Solid waste fine identification method and system based on time sequence multi-posture

Abstract

This invention discloses a method and system for fine solid waste identification based on temporal multi-pose. The method includes steps such as initial pose perception, pose reshaping and temporal acquisition, local salient feature extraction, temporal weighted fusion, and comprehensive decision-making. This invention applies physical perturbation to the solid waste through a multi-stage flow guiding mechanism, causing it to undergo nondeterministic flipping, thereby acquiring a multi-angle temporal image sequence. It then utilizes deep learning algorithms to detect key structural components that are not easily deformable (such as bottle caps and pull rings), and performs weighted fusion of temporal features based on their visibility. This method effectively solves the common problems in solid waste identification, such as deformation interference, random occlusion, and confusion caused by similar materials, achieving accurate identification and sorting of solid waste under complex working conditions.

Description

Technical Field

[0001] This invention relates to the fields of solid waste resource utilization and computer vision technology, specifically to a method for obtaining multi-angle time-series information of solid waste through physical perturbation and using deep learning algorithms for refined identification and sorting. Background Technology

[0002] In actual waste recycling lines, waste plastic bottles (such as PET bottles), metal cans, glass bottles, and other target objects usually exhibit highly unstructured characteristics.

[0003] Existing visual recognition technologies mainly rely on single static photographs, which face the following challenges: deformation interference, as a large number of recycled bottles and cans are squeezed and deformed, resulting in the loss of their overall outline features, making them difficult for traditional algorithms to identify; random occlusion, as solid waste falls randomly on conveyor belts, key classification features (such as labels, bottle caps, and logos) are often pressed to the bottom or back of the object, making it impossible for single-view cameras to obtain effective information; and confusion due to similar materials, as transparent PET bottles with labels removed are visually extremely similar to glass bottles, making it difficult to distinguish them based on a single side view image.

[0004] While existing technologies employ multi-camera shooting solutions, they can only acquire multi-faceted information of the object in its current pose, failing to address the issue of bottom surface occlusion. Furthermore, traditional mechanical flipping devices often lack deep integration with visual algorithms, resulting in a blind flipping process and limited improvement in recognition efficiency. Therefore, there is an urgent need for a technical solution capable of actively altering the pose of solid waste, acquiring temporal multi-pose information, and effectively utilizing this information for interference-resistant recognition. Summary of the Invention

[0005] The purpose of this invention is to provide a method for precise identification of solid waste based on temporal multi-pose. This method actively exposes the hidden features of solid waste through a "pose reshaping" process, and combines a "temporal weighting" algorithm to correct key, non-deformable parts of the solid waste, thereby achieving accurate identification of solid waste under complex working conditions.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for fine identification of solid waste based on temporal multi-pose, comprising the following steps: (1) Initial pose perception: The target solid waste moves continuously on the conveyor belt, and the vision system acquires its initial pose image and performs initial identification assessment; (2) Pose reshaping and temporal acquisition: When the initial identification confidence is lower than a preset threshold, the system controls the solid waste to pass through a multi-level pose disturbance area; during the process of nondeterministic flipping of the solid waste under force, the vision system continuously captures different pose surfaces of the solid waste to construct a temporal multi-pose image sequence of the single solid waste; (3) Local salient feature extraction: Each frame of the image sequence is processed in parallel to extract the texture features of the whole image, and key structural components with rigid features on the surface of the solid waste are detected simultaneously; the key structural components specifically refer to bottle caps, bottle neck threads, or metal can pull rings in the solid waste that are not easily changed with the overall deformation; (4) Temporal weighted fusion: The extracted multi-pose features are input into the temporal fusion network; the network uses the detection results of key structural components to calculate the validity weight of each frame image, assigns high weight to pose frames containing key structural components, and assigns low weight to pose frames with blurred features or back facing the camera; (5) Comprehensive decision: Based on the temporal features after weighted fusion, the final classification result is output and the sorting mechanism is driven to perform sorting.

[0007] Furthermore, the attitude reshaping described in step (2) is achieved through a cascaded flow guiding mechanism, which includes N mechanical flow guiding units arranged sequentially along the conveying direction. The action directions of adjacent units are at an angle to each other, aiming to force the solid waste to generate a cumulative flip of about 360 degrees.

[0008] Furthermore, the time-weighted fusion described in step (4) has dual-modal logic: a significant enhancement mode is adopted when a key structural component is detected, and a statistical voting mode is automatically switched when no key structural component is detected.

[0009] Furthermore, the timing acquisition process described in step (2) introduces a target ID association mechanism: using conveyor belt encoder data and visual target tracking algorithm, images of the same solid waste target captured at different times and in different postures are spatiotemporally associated to construct a feature package of a single solid waste, preventing the image features of adjacent solid waste targets from being confused.

[0010] Furthermore, the key structural component detection in step (3) adopts a strong classification feature extraction method based on geometric descriptors: by identifying the characteristics of the bottle cap thread, the diameter of the bottle mouth, the shape of the sealing ring and the radius of curvature of the top, solid waste categories with similar materials but different top structures are distinguished.

[0011] Furthermore, the method also includes an early stopping mechanism: during the pose reshaping and temporal acquisition process, if the real-time recognition confidence exceeds the high confidence threshold after any level of perturbation, the subsequent image processing flow is immediately terminated and the process directly proceeds to step (5) to perform sorting.

[0012] The present invention also discloses a time-series multi-pose solid waste fine identification system for implementing the above method, comprising: a visual acquisition unit: including a multi-angle light source and a high-speed industrial camera, for capturing image sequences of solid waste at different flipping stages; a multi-level perturbation execution unit: located in the middle section of the conveyor belt, including a multi-level flexible flow guiding device, for applying continuous attitude reshaping torque to the solid waste; and a calculation control unit: equipped with a time-series fusion recognition algorithm and a target tracking algorithm, for processing multi-pose image sequences and generating sorting instructions.

[0013] Compared with the prior art, the present invention has the following beneficial effects: 1. Maximizing information gain: By acquiring multiple poses in a temporal sequence, a two-dimensional image from a single viewpoint is expanded into a four-dimensional spatiotemporal sequence containing information about the entire circumference of the object, effectively solving the occlusion problem.

[0014] 2. Deformation robustness: Utilizing the physical property that rigid structural components such as bottle caps and pull tabs are not easily deformed as strong classification features, combined with attention mechanisms, it solves the problem of not being able to identify bottles after they are flattened.

[0015] 3. High recognition accuracy: Even when the label is detached or severely damaged, the false detection rate and false negative rate can be significantly reduced through comprehensive discrimination under multiple postures. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 The overall process flow diagram of the method of this invention.

[0018] Figure 2 A schematic diagram of the hardware structure layout of the system of the present invention.

[0019] Figure 3 Taking plastic bottle recognition as an example, here is a block diagram illustrating the principle of the time-series fusion recognition algorithm.

[0020] Figure 4 Taking plastic bottle recognition as an example, here is a schematic diagram of temporal multi-pose image sequences and weight allocation. Detailed Implementation

[0021] The specific embodiments of the present invention will be briefly described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some, not all, embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0022] Example As attached Figure 1 and attached Figure 2 As shown in the figure, the solid waste fine identification system based on temporal multi-pose provided in this embodiment mainly includes a high-speed conveyor belt, a vision acquisition unit, a multi-level disturbance execution unit, and a computing control unit.

[0023] The specific workflow of this embodiment is as follows: Step 1: Initial attitude perception; mixed solid waste moves along the conveyor belt. A camera at the entrance captures an initial top-down view of the solid waste. If the solid waste features are clear (e.g., label side up) and the recognition confidence is high, the target is directly locked without disturbance.

[0024] Step 2: Attitude Reshaping and Timing Acquisition; When the initial identification confidence is low (e.g., bottle bottom up or severely crushed), the system activates a multi-level perturbation process. Solid waste enters the perturbation zone consisting of three sets of flexible guide levers.

[0025] The first-stage lever applies a leftward tangential force, forcing the solid waste to tip over. The second-stage lever applies a rightward tangential force, forcing the solid waste to roll back; The third-level lever applies a top frictional force, forcing the solid waste to rotate axially. Within the stable range after each lever, the camera array triggers a snapshot. The system uses a target tracking algorithm to correlate the initial image with subsequent snapshots, forming a temporal sequence containing four frames of images at different poses.

[0026] Step 3: Local salient feature extraction; the system processes these 4 images in parallel. In addition to extracting conventional full-image texture features, a specific detection algorithm is run to find "key structural components." In this embodiment, key structural components are defined as: the threaded neck / cap of a PET bottle, the pull tab / concave base of an aluminum can, and the thickened neck of a glass bottle. These parts can maintain a good geometric shape even when solid waste is violently compressed.

[0027] Step 4: Temporal weighted fusion; the system performs weighted fusion on the extracted features.

[0028] Scenario A (Successful Flip): Assume that the PET bottle cap is clearly captured in the 3rd frame, while the other 3 frames show the flattened bottle body. The algorithm will automatically assign a very high weight to the 3rd frame (e.g., 0.8), ignoring the other low-quality frames, thus confirming that the object is a PET bottle.

[0029] Scenario B (Feature Missing): Assume the solid waste is severely damaged and the bottle cap is missing; no key structural components are detected in any of the four images. The algorithm automatically switches to "statistical voting mode," combining the average prediction probabilities of the four images and using texture statistics from multiple perspectives to make the optimal judgment.

[0030] Step 5: Comprehensive decision-making; the system outputs the final category instruction, driving the downstream sorting mechanism (such as an air nozzle array) to spray the identified target into the corresponding recycling box.

[0031] In production lines processing waste plastic bottles that have been compressed and packaged and then scattered, the method described in this invention improves the accuracy of identifying flattened bottles and unlabeled bottles from 72% to over 98% compared to traditional single-view identification, effectively reducing the pressure on manual sorting at the back end.

[0032] The number and form of the disturbance mechanism described in this embodiment can be adjusted according to the actual situation of the production line. The core of this invention lies in obtaining temporal multi-pose information through physical means and using an algorithm to perform weighted processing on the temporal information based on key features. This technical concept is within the protection scope of this invention.

[0033] In the description of this invention, unless otherwise stated, "a plurality of" means two or more; the terms "upper," "lower," "left," "right," "inner," "outer," "front end," "rear end," "head," "tail," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0034] The above description of the disclosed embodiments enables those skilled in the art to make and use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit and scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for fine identification of solid waste based on temporal multi-pose, characterized in that, Includes the following steps: Step 1: Initial posture perception: The target solid waste moves continuously on the conveyor belt, and the vision system acquires its initial posture image and performs initial recognition and assessment. Step 2, Posture Reconstruction and Temporal Acquisition: When the initial recognition confidence is lower than the preset threshold, the system controls the solid waste to pass through a multi-level posture disturbance area; during the process of the solid waste undergoing nondeterministic flipping under force, the vision system continuously captures different posture surfaces of the solid waste to construct a temporal multi-posture image sequence of the single solid waste. Step 3, Local Satisfactory Feature Extraction: Each frame of the image sequence is processed in parallel to extract the texture features of the entire image and simultaneously detect key structural components with rigid features on the solid waste surface; the key structural components specifically refer to bottle caps, bottle neck threads, or metal can pull rings in the solid waste that are not easily changed by overall deformation. Step 4, Temporal Weighted Fusion: Input the extracted multi-pose features into the temporal fusion network; the network uses the detection results of key structural components to calculate the validity weight of each frame image, assigning high weight to pose frames containing key structural components and low weight to pose frames with blurred features or with the back facing the camera. Step 5, Comprehensive Decision-Making: Output the final classification result based on the weighted fusion of temporal features, and drive the sorting mechanism to perform sorting.

2. The method for fine identification of solid waste based on temporal multi-pose as described in claim 1, characterized in that, The attitude reshaping process described in step two is achieved through a cascaded flow guiding mechanism: the cascaded flow guiding mechanism includes N mechanical flow guiding units arranged sequentially along the conveying direction, where N is an integer greater than or equal to 2; the mechanical action directions of adjacent mechanical flow guiding units are designed to be at an angle to each other, forcing the solid waste passing through to generate a cumulative flipping effect of about 360 degrees during the conveying process, thereby ensuring that the key structural components or label information on the surface of the solid waste are oriented towards the visual system in at least one frame of the time-series image sequence.

3. The method for fine identification of solid waste based on temporal multi-pose as described in claim 1, characterized in that, The temporal weighted fusion described in step four employs a dual-modal logic: Significant enhancement mode: When a key structural component is detected in the temporal image sequence, the frame containing the key structural component takes precedence, suppressing background noise interference from other frames; Statistical voting mode: When no key structural components are detected in all frames of the time-series image sequence, it automatically switches to multi-frame average voting mode, using the average prediction probability from multiple angles as the final classification basis.

4. The method for fine identification of solid waste based on temporal multi-pose as described in claim 1, characterized in that, Step 2 introduces a target ID association mechanism in the time-series acquisition process: using conveyor belt encoder data and visual target tracking algorithms, images of the same solid waste target captured at different times and in different postures are spatiotemporally associated to construct a feature package for a single solid waste.

5. The method for fine identification of solid waste based on temporal multi-pose as described in claim 1, characterized in that, Step 3 describes the detection of key structural components using a strong classification feature extraction method based on geometric descriptors: by identifying the characteristics of the bottle cap thread, the diameter of the bottle mouth, the shape of the sealing ring, and the radius of curvature at the top, solid waste categories with similar materials but different top structures can be distinguished.

6. The method for fine identification of solid waste based on temporal multi-pose as described in claim 1, characterized in that, The method also includes an early stopping mechanism: during the attitude reshaping and temporal acquisition process, if the real-time recognition confidence exceeds the high confidence threshold after any level of perturbation, the subsequent image processing flow is immediately terminated and the process directly proceeds to step five for sorting.

7. A solid waste fine identification system based on temporal multi-pose, implementing the method of any one of claims 1 to 6, characterized in that, include: Visual acquisition unit: Includes multi-angle light sources and high-speed industrial cameras, used to capture image sequences of solid waste at different stages of turning; Multi-stage disturbance execution unit: Located in the middle section of the conveyor belt, it includes a multi-stage flexible flow guiding device for applying continuous attitude reshaping torque to solid waste; Computational control unit: Deploys time-series fusion recognition algorithm and target tracking algorithm to process multi-pose image sequences and generate sorting instructions.

Images