or endangered primate in China and is listed as a first-class important wild animal protection in China, and the dynamic monitoring of population quantity and survival state is one of core tasks of biodiversity protection and ecological system assessment. The traditional monitoring relies on manual inspection and manual interpretation after fixed-point shooting of an infrared camera, is time-consuming, labor-consuming and limited in coverage area, is easily affected by factors such as complex terrain, long observation distance and the like, and is difficult to realize continuous and accurate monitoring. With the development of computer vision and edge computing technology, an image-based automatic detection method becomes the main flow direction of wild animal monitoring, so that the space-time limit of manual monitoring can be broken through, and efficient and objective data support is provided for population dynamic analysis. However, the field living environment and species characteristics of the golden-silk monkeys lead to multiple technical bottlenecks in that firstly, the background interference is extremely strong, the golden-silk monkeys inhabit in subtropical mountain needle-leaved forest and needle-leaved hybrid forest, golden hairs of the golden-silk monkeys are highly similar to the environment with the vegetation textures and light shadows in the forest, the difference between the gray level and the texture characteristics of the target and the background is weak, and the vegetation is easily misjudged as the target or the true target is missed by the traditional detection algorithm; the method has the advantages that the method is simple in structure, convenient to use, and easy to operate, the problem of shielding is outstanding, the blocking problem is outstanding, the Sichuan monkey is a social animal, overlapping blocking among individuals is frequent during group activities, the blocking of branches and blades further damages the complete outline of a target, so that effective characteristics are difficult to extract by an algorithm, the small target detection difficulty is high, an infrared camera is often deployed at a remote observation point in field monitoring, the Sichuan monkey occupies extremely small proportion in an image, characteristic information is seriously lost, the recognition rate of the small target by a conventional algorithm is extremely low, the edge deployment requirement is severe, field monitoring equipment is mostly limited in computing power, edge terminals (such as portable infrared cameras and field deployed edge computing nodes) are limited in computing power, the problems of large parameter and computing redundancy exist in the existing general target detection algorithm (such as FasterR-CNN and YOLO series basic models), real-time detection on edge equipment is difficult to realize, for example, a YOLO series characteristic fusion network (such as PANet) is generally operated by adopting channel splicing (Concat), the number of channels is expanded, the calculated amount is increased, and a standard convolution module is not optimized for blocking scenes and fuzzy boundaries, and the efficiency cannot be balanced in complex environments. In summary, the existing detection technology is difficult to simultaneously meet the core requirements of 'complex environment adaptability is strong, small targets and shielding scene detection are accurate, lightweight adaptation edge deployment', and a lightweight and high-precision detection method aiming at the field monitoring scene customization development of the golden monkey is needed to solve the pain point of the existing technology and promote the automation and intelligent upgrading of the golden monkey monitoring. Disclosure of Invention 1. Technical problem to be solved by the invention The invention aims to provide a method for detecting the light-weight of a golden monkey based on an edge-aware double-fusion network, which solves the following problems in the prior art: (1) The detection precision of the golden-silk monkey in the complex field environment is insufficient, and the golden-silk monkey field living environment has the problems of strong background interference, frequent shielding, missing of small target characteristics and the like, so that the traditional detection algorithm is difficult to accurately identify. The golden hair is highly similar to the vegetation texture and the shadow environment under the forest, the object is easy to be confused with the background, the monkey group is easy to be subject to overlap and shelter, the tree branch blades furth