CN-121998868-A - Complex low-illumination gangue image restoration method and system integrating frequency domain separation and multi-scale enhancement

CN121998868ACN 121998868 ACN121998868 ACN 121998868ACN-121998868-A

Abstract

The invention discloses a complex low-light gangue image restoration method and system integrating frequency domain separation and multi-scale enhancement, wherein the method comprises the following steps of firstly obtaining an original gangue image in a low-light dust environment in a coal mine separation scene, secondly constructing a LFMNet restoration model containing ASID, FLDS, MMSL, thirdly transmitting the original image into a LFMNet restoration model, capturing global characteristic association through ASID lightweight, FLDS adaptively separating low-frequency illumination and high-frequency dust noise, MMSL multi-scale enhancement texture details, outputting a restoration image, fourthly outputting the restored image to a downstream gangue detection system, and outputting the obtained gangue detection result to a mechanical arm control unit, thereby realizing automatic gangue separation. The invention has the characteristics of high repair precision and light weight, can realize the triple targets of dust suppression, illumination balance and texture retention, and improves the accuracy of downstream gangue detection and separation.

Inventors

GAO ZHIJUN
ZHANG YU
CHEN KAIYUN
LIU JINGQI
WANG YANWEI
HAN LONG
ZHAO YANQIN
YIN XILIANG
LI YI
SHENG CHENYUAN

Assignees

黑龙江科技大学

Dates

Publication Date: 20260508
Application Date: 20260115

Claims (9)

1. A complex low-illumination gangue image restoration method integrating frequency domain separation and multi-scale enhancement is characterized by comprising the following steps: step one, image acquisition: adopting an image acquisition scheme of combining a 0-degree annular light source and strip light sources at two sides of the direction of a conveyor belt to acquire an original image of coal gangue in a low-illumination dust environment in a coal mine separation scene; Modeling and training an image restoration model: step two, constructing LFMNet repair models: The LFMNet repair model employs an encoder-intermediate module-decoder three-phase collaborative architecture, wherein: The encoder adopts a three-level progressive downsampling structure, realizes feature dimension increase through E_block1-E_block3, receives an input image through E_block1, serially connects attention sharing and information distillation modules ASID after being processed by a convolution layer and PReLU activation functions, captures long-distance illumination dependence through local feature preprocessing, channel segmentation, cross-block attention sharing and feature aggregation residual fusion, screens gangue texture related features, then introduces an adaptive frequency domain optical dust separation module FLDS at the output end of the E_block1, carries out global average pooling on the ASID output features, predicts a filtering intensity parameter sigma through double 1×1 convolution, converts the spatial domain features into a frequency domain through 2D-DCT transformation, generates Gaussian dynamic masks M, reserves low-frequency illumination and high-frequency gangue textures, inhibits high-frequency dust noise, and finally converts back to a spatial domain through 2D-IDCT inverse transformation to obtain dust removal purification features and serves as input of E_block2; The middle module is provided with a supervision path and a feature refining path, the supervision path generates side output to provide additional training supervision, and the feature refining path optimizes coding features through a two-stage multi-scale processing unit; The decoder adopts an up-sampling architecture combining dynamic convolution and residual connection, high-resolution image reconstruction is realized through D_block1-D_block3, light-weight multi-scale enhancement branches MMSL are integrated, branch output characteristics are spliced through channels, then the branch output characteristics are input into an SPPF serial pooling module through a 1X 1 convolution compression channel to realize light-weight global aggregation, and finally the integrated data are fused with input characteristic residual errors, so that characteristic degradation is avoided; Training the constructed LFMNet repair model; step three, image restoration reasoning: transmitting the original image acquired in the first step into a LFMNet restoration model which is completed by training, capturing global feature association through ASID lightweight, FLDS adaptively separating low-frequency illumination and high-frequency dust noise, MMSL multi-scale enhancement of texture details, and outputting restoration images with balanced brightness, no dust interference and clear textures; Step four, outputting and sorting results: and outputting the repaired image to a downstream gangue detection system for gangue detection, outputting the obtained gangue detection result to a mechanical arm control unit, realizing automatic gangue sorting, and providing high-fidelity visual input for target identification.
2. The complex low-light gangue image restoration method integrating frequency domain separation and multi-scale enhancement according to claim 1, wherein the specific steps of the first step are as follows: Arranging strip-shaped light sources on two sides of the conveyor belt along the direction of the conveyor belt; step one, installing an industrial-level area array camera above a gangue sorting belt, and installing a 0-degree annular light source below a lens; Step one, acquiring real-time videos of coal and gangue on a conveyor belt by using a camera, and transmitting the videos to a server; and step four, screening, labeling and data enhancement processing are carried out on the acquired images, and finally a dataset for LFMNet repair model training is obtained.
3. The complex low-illumination coal gangue image restoration method integrating frequency domain separation and multi-scale enhancement according to claim 2, wherein in the step one, the strip-shaped light source irradiates with an inclination angle of 30-60 degrees.
4. The complex low-light gangue image restoration method integrating frequency domain separation and multi-scale enhancement according to claim 2, wherein in the step one, the illumination intensity of the acquisition environment is 1-5 lux, and the dust density is 5-40%.
5. The method for restoring the complex low-light gangue image by fusing frequency domain separation and multi-scale enhancement according to claim 1, wherein in the first step, MMSL comprises three differential branches, namely a small-scale branch is used for focusing local dust and micro textures, a middle-scale branch is used for balancing regional illumination gradual change, and a large-scale branch is used for covering global illumination distribution.
6. The complex low-light gangue image restoration method integrating frequency domain separation and multi-scale enhancement according to claim 1, wherein the specific steps of the second step are as follows: (1) Collecting low-illumination dust coal gangue images of a real scene and a simulated scene of a coal mine, marking normal illumination dust-free truth images after screening, and manufacturing a tag file; (2) Dividing the marked image data into a training set, a verification set and a test set, and carrying out data enhancement processing on the divided data; (3) Inputting the training set and the verification set into LFMNet repair models, setting training parameters and training, verifying on the verification set after each Epoch training is completed, and recording and storing optimal model weights; (4) And detecting the images in the test set by using the optimal model weight, and checking LFMNet the effect of repairing the model.
7. The method for restoring complex low-light gangue images by fusing frequency domain separation and multi-scale enhancement according to claim 6, wherein in the step (3), training parameters are set to be based on PyTorch frames, an Adam optimizer is adopted, initial learning rate is 1e-4, batch size is 8, total training round is 100, and a combined loss function of 0.5×L1 loss+0.3×SSIM loss+0.2×frequency domain loss is adopted.
8. The complex low-light gangue image restoration method integrating frequency domain separation and multi-scale enhancement according to claim 1, wherein the specific steps of the fourth step are as follows: Inputting LFMNet repaired image sequences into Yolov n, yolov10n or Yolov n series detectors, outputting position marking information and category confidence coefficient of the gangue, and transmitting the position marking information and category confidence coefficient to a mechanical arm and a mechanical arm in real time; And step four, the mechanical arm performs accurate sorting operation based on the position marking information of the gangue output by the step four, and realizes automatic separation of coal and gangue.
9. A gangue image restoration system for implementing the complex low-light gangue image restoration method integrating frequency domain separation and multi-scale enhancement as claimed in any one of claims 1-8, characterized in that the system comprises an image input module, a gangue LFMNet restoration module, and a result output and sorting module, wherein: the image input module adopts an industrial-level area array camera and is used for acquiring real-time images or video frames of low-illumination dust coal gangue on the coal mine sorting conveyor belt and transmitting the real-time images or video frames to the end-side computing unit; The gangue LFMNet restoration module is used for restoring the real-time image or video frame acquired by the image acquisition module by using the trained LFMNet restoration model; the result output and sorting module is used for transmitting the repaired image to the downstream detection model, receiving the detection result and forwarding the detection result to the mechanical arm control unit to drive automatic sorting.

Description

Complex low-illumination gangue image restoration method and system integrating frequency domain separation and multi-scale enhancement Technical Field The invention belongs to the technical field of image enhancement and industrial intelligent separation, relates to a coal gangue image processing method in a low-light dust environment, and particularly relates to a complex low-light coal gangue image restoration method and system integrating frequency domain separation and multi-scale enhancement. Background The intelligent separation of coal is a key direction of industry transformation, and the coal gangue separation link is limited by complex environments such as underground low illumination (illumination intensity is less than 5 lux), high dust and the like of a coal mine, so that the acquired coal gangue images have the problems of uneven brightness, fuzzy textures, confusion of dust and gangue characteristics and the like. The gray level difference between the coal and the gangue is covered, the identification accuracy of the traditional separation equipment is suddenly reduced to below 60%, and the intelligent separation efficiency is seriously affected. The existing low-illumination image enhancement method can solve the problem of partial brightness, but is difficult to adapt to special scenes of 'low illumination and dust interference' of a coal mine, firstly, the scene adaptation is insufficient, the general method cannot distinguish high-frequency dust noise from gangue textures, so that the enhanced textures are lost or dust residues, secondly, the performance and efficiency are unbalanced, the deep learning model has the problems that the parameter quantity is too large (more than 10M), the reasoning speed is insufficient by 10 FPS, the deep learning model cannot be deployed on embedded end-side equipment of the coal mine, thirdly, the multi-scale characteristics are not utilized enough, the overall illumination adjustment and the local texture retention are difficult to balance, and local overexposure or detail blurring easily occurs. The defects cause that the downstream gangue detection task is missed in detection and has high false detection rate, so that the floor application of the intelligent coal mine separation system is restricted. Disclosure of Invention Aiming at the problems of confusion of light dust characteristics, texture loss, difficult model deployment and the like faced by gangue image restoration in a low-light dust environment, the invention provides a complex low-light gangue image restoration method and system integrating frequency domain separation and multi-scale enhancement, which have the characteristics of high restoration precision and light weight, can realize triple targets of dust suppression, illumination balance and texture retention, adapt to the deployment requirements of coal mine end-side equipment, improve the accuracy of downstream gangue detection and separation, are suitable for gangue image optimization in low-light and high-dust interference in a coal mine separation scene, and provide high-fidelity visual input for downstream gangue detection and separation. The invention aims at realizing the following technical scheme: a complex low-illumination gangue image restoration method integrating frequency domain separation and multi-scale enhancement comprises the following steps: step one, image acquisition: The method comprises the following steps of performing image acquisition by adopting an image acquisition scheme of combining a 0-degree annular light source and strip-shaped light sources on two sides of a conveyor belt direction, and acquiring an original image of coal gangue in a low-illumination dust environment in a coal mine separation scene, wherein the specific steps are as follows: Arranging strip-shaped light sources on two sides of the conveyor belt along the direction of the conveyor belt; step one, installing an industrial-level area array camera above a gangue sorting belt, and installing a 0-degree annular light source below a lens; Step one, acquiring real-time videos of coal and gangue on a conveyor belt by using a camera, and transmitting the videos to a server; Step four, screening, labeling and data enhancement processing are carried out on the acquired images, and finally a dataset for LFMNet repair model training is obtained; Modeling and training an image restoration model: step two, constructing LFMNet repair models: The LFMNet repair model adopts a three-stage collaborative architecture of an encoder-middle module-decoder, and the core comprises three special modules, namely an attention sharing and information distillation module (ASID), a self-adaptive frequency domain optical dust separation module (FLDS) and a light-weight multi-scale enhancement branch (MMSL), wherein the ASID is used for optimizing long-distance illumination dependency modeling, FLDS is used for realizing the accurate separation of low-frequency illu