A multi-target lightweight visual real-time recognition system in a complex environment

Abstract

The application discloses a multi-target lightweight visual real-time identification system in a complex environment, which comprises a field image acquisition module for continuously acquiring original image data by shooting a field picture of a coke oven slag removal operation area in real time; an environment interference adaptive correction module for pre-processing the collected original image in view of the high dust, uneven light and picture blur interference problems of the coke oven operation scene; and a simplified and optimized convolutional neural network structure; the application has the beneficial effects that the device investment cost and deployment threshold of the coke oven field intelligent transformation are effectively reduced, the application of the coke oven slag removal equipment of various types of large, medium and small sizes is adapted, various types of environmental noises on site can be effectively filtered out, and the details of the degraded image are repaired; the defects of the prior art, such as target feature loss, recognition missed detection and false detection, in a complex interference scene are solved, and the accuracy, stability and anti-interference ability of multi-target identification under complex working conditions are greatly improved.

Description

Technical Field

[0001] This invention belongs to the field of visual recognition technology, specifically relating to a lightweight real-time visual recognition system for multiple targets in complex environments. Background Technology

[0002] The environment of intelligent slag removal operation in coke ovens is complex, with common problems such as heavy dust, uneven lighting, overlapping and obscuring of coal slag targets, and complicated background interference. Accurate and real-time identification of multiple targets such as coal slag and impurities in the furnace is the core prerequisite for ensuring the stable and efficient operation of automated slag removal.

[0003] Currently, there are mature related technologies in the field of intelligent control of coke oven slag removal. Among them, the intelligent control system and method for coke oven slag removal with embodied perception and multimodal fusion, which has patent publication number CN122194681A, has an overall architecture that is divided into a multimodal perception layer, a perception decision layer, a cleaning execution layer, a cleaning control layer, and a communication network layer. The multimodal perception layer relies on the embodied perception system to collect multimodal information on the coke oven environment and slag status and complete the preprocessing to generate a multimodal feature matrix.

[0004] However, the existing technology has high requirements for edge device hardware configuration, high deployment costs, and weak adaptability, making it difficult to adapt to the deployment scenarios of lightweight devices. At the same time, the data processing process of this technology is cumbersome and the calculation is time-consuming. It has a high latency in recognizing multiple types of targets that are dynamically changing in coke ovens, which cannot meet the real-time control requirements of high-speed slag cleaning operations. Furthermore, it is difficult to balance the recognition accuracy and real-time performance under complex interference scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a lightweight real-time visual recognition system for multiple targets in complex environments, so as to solve the problems of target missed detection and false detection.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a lightweight real-time visual recognition system for multiple targets in complex environments, comprising... A field image acquisition module used to capture real-time footage of the coke oven slag removal operation area and continuously acquire raw image data. To address the issues of high dust levels, uneven lighting, and blurred images in coke oven operation scenarios, an adaptive environmental interference correction module is implemented for the preprocessing of the acquired raw images. A lightweight feature extraction module is developed that uses a simplified and optimized convolutional neural network structure to quickly extract multi-dimensional target features from preprocessed image data. Based on the extracted target features, the multi-target accurate identification and judgment module distinguishes between coal slag and debris targets in the coke oven and simultaneously completes target localization and classification judgment. A real-time output interaction module that receives identification and positioning results in real time and outputs information with low latency.

[0007] As a preferred technical solution of the present invention, raw image data containing coal slag, furnace impurities, and obstructing debris is continuously acquired.

[0008] As a preferred technical solution of the present invention, in order to address the interference problems of high dust, uneven lighting, and blurred images in coke oven operation scenarios, noise reduction, brightening, and defogging preprocessing are performed on the acquired original images.

[0009] As a preferred technical solution of the present invention, based on the extracted target features, coal slag and debris targets of different sizes, shapes and overlapping and obstructing each other in the coke oven are distinguished.

[0010] As a preferred technical solution of the present invention, the identification and positioning results are received in real time, and the target location, type, and distribution status information are output with low latency.

[0011] As a preferred technical solution of the present invention, the lightweight feature extraction module is implemented as follows: The image of the coke oven site after noise reduction, defogging, and light correction is subjected to uniform size adaptation processing; the pixel ratio of the image is finely adjusted to match the input size of the simplified convolutional network; image features are extracted layer by layer using multi-layer lightweight convolutional units, and the entire coke oven image is scanned step by step using small-sized convolutional kernels; after each layer of convolution calculation is completed, feature dimensionality reduction is performed simultaneously, automatically filtering out duplicate, overlapping, and meaningless feature data; the multi-class effective information extracted and filtered layer by layer is integrated into a simplified feature data set.

[0012] As a preferred technical solution of the present invention, the multiple types of effective information include the morphology of coal slag, target location, debris characteristics, and characteristics of the obstructed area.

[0013] Compared with the prior art, the beneficial effects of the present invention are: Effectively reduces equipment investment costs and deployment thresholds for on-site intelligent transformation of coke ovens, and is compatible with the application of slag removal equipment for large, medium and small coke ovens; It can effectively filter out various environmental noises on site and restore degraded image details; It solves the shortcomings of existing technologies that are prone to target feature loss, missed detection, and false detection in complex interference scenarios, and greatly improves the accuracy, stability and anti-interference ability of multi-target recognition under complex working conditions; Employing a lightweight and rapid computing process, it significantly reduces the time required for image feature extraction and target recognition, resulting in lower overall recognition latency and faster response speed. It can output effective information such as the location, type, shape, and distribution status of targets inside the furnace in real time and continuously, accurately matching the high-speed operation and real-time control requirements of automated coke oven slag cleaning equipment. It provides stable and reliable real-time data support for parameter adjustment, intelligent operation, and safety management of the slag cleaning mechanism, effectively improving the automation level, operational efficiency, and on-site safety of coke oven slag cleaning operations. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the multi-target lightweight real-time visual recognition system under complex environments according to the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Please see Figure 1 This invention provides a lightweight real-time visual recognition system for multiple targets in complex environments, including... The on-site image acquisition module is used to capture real-time images of the coke oven slag removal operation area and continuously acquire raw image data containing coal slag, furnace impurities, and obstructing debris. The environmental interference adaptive correction module communicates with the on-site image acquisition module. In response to interference problems such as high dust, uneven lighting, and blurry images in the coke oven operation scene, the module performs noise reduction, brightening, and defogging preprocessing on the acquired raw images, eliminating invalid image information caused by environmental interference and retaining clear and effective target image features. The lightweight feature extraction module adopts a simplified and optimized convolutional neural network structure to quickly extract multi-dimensional target features from preprocessed image data, while significantly reducing the model's computational load and power consumption, thus adapting to the lightweight deployment requirements of edge terminal devices. The multi-target accurate identification and judgment module is connected to the lightweight feature extraction module. Based on the extracted target features, it distinguishes between coal slag and debris targets of different sizes, shapes and overlapping and obscuring each other in the coke oven, and simultaneously completes target positioning and classification judgment, effectively avoiding the problem of missed detection and false detection of multiple targets in complex environments. The real-time output interaction module is connected to the multi-target accurate identification and judgment module and the coke oven slag cleaning control terminal. It receives the identification and positioning results in real time and outputs the target location, type and distribution status information with low latency.

[0017] Furthermore, the lightweight feature extraction module is implemented as follows: After noise reduction, defogging, and light correction, the coke oven site image is subjected to uniform size adaptation processing; complex deformation calculations involving image enlargement and reduction are unnecessary, only the pixel ratio of the image is finely adjusted to unify it to the input size of the simplified convolutional network, avoiding redundant calculations due to inconsistent image specifications and reducing initial computational power consumption; the entire coke oven image is scanned step-by-step using small-sized convolutional kernels. Shallow convolutional units mainly capture basic visual features, including the outline boundaries of coal slag inside the furnace, surface texture, and the shape outline of debris; mid-level convolutional units integrate basic features, refining the size and shape of coal slag, occlusion edges, and distinguishing features between debris and the furnace wall; deep convolutional units further filter effective features, eliminating invalid interference features such as dust residue and light and shadow noise, retaining only those usable for target recognition. Other core effective features; after each layer of convolution calculation is completed, feature dimensionality reduction is performed simultaneously, automatically filtering out duplicate, overlapping, and meaningless feature data, and not retaining redundant feature parameters; at the same time, the large number of parameter iteration calculation processes in traditional networks are eliminated, the core feature calculation weights are fixed, and there is no need to repeatedly correct parameters, which greatly reduces the computational workload of the equipment; the effective information such as coal slag morphology, target location, debris features, and occlusion area features after layered extraction and screening is integrated into a simplified feature data set, which is directly output to the subsequent multi-target accurate identification and judgment module; the entire extraction process eliminates redundant calculation steps, and while ensuring the completeness of multi-target feature extraction in coke ovens and without losing key information, it greatly improves the feature extraction speed, achieves a lightweight feature extraction effect with low computing power and high efficiency, and meets the core requirements of real-time identification in coke oven slag cleaning operations.

[0018] Although embodiments of the invention have been shown and described (see the detailed description above), it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A lightweight real-time visual recognition system for multiple targets in complex environments, characterized in that: include A field image acquisition module used to capture real-time footage of the coke oven slag removal operation area and continuously acquire raw image data. To address the issues of high dust levels, uneven lighting, and blurred images in coke oven operation scenarios, an adaptive environmental interference correction module is implemented for the preprocessing of the acquired raw images. A lightweight feature extraction module is developed that uses a simplified and optimized convolutional neural network structure to quickly extract multi-dimensional target features from preprocessed image data. Based on the extracted target features, the multi-target accurate identification and judgment module distinguishes between coal slag and debris targets in the coke oven and simultaneously completes target localization and classification judgment. A real-time output interaction module that receives identification and positioning results in real time and outputs information with low latency.

2. The lightweight real-time visual recognition system for multiple targets in complex environments according to claim 1, characterized in that: Continuously acquire raw image data containing coal slag, furnace impurities, and obstructing debris.

3. The lightweight real-time visual recognition system for multiple targets in complex environments according to claim 1, characterized in that: To address the interference issues of high dust levels, uneven lighting, and blurred images in coke oven operation scenarios, noise reduction, brightening, and defogging preprocessing are performed on the acquired raw images.

4. The lightweight real-time visual recognition system for multiple targets in complex environments according to claim 1, characterized in that: Based on the extracted target features, we can distinguish between coal slag and debris targets of different sizes, shapes, and overlapping and obscuring objects inside the coke oven.

5. The lightweight real-time visual recognition system for multiple targets in complex environments according to claim 1, characterized in that: It receives identification and positioning results in real time and outputs target location, type, and distribution status information with low latency.

6. The lightweight real-time visual recognition system for multiple targets in complex environments according to claim 1, characterized in that: The lightweight feature extraction module is implemented as follows: After noise reduction, defogging, and light correction, the coke oven site image is subjected to uniform size adaptation processing; the pixel ratio of the image is finely adjusted to match the input size of the simplified convolutional network; multi-layer lightweight convolutional units are used to extract image features layer by layer, scanning the entire coke oven image step by step using small-sized convolutional kernels; after each layer of convolution calculation is completed, feature dimensionality reduction is performed simultaneously, automatically filtering out duplicate, overlapping, and meaningless feature data; the multi-class effective information extracted and filtered layer by layer is integrated into a simplified feature data set.

7. A lightweight real-time visual recognition system for multiple targets in complex environments according to claim 6, characterized in that: Multiple types of effective information include coal slag morphology, target location, debris characteristics, and characteristics of obstructed areas.

Images