CN-119107270-B - Image enhancement processing method for fishery resource statistics

Abstract

The invention provides an image enhancement processing method for fishery resource statistics, which belongs to the field of image enhancement and aims to realize the effective fusion of color information in low-frequency information and texture and detail characteristics in high-frequency information by constructing a color priori feature enhancement network, wherein the color priori feature enhancement network comprises a high-frequency information enhancement module and a low-frequency information enhancement module, the high-frequency information enhancement module enhances detail characteristics in the fishery resource statistics image, the low-frequency information enhancement module focuses on the accurate correction of image colors, the integral content and color expression of the fishery resource statistics image are better reserved, a frequency information interaction module is constructed, the interaction between the high-frequency information and the low-frequency information is realized, and the fishery resource statistics image enhancement is realized.

Inventors

• KE HAN
• LIU HONGCAI
• WANG BINGLI
• WANG QINGQING
• LIU YA
• Xiao Shanshi

Assignees

• 山东省淡水渔业研究院（山东省淡水渔业监测中心）

Dates

Publication Date

20260508

Application Date

20241108

Claims (2)

1. 1. An image enhancement processing method for fishery resource statistics is characterized by comprising the following steps: S1, collecting fishery resource statistical images, mounting an underwater camera on an underwater robot, cruising and shooting in a specific water area to obtain the image of the underwater fishery resource statistics, manually adding a label into the obtained fishery resource statistical image, and preparing a data set; S2, constructing a color priori feature enhancement network, wherein the color priori feature enhancement network is composed of N feature information enhancement modules, each feature information enhancement module comprises a low-frequency information enhancement module, a high-frequency information enhancement module and a frequency information interaction module, the low-frequency information enhancement module comprises extraction of multi-scale color information, and in order to independently process and extract information of each color channel, the input fishery resource statistical image features are firstly counted The color channels of R, G, B are extracted respectively, , Representing the separation of color channels, extracting local features in the color channels by using a 3X 3 convolution layer and a ReLU activation function, then giving different weights to R, G, B color channels by using an SE attention mechanism, enhancing important features, inhibiting unimportant features, carrying out color feature fusion on the color features processed by the channel attention mechanism by using element addition, carrying out layer normalization processing on the fused color features to obtain multi-scale color information, , A3 x 3 convolutional layer is shown, Representing the function of the ReLU activation, Representing the mechanism of attention of the SE, Normalization processing of a representation layer; S3, constructing a low-frequency information enhancement module which comprises an SE attention mechanism, regularization, GELU activation function, a full-connection layer and a multi-scale color information extraction module, realizing low-frequency information color correction of the fishery resource statistical image, and inputting the fishery resource statistical image low-frequency information into the low-frequency information enhancement module H, W and C respectively represent height, width and channel, firstly, SE attention mechanism self-adaption distributes weights for different characteristic channels, enhances important characteristics, suppresses unimportant characteristics, regularization is carried out by using Dropout to prevent overfitting of model training, 3X 3 convolution layer and GELU activation function act together to capture low-frequency information and improve representation capability of the model at the same time, and optimized low-frequency information is obtained , Representing the mechanism of attention of the SE, Representing a Dropout regularization, A3 x 3 convolutional layer is shown, Representation GELU activates a function and then obtains more comprehensive color information using a multi-scale color information extraction module , Representing a multi-scale color information extraction module, and finally Low frequency information enhanced by input to full connection layer ; S4, constructing a high-frequency information enhancement module which comprises a1 multiplied by 1 convolution layer, a depth convolution layer, a attention mechanism and a multi-layer perceptron, realizing the high-frequency information detail enhancement of the fishery resource statistical image, and for the high-frequency information enhancement module, firstly, carrying out high-frequency information on the fishery resource statistical image Performing layer normalization processing H, W and C represent height, width and channel respectively, Representation layer normalization processing, and the processed features are respectively obtained by a 1×1 convolution layer and a 7×7 depth convolution layer, a 1×1 convolution layer and a 5×5 depth convolution layer, and a 1×1 convolution layer and a3×3 depth convolution layer , , , Respectively representing a query, a key and a value, A1 x 1 convolutional layer is shown, A 7 x 7 depth convolution layer is shown, A 5 x 5 depth convolution layer is shown, Representing a3 x 3 depth convolution layer, then performing self-attention calculation, layer normalization processing and multi-layer perception operation on Q, K and V, calculating the weighted value of the feature, improving the attention of the model to the important feature, finally introducing residual connection to improve the training stability of the model, , The mechanism of self-attention is represented, The normalization process of the presentation layer is performed, A multi-layer sensing operation is represented, Representing element addition; S5, constructing a frequency information interaction module which comprises a1 multiplied by 1 convolution layer, a3 multiplied by 3 depth convolution layer and a softmax activation function to realize mutual reinforcement of high-low frequency information, and for the frequency information interaction module, the function is to promote information interaction between the low frequency information and the high frequency information, realize mutual reinforcement of the high-low frequency information and input the high frequency information And low frequency information H, W and C represent height, width and channel, respectively, high frequency information Obtained by a 1X 1 convolution layer , A1 x 1 convolutional layer is shown, Representation of Is to search for high frequency information Obtained by passing through a1×1 convolution layer and a 3×3 depth convolution layer , A1 x 1 convolutional layer is shown, A3 x 3 depth convolution layer is shown, Representation of Value of (2), low frequency information Obtained by a 1X 1 convolution layer , A1 x 1 convolutional layer is shown, Representation of Key of (C), low frequency information Obtained by passing through a1×1 convolution layer and a 3×3 depth convolution layer , A1 x 1 convolutional layer is shown, A3 x 3 depth convolution layer is shown, Representation of Is then to And Normalizing softmax to obtain , Representing a softmax activation function, which will ultimately And Element multiplication is carried out to obtain interacted high-frequency information , Representing element multiplication, will And Element multiplication is carried out to obtain interacted low-frequency information , Representing element multiplication; S6, processing the fishery resource statistical image, inputting the fishery resource statistical image to be processed into a color priori feature enhancement network, and obtaining the enhanced fishery resource statistical image.
2. 2. The image enhancement processing method for fishery resource statistics according to claim 1, wherein in S2, for the color a priori feature enhancement network, a fishery resource statistical image is input H, W and C respectively represent Is to be transformed using discrete wavelets Is decomposed into high-frequency information and low-frequency information, , Representing the discrete wavelet transform of the object, Low frequency information representing the input of the statistical image of the fishery resource, including content and color information of the input of the statistical image of the fishery resource, The high-frequency information representing the input fishery resource statistical image comprises detail information of global structure and texture, then the high-frequency information and the low-frequency information are input into a characteristic information enhancement module, the characteristic information enhancement module consists of a low-frequency information enhancement module, a high-frequency information enhancement module and a frequency information interaction module, and the low-frequency information is input into a characteristic information enhancement module Obtained by a low-frequency information enhancement module High frequency information Obtained by a high-frequency information enhancement module , And The input to the frequency information interaction module realizes the information interaction between the high-frequency information and the low-frequency information, , , The frequency information interaction module is represented as such, Representing the low frequency information after the interaction, Representing the interacted high frequency information, then And Adding elements to obtain , The elements are represented to be added up, And Adding elements to obtain , The elements are represented to be added up, And As the input of the next characteristic information enhancement module, the N characteristic information enhancement modules are used for obtaining And , And Generating final image by inverse discrete wavelet transform , Representing an inverse discrete wavelet transform.

Description

Image enhancement processing method for fishery resource statistics Technical Field The invention belongs to the field of image enhancement, and particularly relates to an image enhancement processing method for fishery resource statistics. Background In fishery resource management, it is important to accurately acquire the quantity, distribution and behavior characteristic information of fish populations, and due to the complexity of the underwater environment, the quality of fishery resource statistical images is greatly affected, so that problems of low contrast, high noise and color distortion exist in the underwater images, and further the accuracy of fish identification and statistics is affected, therefore, how to effectively enhance the images acquired by the underwater imaging equipment, the method is a key technology for improving the statistical accuracy of fishery resources, and the image enhancement technology mainly comprises denoising, contrast improvement, color correction and edge detection, and can recover the details and color information of images and enhance the visibility of fishes by removing scattered light effects and illumination non-uniformity in an underwater environment, so that the accuracy of subsequent fish identification, classification and quantity statistics is greatly improved, and a reliable data base is provided for the accurate evaluation of the fishery resources. In recent years, outstanding performances of diffusion-based methods in image restoration tasks are widely focused, although standard diffusion models show enough capability, unpredictable artifacts may occur due to the fact that diversity is introduced in the sampling process from random Gaussian noise to images, in addition, the diffusion models have limited capability of focusing on fine granularity information, textures and details are ignored, the previous methods are mostly based on original pixel space of images, exploration of underwater image frequency space characteristics is limited, representation capability of depth models cannot be effectively utilized to generate high-quality images, and in order to solve the problems, the invention provides an image enhancement processing method for fishery resource statistics, which is used for constructing a characteristic information enhancement module to fully explore frequency domain information in fishery resource statistical images, and simultaneously, realize color enhancement of high-frequency information and color correction of low-frequency information, and further enhance image quality. Disclosure of Invention The invention provides an image enhancement processing method for fishery resource statistics, which aims to divide an image into high-frequency information containing global structure and texture details and low-frequency information containing content and colors by using discrete wavelet transformation through constructing a color priori feature enhancement network, enhance the frequency information in a wavelet space, separate the high-frequency information from the low-frequency information, respectively realize detail enhancement and color correction, construct a frequency information interaction module, realize interaction of the high-frequency information and the low-frequency information through the module and realize image enhancement. The invention aims to provide a color priori feature enhancement network, and provides an image enhancement processing method for fishery resource statistics, which comprises the following steps. S1, collecting fishery resource statistical images, installing an underwater camera on an underwater robot, cruising and shooting in a specific water area, obtaining the images of the underwater fishery resource statistics, adding labels to the obtained fishery resource statistical images, and manufacturing a data set. S2, constructing a color priori feature enhancement network, wherein the color priori feature enhancement network is composed of N feature information enhancement modules, and the feature information enhancement modules comprise a low-frequency information enhancement module, a high-frequency information enhancement module and a frequency information interaction module. S3, constructing a low-frequency information enhancement module which comprises an SE attention mechanism, regularization, GELU activation functions, a full-connection layer and a multi-scale color information extraction module, and realizing low-frequency information color correction of the fishery resource statistical image. S4, constructing a high-frequency information enhancement module which comprises a1 multiplied by 1 convolution layer, a depth convolution layer, a attention mechanism and a multi-layer perceptron, and enhancing the high-frequency information details of the fishery resource statistical image. S5, constructing a frequency information interaction module which comprises a1 multiplied by 1 convolution layer, a