Method and device for the analysis of carcasses or parts thereof, method for training at least one neural network for the analysis of carcasses or parts thereof

Abstract

The present invention relates to a method (100) for the analysis of carcasses (11) comprising the steps of conveying (101) the carcasses (11), detecting (102) the carcasses (11) by means of a first detection unit (14), generating (103) numerical data, processing (104) the numerical data by means of a computer unit (15), wherein characteristic attributes in a predetermined region are selected (105), or wherein the numerical data are compared (106) by means of a computer unit (15) with numerical database data contained in a database (16) in order to identify (107) characteristic attributes, providing (108) the characteristic attributes for display (109) and / or transmission (110) to at least one downstream machine (20) for assessment (111) or processing (112) of the carcasses (11) or carcass regions on the basis of the characteristic attributes determined. The invention also relates to a method (200) for training at least one neural network (17, 18), to a non-volatile computer-readable storage medium and also to a corresponding device (10) for the analysis of carcasses (11).

Description

[0001] Method and device for the analysis of carcasses or parts thereof, method for training at least one neural network for the analysis of carcasses or parts thereof

[0002] The invention relates to a method for analyzing carcasses or parts thereof, in particular poultry carcasses or fish carcasses.

[0003] Furthermore, the invention relates to a method for training at least one neural network for the analysis of carcasses or parts thereof, in particular poultry carcasses or fish carcasses.

[0004] Furthermore, the invention relates to a device for analyzing carcasses or parts thereof, in particular poultry carcasses or fish carcasses.

[0005] Such methods and devices are used in the automated processing of carcasses, particularly poultry or fish carcasses. Devices and methods for processing carcasses have been known for many years and are used primarily for the efficient mechanical processing of such food products, ensuring quick, clean, and reliable processing. In recent years, increased demands have been placed on the mechanical processing of carcasses and the resulting product recovery, which are to be implemented primarily through the use of more precise, reliable, and faster processes and corresponding devices.In addition, automatic and automated systems, such as robot-based processing equipment, are increasingly being used in processing stations, whereby coordination between the individual processing steps and device components must be ensured. Furthermore, comprehensive information is required during transport and processing of the carcasses in order to carry out processing at the right time and / or in the intended location, so that processing-related quality and thus value reductions can be avoided. Particularly due to the increased transport speeds and the increased accuracy requirements in the use of processing stations, errors such as inaccuracies or delays can occur during the processing of the carcasses, which can lead to rejects or time-consuming and costly reprocessing.

[0006] Furthermore, due to a shortage of skilled workers and rising personnel costs, reduced personnel deployment for such processes or the use of corresponding equipment is required or desired, so that the dependency on personnel when using corresponding processes and equipment should be reduced. Furthermore, a standard for evaluating corresponding carcasses / products is desired that is not based on subjective criteria. In order for automated processing of carcasses to be possible at all, appropriate carcass detection must be available. In principle, mechanical devices and methods are known in particular for detecting the presence of carcasses on a conveyor system, e.g., to record the dimensions or position of the carcasses.However, in order to subsequently record carcasses in such a way that their condition can be recorded or further processed, it is necessary to obtain / generate the specific condition, position, or other carcass-specific information in order to be able to carry out the characteristic processing by suitable personnel or specialized machines. Of particular interest are the positions and / or orientation of the carcasses, as well as any changes or conditions present in the carcasses.

[0007] A further disadvantage of the state of the art is that the known methods and devices only record the carcass at a specific time or at a specific location in the system, without any exchange of the respective recorded information. This places particular demands on the machine software, which must be individually adapted for each application and does not allow complete control. Furthermore, the evaluation of corresponding recordings or the machine control based on them requires highly qualified specialist personnel or pre-defined algorithms, making such recordings costly and maintenance-intensive. If multiple carcass recordings are planned in one device or method, interaction between these recordings is not easily possible because they are not available on a uniform basis or in a common format.To consolidate this data, an interface solution must be available that uniquely identifies the recorded carcasses. However, such solutions are generally only used in conjunction with carcass tracking throughout an entire facility, tracking individual carcasses throughout the entire processing period. This requires complex computational operations, and the resulting information cannot be used for other purposes, as they are only designed to identify or re-identify a single carcass. Furthermore, such carcass tracking options are prone to errors, as small changes have major impacts on the entire system. Mechanical and human monitoring are always required.This leads to increased demands on personnel and resources and the associated costs, in addition to the disadvantages resulting from the impossibility of objective and consistent monitoring.

[0008] The known devices and methods are generally designed to perform specific operations, e.g., removing or processing certain parts of a carcass. Specific processing of carcasses is only possible within narrow limits with the available methods and devices, e.g., with special cutting devices where an area to be removed has been previously selected manually or automatically. Integration into a comprehensive system or a comprehensive process is not feasible with the available options, as this type of information cannot be recorded and no interaction or communication takes place between the individual process steps, meaning this information is neither of interest nor usable.

[0009] Furthermore, the known devices or methods do not allow the handling of or changes to the carcass during processing. In particular, it is not possible to determine at what point in time, for example, a change in the carcass structure or quality occurred in order to trace this back to the cause. Nor can it be traced at what point in time or on which component of the device proper or improper processing took place in order to make the resulting processing quality visible. The monitoring and intervention options for the known methods and devices are therefore limited to an assessment of the respective actual condition, without providing any concrete information about where, in what manner or to what extent a carcass exhibited a corresponding condition and what the cause of the respective condition is.A current, ongoing recording of the respective status during a process or during processing in a device cannot be provided with the known means.

[0010] It is therefore an object of the present invention to provide corresponding methods and devices which enable an analysis of carcasses or parts thereof in order to carry out an evaluation of the recorded information and at the same time to recommend a resulting further processing of the carcasses.

[0011] The object is achieved by the method for analyzing carcasses mentioned at the outset, comprising the steps of conveying the carcasses in a conveying direction by means of a conveyor device, detecting the carcasses or a carcass area by means of a first detection unit for each of the carcasses or the carcass area that is guided past the first detection unit, generating numerical data, in particular digital images, of the carcasses or carcass areas detected by the first detection unit by means of the first detection unit and / or by means of a computer unit, processing the numerical data of the detected carcasses or carcass areas by means of a computer unit, wherein the carcasses or carcass areas are selected for the extraction of characteristic expressions in a predetermined region,or wherein the numerical data of the recorded carcasses or carcass areas are compared by means of a computer unit with numerical database data contained in a database, in particular digital database images, in order to identify characteristic expressions in a predetermined region of the carcasses or carcass areas, providing the characteristic expressions for display and / or forwarding to at least one downstream machine for evaluating or processing the carcasses or carcass areas on the basis of the determined characteristic expressions.,

[0012] The method according to the invention offers the advantage that numerical data are generated from the recorded carcasses, which can be used for specific analysis. The disclosed method enables an assessment / analysis of the performance of a device without being directly connected to the device or the carcass. The information on the condition of the carcass is available as numerical data and can be used and further processed individually and specifically, for example by focusing on individual sub-areas. The performance of the method is evaluated by processing numerical data generated by the recording unit. In this way, a change or a processing step on the carcass can be represented by the data measured with the at least one recording unit.Carcasses can include poultry, fish, pork or beef or parts thereof.

[0013] After the numerical data has been processed by the computing unit, characteristic characteristics can be extracted within a predefined range in order to carry out subsequent actions taking the characteristic characteristics into account. The characteristic characteristics can be reliably determined using the recording unit, taking into account the characteristic characteristics to be extracted, in particular within the corresponding predefined range. For this purpose, at least one recording unit is selected accordingly in order to determine the respective characteristic characteristics based on the numerical data. The provided characteristic characteristics thus represent corresponding information, in particular data, for display or further processing in a downstream machine for evaluating or processing the carcasses, in particular in the predefined areas with the characteristic characteristics.This makes it possible to create and establish an assessment or planned processing of the carcasses based on the determined trait characteristics. This creates a correlation between the carcasses recorded by the recording unit, the numerical data generated from them, and the extraction of the trait characteristics to form an action based on them. The corresponding database images can include, for example, a visible spectrum or data based on UV, IR, X-ray, or 3D data.

[0014] The term "conveying device" within the meaning of the invention refers to all conveying devices designed and configured to convey carcasses, at least in certain areas. The conveying devices preferably have at least one conveying element for conveying the carcasses. The conveying device in connection with the method according to the invention preferably has a single conveying element designed and configured as an endless belt and can alternatively also be referred to as a conveyor belt, conveyor belt, belt conveyor, transport belt, etc. As conveying devices for poultry carcasses or parts thereof, conveyor elements designed as so-called "shackles" are preferably used, in which the legs of the poultry carcasses or parts thereof can be arranged accordingly. The respective conveying element can in turn consist of several or a plurality of components.The material of the transport element is not relevant to the invention, but in preferred embodiments it can consist of metal, plastic, rubber or a combination thereof.

[0015] In principle, any detection unit that enables the detection of carcasses, carcass areas, or parts thereof can be provided as a "detection unit." For this purpose, cameras, sensors, X-ray machines, photoelectric devices, etc., can be used to detect the carcasses and subsequently image them as numerical data. The numerical data is preferably provided or generated in the visible spectral range and / or on the basis of UV, IR, X-ray, or 3D data by means of the respective at least one detection unit. The detection unit is preferably designed and configured as an imaging detection unit. The detection unit is not limited to consisting of a single component, but can preferably consist of several detection elements.The detection unit is preferably provided for detecting / determining the geometry, color information, and / or texture properties of the respective carcass or carcass area. The detection unit is further preferably arranged in a stationary manner in the area of ​​the conveyor device. Further preferably, several detection units are provided at different positions to form a multi-part detection of the carcasses. In preferred embodiments, the computer unit can already be integrated into the detection unit to process the numerical data. In further advantageous embodiments, the computer unit is a separate component to perform corresponding computing operations.

[0016] "Characteristic characteristics" generally include any characteristics of the numerical data of the recorded carcasses that enable extraction based on predefined or trained characteristics using the computer unit. A characteristic characteristic can, for example, be an optically detectable characteristic such as a notch, a bulge, a thickening, a discoloration, a recess, a graphic design (code), a point, etc. The characteristic characteristics are selected from a predetermined region of the numerical data, i.e., the carcass. The predetermined region can also be the entire carcass, in order to provide or support the recording / detection of the characteristic characteristics. However, the predetermined region is preferably a sub-area of ​​the numerical data or the carcass in order to select the characteristic characteristics.

[0017] The "downstream machine" within the meaning of the invention can be any machine or station that designs and sets up an assessment or processing of the carcasses. Such an assessment can, for example, be a qualitative or quantitative classification. Processing can, in particular, involve cutting, dividing, separating, etc., of the carcasses or parts thereof using suitable means. However, an assessment or processing of the carcasses can also include a further processing step, e.g., reloading, further transport, gripping, weighing, sorting, etc. Such machines are also referred to as processing stations, cutting devices, evisceration devices, separating devices, distribution devices, transfer devices, etc. The machines comprise at least one means that can be brought into engagement and / or interaction with the carcasses.

[0018] A useful embodiment of the invention is characterized in that the numerical data are provided as input data to a first, in particular pre-trained, neural network, wherein the first neural network has been trained to localize and / or evaluate characteristic expressions in a predetermined region of the carcass or carcass areas. This has the advantage that a comprehensive evaluation or learning of the localization of characteristic expressions in a predetermined region of the carcass only needs to take place during the training or learning phase. After the neural network has been trained, a localization and / or evaluation of characteristic expressions in a predetermined region can be generated based on the generated or recorded numerical data.Overall, the use of a pre-trained neural network enables precise and efficient processing of numerical carcass data, resulting in accelerated and reliable carcass analysis. This approach contributes to reducing labor, improving the accuracy of results, and overall increasing the quality of feature localization and evaluation.

[0019] A further expedient embodiment of the invention is characterized in that the numerical data are provided as input data to the first neural network and / or a second neural network, wherein the first neural network and / or the second neural network is provided for the local detection and / or segmentation of the carcass, in particular of predetermined regions of the carcass or carcass areas. Preferably, more than two neural networks are provided to carry out corresponding detection and / or segmentation of carcasses. Due to this combination, efficient and accurate processing of the numerical data in various contexts is possible. The interaction of one and / or the two neural networks leads to improved local analysis and segmentation of the carcass, which further increases the accuracy and reliability of the findings.In particular, this makes it possible to record only individual sub-areas of the respective carcass in isolation or to use them segment by segment to identify characteristic characteristics. By using at least two neural networks, the analysis can be broken down further, whereby one or more neural networks can break down a specific object, i.e. in particular a carcass, into sub-areas, while one or more further neural networks can carry out analysis tasks for relevant sub-areas of the object. The individual results of each neural network can then be combined into an overall analysis, in particular a qualitative assessment of the processing quality. This enables a multi-layered assessment of the carcasses to be analyzed and integration into more complex systems, devices or machines in order to provide continuous assessment / analysis of the processing.

[0020] According to a further preferred embodiment of the invention, the predetermined region is selected depending on the carcass or carcass area and / or depending on the design and / or positioning of the detection unit. This enables a customized adaptation of the region to be analyzed depending on individual characteristics of the carcass or specific areas of the carcass. This customized selection enables greater analytical accuracy and efficiency, since the regions to be examined can be specifically defined or selected based on needs, particularly depending on the respective characteristic characteristics. Furthermore, by taking into account the detection unit, its design and / or positioning, optimized recording and acquisition of corresponding data is ensured.Thus, this alternative embodiment increases the flexibility, accuracy and applicability of the entire analysis procedure for carcasses or carcass areas.

[0021] In a preferred development, the method comprises at least one further detection unit, wherein the at least one further detection unit is arranged downstream of the first detection unit in the conveying direction in order to detect each of the carcasses or carcass regions conveyed past the at least one further detection unit, in order to generate, by means of the at least one further detection unit and / or by means of the computer unit, further numerical data, in particular further digital images, of the carcasses or carcass regions detected by the at least one further detection unit. This allows each carcass or each region of a carcass that passes at least one further detection unit to also be detected.This leads to the generation of further numerical data, in particular additional digital images, of the recorded carcasses or carcass areas for appropriate processing. The use of this additional recording unit in conjunction with the computer unit opens up the possibility of creating a broader database and obtaining more detailed information about the carcasses or carcass areas to be analyzed, particularly in the period between the first recording unit and the at least one additional recording unit.

[0022] A further preferred development of the invention is characterized in that the further numerical data generated by the at least one further recording unit are further processed by the computer unit, wherein the carcasses or carcass areas are selected for the extraction of feature characteristics in a predetermined region, or wherein the further numerical data of the carcasses or carcass areas recorded by the at least one further recording unit are compared by the computer unit with numerical database data contained in a database, in particular digital database images, in order to identify feature characteristics, which are recorded in particular by the at least one pre-trained neural network, in a predetermined region of the carcasses or carcass areas, or wherein the further numerical data, in particular the feature characteristics,the carcasses or carcass areas recorded by the at least one further recording unit are compared by means of the computer unit with the numerical data, in particular the characteristic values, of the first recording unit, wherein the characteristic values ​​are provided for display and / or forwarding to a downstream machine for evaluating or processing the carcasses or carcass areas based on the determined characteristic values. In this way, with the present embodiment of the method according to the invention, for example, the time periods between two consecutive recording units on the same conveyor device (e.g. conveyor belt or conveyor chain) can be used to compare the data from recording units of the same carcass or parts of the same carcass, e.g. before and after a machine intended for processing.to compare. In this way, for example, the processing quality and performance of the machine can be monitored by comparing the result of the application of neural networks to the information from the detection units that are recorded before and after the specific machine on the conveyor device, e.g., the conveyor chain, the conveyor belt, etc. Furthermore, it is advantageous if all recorded values ​​of the at least one detection unit are fed into the respective neural network. A further expedient embodiment of the invention is characterized in that at least one of the detection units comprises at least one sensor element for detecting carcasses or carcass areas with numerical data, in particular digital images. By using suitable sensor elements for detecting numerical data, in particular digital images, a detection unit is provided,which ensures the needs-based recording of carcasses. Recording carcasses using the numerical data to be recorded enables more efficient analysis, accelerates the process, and reduces the analytical effort. The objective evaluation of carcasses based on this sensor data eliminates subjective factors and enables rational and consistent decision-making. This is particularly important for improved quality control, production planning, and resource allocation. Such flexibility allows adaptation to specific requirements, while simultaneously recording the quality of the carcasses and specifically improving it through appropriate machining, further reducing waste. This can give users a competitive advantage. In addition, precise analysis contributes to improving food safety.by detecting potential health risks or undesirable components of carcasses at an early stage and, if necessary, removing them or further processing them mechanically, e.g., cutting them up. Preferably, technologies and systems with multiple sensors are used to generate performance and quality-related data from carcasses, which can be used in particular to evaluate each individual processing step in order to react early to potential defects and, at the same time, maximize the quality of the carcasses to be analyzed and reduce any type of waste. The at least one sensor element can be, but is not limited to, monochromatic or color-measuring sensors, such as optical image processing systems, radiation-measuring sensors, such as X-ray detectors, area- or distance-based measuring sensors, such as depth sensors, mechanical systems,that respond to direct physical contact in a specific spatial position (distance), or temperature sensors, which are implemented, for example, as imaging devices with infrared sensors. The data is preferably processed accordingly by the at least one computer unit, for example, to determine the machine performance by comparing the respective sensor information, or the numerical data generated thereby, before and after a machine that performs a processing step. Accordingly, several sensor elements or detection units are required for this purpose.

[0023] According to a further preferred embodiment, the characteristic characteristics are recorded and compared with reference characteristics of known and defined anomalies stored in the database and / or with the characteristic characteristics determined by the first recording unit or at least one further recording unit. This enables precise identification of deviations and irregularities in the corresponding carcasses or characteristic characteristics. This is particularly useful for reliable quality control and efficient defect detection, as well as for possible classification or intended further processing. Linking the recorded characteristics with the reference data contributes to data-supported analysis and enables targeted further development of the processes. Reference characteristics are preferably characteristics of known damaged areas on carcasses, which are stored in the database.

[0024] An advantageous further development is characterized by the fact that, based on the comparison of the feature values, in particular using the first neural network, an anomaly status of the feature values ​​is determined. In this way, it is possible to determine the anomaly status of feature values ​​as precisely as possible. This is achieved through a detailed comparison of the feature values, whereby the first neural network is used in particular. This approach enables the reliable identification of deviations and anomalies that might have been overlooked using conventional methods or could only be carried out with considerable personnel effort. The use of a neural network brings increased accuracy, consistency, and sensitivity to the analysis process. This leads to an increased ability to detect even subtle deviations from the reference features or the determined feature values.This improves quality assurance because potential defects or unusual patterns can be reliably detected. The data-based processing by the first neural network also enables continuous improvement over time. The system can learn from previous analyses and adapt to new findings and changes. This leads to constant optimization of anomaly detection and supports the continuous refinement of the analysis and production processes. A further advantage lies in the efficiency and speed of the neural network used, which can process large amounts of data in near real time to determine anomaly status within the shortest possible time. This minimizes potential delays and enables rapid response to deviations, resulting in the most just-in-time processing of carcasses possible.In addition, this method applies the same level of analysis to each object on the conveyor, which is not consistently possible when using human inspectors due to movement and attention constraints. With the method according to the invention, it is not necessarily necessary to focus on only one defect, which would lead to neglecting the other objects. With the known methods, this results in objects located behind the defective inspection object being inspected less intensively. In addition, the method according to the invention offers holistic evaluation / reporting for data acquisition and process optimization.

[0025] A useful embodiment of the invention is characterized in that an action is executed depending on the characteristic characteristics and / or the anomaly status, wherein the action particularly comprises the automatic removal of the carcass or carcass area from the conveyor system. This enables the automatic implementation of suitable measures depending on the detected characteristic characteristics and / or the anomaly status. This is particularly advantageous for the meat processing industry, where contamination or quality defects can lead to far-reaching subsequent problems. The integration of automatic action mechanisms raises quality assurance to a new level of performance. Potential deviations or anomalies can be detected immediately, and specifications can be implemented precisely and promptly.This not only helps prevent quality problems and waste, but also increases overall production efficiency. Furthermore, targeted rejection can avoid unnecessary work steps, thus reducing the energy consumption of the overall process. The automatic rejection of carcasses or carcass sections from the conveyor system, in particular, represents progress. Defective or abnormal products can be immediately isolated and separated from further processing before they can impact the production process. This leads to a significant reduction in waste and contributes to cost savings. Another advantage is the system's real-time response to characteristic variations and anomalies. This enables rapid correction and adjustment, minimizing production interruptions and maintaining high production capacity.This is especially important in an industry where efficiency and on-time delivery are of utmost importance.

[0026] In a preferred refinement of the method, the action includes the automatic sorting of the carcasses or carcass areas into several categories based on the anomaly status and / or the characteristic expressions. Automatic sorting based on anomaly status and characteristic expressions primarily contributes to a significant improvement in quality control and process control. Products can be classified in real time and with high accuracy, allowing, for example, defective or deviating carcasses to be identified and separated early and / or processing of these areas by downstream machines to be carried out or omitted. This leads to a reduction in quality problems and waste.The ability to sort carcasses into several specific categories allows companies using the inventive method to optimize their production and meet the needs of different markets and customers. This contributes to the flexibility and adaptability of production processes and makes it possible to offer customized products that meet diverse requirements. The automation of this sorting process contributes to increasing overall efficiency. By eliminating manual sorting steps, labor is reduced and production capacity is increased. This helps optimize production flow and minimize bottlenecks. Furthermore, automated sorting enables precise data collection and analysis.The continuous acquisition of information on the anomaly status and the trait expressions enables continuous improvement in classification accuracy. The system can learn from past analyses and adapt to changes, leading to continuous optimization of the sorting accuracy and the available data of the neural network. A further preferred development of the invention is characterized in that the action includes an assessment of the quality of the carcasses or carcass areas based on the recorded trait expressions, the anomalies, and / or the anomaly status. The integration of this quality assessment enables an accurate and objective assessment of the quality of the carcasses. Traditionally subjective assessments are replaced by a data-driven approach.This helps minimize human errors and differences in interpretation and leads to consistent and reliable quality assessment. Comprehensive analysis allows potential quality problems to be identified and addressed early. Irregularities, deviations, or anomalies are accurately recorded and evaluated, leading to improved quality assurance. Quality assessment based on the recorded characteristic values ​​and anomalies enables data-driven optimization of production processes. Targeted improvements can be derived from the information obtained to further increase product quality. This contributes to continuous improvement and adaptability to changing market requirements.

[0027] A useful embodiment of the invention is characterized in that at least one further detection unit is used to detect further feature characteristics, such as geometry, color information, and / or texture properties of the carcass or carcass area, in order to perform an additional assessment of the anomaly status using at least one further neural network. This advanced approach enables a more comprehensive analysis and assessment of the anomaly status, which brings with it a number of significant advantages. The integration of further feature characteristics such as geometry, color information, and texture properties enables a more precise and detailed characterization of carcasses or carcass areas. This leads to an improved ability to detect and assess anomalies, especially those that may have been overlooked using previous techniques.The use of at least one additional neural network for the additional assessment of the anomaly status increases the accuracy and sensitivity of the analysis method. This at least one neural network can, for example, be specifically trained on the recorded characteristics in order to recognize complex relationships and patterns. This not only makes the quality assessment even more precise and reliable, but also allows the at least one additional neural network to be used for a specific purpose in the analysis or assessment. The more comprehensive analysis allows a broader spectrum of quality characteristics to be considered and assessed. The combination of different characteristic characteristics and neural networks opens up new possibilities for error detection and prevention and offers greater assessment depth for assessing carcasses or carcass cuts at different depths or stages or for subsequent further processing.The use of at least one additional detection unit and at least one additional neural network increases the adaptability of the system. The system can be tailored to different carcasses or carcass areas and is capable of adapting to changing production conditions.

[0028] A further preferred embodiment of the invention is characterized in that the recorded feature values ​​are processed or evaluated by a pre-trained machine learning model. The machine learning model is preferably at least one neural network. The use of a pre-trained machine learning model enables a particularly efficient and precise analysis of the recorded feature values. The model has already learned / been provided with a large amount of data and can use this knowledge to recognize patterns and relationships that are crucial for quality assessment. This leads to increased evaluation speed and improved error detection efficiency.The use of a pre-trained model reduces the use of time and resources, as the complex training process can be prepared separately and specifically structured using the pre-trained learning model. This accelerates implementation in the method according to the invention. Furthermore, the pre-trained model is already prepared for a variety of feature values, which increases its adaptability. The use of a pre-trained machine learning model contributes to the consistency and objectivity of the method, as the evaluation of the respective numerical data is again based on empirical data.

[0029] In a preferred embodiment of the invention, the method is carried out essentially in real time to enable continuous monitoring of the conveyance of the carcasses or carcass areas. This leads to a reaction to the resulting analysis or corresponding recommendations for action that is as delay-free as possible. The term "real time" in the sense of the invention includes certain technically induced delays, which, however, essentially enable direct intervention in the carcasses or carcass areas due to the continuous monitoring. The preferred implementation of the method in real time preferably also includes batch processing, in that batch recording is carried out first, although the data is recorded essentially in real time.

[0030] The object is also achieved by the method mentioned above for training at least one neural network, comprising the following steps: providing a plurality of carcasses or parts thereof, in particular poultry carcasses or fish carcasses, detecting the carcasses or a carcass area by means of a first detection unit, generating numerical data, in particular digital images, of the carcasses or carcass areas detected by the first detection unit by means of the first detection unit and / or by means of a computer unit, determining at least one reference value for identifying the positions and / or the variation of at least one characteristic expression in a predetermined region of the carcass or carcass area based on the numerical data,Entering the numerical data as input data and the at least one reference value as target data as training data for the neural network, repeatedly adjusting the weights of the neural network based on the difference between the target data and the output data generated by the neural network. The method according to the invention offers the advantage that the characteristic characteristics are recorded specifically during the training or learning phase. After the neural network has been trained using the method according to the invention, the respective characteristic characteristics of the carcasses or carcass areas can be determined based on the recording of the numerical data by means of the at least one recording unit. For this purpose, the numerical data and the target data are compared accordingly. It is therefore possible,Based on the difference between the target data and the output data generated by the neural network, to identify the positions and / or the variation of at least one characteristic expression in a predetermined region of the carcass or carcass area in order to classify the characteristic expression. In this way, a correlation is created between the reference values ​​and the target values ​​in order to provide an output as a training model. Preferably, at least one neural network, particularly preferably several neural networks, is pre-trained to extract precise information that can be used to evaluate the current quality of the carcass or areas of the carcass based on the data acquired by the at least one acquisition unit, wherein they are also supplied with information (target data) about,How the result (reference value) of a specific trait expression or, preferably, a processing step is to be evaluated. All information extraction tasks can be performed at the lowest level of the products to be processed, e.g., in the case of poultry processing, at the level of a flock or an individual chicken, down to a specific product such as a poultry wing, breast, or leg, or even parts thereof (e.g., feet, neck).

[0031] To avoid repetition, in connection with the method according to the invention for training at least one neural network, particular reference is made to the advantages already explained in detail in connection with the method according to the invention for analyzing carcasses or parts thereof. This also applies analogously to the method according to the invention for training at least one neural network described below.

[0032] In a further advantageous embodiment of the invention, the reference values ​​are sorted into known or defined anomalies, whereby an anomaly status of the feature values ​​is created and established. Sorting the reference values ​​into known or defined anomalies enables targeted identification and evaluation of quality deviations and thus targeted training of the corresponding neural network. By establishing an anomaly status, deviations from the norm can be detected quickly and precisely, leading to improved application. The ability to sort reference values ​​into known or defined anomalies also facilitates the interpretation of the recorded feature values. Clear categories of anomalies are created, which enable and accelerate targeted evaluation.Finally, the creation and implementation of an anomaly status for the characteristic values ​​allows for continuous monitoring and analysis of deviations. Regularly updating the anomaly status allows the system to react to changing production conditions, adapt to new quality characteristics, and develop further learning processes.

[0033] A useful embodiment of the invention is characterized in that, after the carcass or carcass area has been recorded by the first recording unit, a modification of the carcass or carcass area is carried out in the range of the at least one reference value. The targeted modification in the range of the reference value makes it possible to react directly to anomalies or deviations, particularly in the recorded characteristic characteristics. The specific modification of the carcass or carcass area makes it possible to identify potential conditions of the carcass at an early stage and influence them accordingly, before they can, for example, affect the production process. The data obtained during recording serves as the basis for the modification and thus enables targeted and precise adaptation. This contributes to greater efficiency and accuracy of the method.The system can respond flexibly to new challenges and ensures that the manufactured products always meet the required standards. By modifying the reference value, the system can respond to new requirements. Deviations from the norm can be immediately detected and corrected, leading to a significant reduction in scrap and losses.

[0034] A further preferred embodiment of the invention is characterized in that the carcasses or carcass regions are sorted, evaluated, and / or designated for further processing based on the detected characteristic expressions, anomalies, and / or anomaly status. This offers the advantage that the carcasses or carcass regions do not need to be pre-sorted with regard to characteristic expressions, anomalies, and / or anomaly status. Regardless of the respective conditions of the carcasses or carcass regions, appropriate sorting or classification can thus be carried out fully automatically.A further expedient embodiment of the invention is characterized by detecting the carcass or carcass region after the modification by means of a further detection unit in order to generate, by means of the at least one further detection unit and / or the computer unit, further numerical data, in particular further digital images, of the carcasses or carcass regions detected by the at least one further detection unit. As a result, each carcass or each region of a carcass that passes at least one further detection unit can also be detected. This leads to the generation of further numerical data, in particular additional digital images, of the detected carcasses or carcass regions in order to utilize them accordingly.The use of this additional recording unit in conjunction with the computer unit opens up the possibility of creating a broader database and obtaining more detailed information on the carcasses or carcass areas to be analysed, in particular in the period between the first recording unit and the at least one additional recording unit.

[0035] A preferred development of the invention is characterized by establishing a modification classification for evaluating the modification carried out, in particular the qualitative modification, in the range of the at least one reference value based on the further numerical data from the further recording unit. Establishing a change classification offers a clear and objective way of evaluating the effects of the modification carried out using the method according to the invention. By including additional numerical data, a comprehensive basis for the evaluation is created, thereby minimizing human factors and subjective assessments. The use of the further numerical data from the at least one further recording unit allows a comprehensive analysis of the modification, in particular for comparing any processing carried out on the carcasses.The targeted determination of a modification classification based on additional numerical data contributes to objective traceability. The training model can learn from past modifications and further develop its evaluation capabilities. This leads to continuous optimization of the modification processes and product quality. A useful embodiment of the invention is characterized by inputting the additional numerical data from the additional acquisition unit as additional input data and the modification classification as additional target data in order to perform a comparison, in particular a qualitative comparison, between the implemented modification classification and the implemented modification, and repeatedly adjusting the weights of the neural network based on the difference between the additional target data and the output data generated by the neural network.The neural network is preferably a multi-layer neural network with a corresponding number of hidden layers. The weights are preferably adjusted during training using the stochastic gradient descent method. The loss function used is, for example, the mean square error, the cross entropy, or the Jaccard coefficient (intersection over union) calculated from the difference between the target data and the output data generated by the neural network. This allows for a detailed and comprehensive evaluation of the modification. The neural network uses the additional data to recognize patterns and correlations and to perform a qualitative comparison between expected and actual results.The structure of such neural networks and the adjustment of the neural network's weights based on the error between the desired output data and the target data are well known, so further details will be omitted here. The possibility of a qualitative comparison between the modification classification and the implemented modification offers a high level of quality assurance and error detection. Potential deviations or inconsistencies can thus be quickly identified. The repetitive adjustment of the neural network's weights based on the difference between the further target data and the generated output data enables continuous improvement of the evaluation capabilities. The neural network learns from past adjustments and optimizes itself over time.

[0036] An advantageous further development is characterized in that the at least one reference value is at least one value that references the position or qualitative nature of the characteristic characteristics and / or the anomaly status. The assignment of reference values ​​to specific positions or qualities enables detailed analysis and evaluation, leading to increased accuracy. The targeted referencing of positions or qualitative properties of the characteristic characteristics and / or the anomaly status contributes to the creation of a comprehensive reference system. This allows deviations from the norm to be precisely identified and evaluated, which in particular leads to improved error detection and evaluation, which in turn enables simplified further measures, e.g., rejection or machine processing.The inclusion of reference values ​​in the quality assessment process promotes the uniformity and objectivity of the assessment. The use of standardized reference values ​​minimizes human uncertainty and differences in interpretation and leads to reliable and traceable results.

[0037] A further advantageous embodiment of the invention is characterized in that the at least one reference value for the qualitative quality is based on database data stored in the neural network for classifying the qualitative assessment of the quality of the characteristic characteristics and / or the anomaly status. The use of database-supported reference values ​​enables a comprehensive and objective classification of the qualitative quality. The neural network draws on a large number of data, which increases the accuracy and reliability of the assessment. The integration of database information for qualitative classification enables adaptations and updates to changing quality requirements and thus supports flexible and prospective quality control.The use of stored data accelerates automated quality assessment by allowing the neural network to efficiently compare the collected information with reference values ​​from the database. This enables a rapid qualitative assessment of feature values ​​and / or anomaly status.

[0038] The object is also achieved by the aforementioned non-volatile computer-readable storage medium, comprising a program that includes instructions for causing the computer to execute the inventive method for training at least one neural network. All common storage types are suitable as storage media, for example, CD-ROMs, DVDs, memory sticks, hard drives, or cloud storage services.

[0039] In addition, the task is also achieved by the device mentioned above for

[0040] Analysis of carcasses or parts thereof, comprising a conveyor device configured to convey the carcasses or parts thereof in a conveying direction, a first detection unit designed and configured to detect the carcasses or a carcass area, for each of the carcasses or the carcass area that can be guided past the first detection unit, wherein by means of the first detection unit and / or by means of a computer unit, numerical data, in particular digital images, of the carcasses or carcass areas detected by the first detection unit can be generated, a computer unit provided for processing the numerical data, wherein the carcasses or carcass areas can be selected for the extraction of characteristic expressions in a predetermined region, or wherein the numerical data of the detected carcasses or carcass areas can be processed by means of a computer unitare comparable with numerical database data contained in a database, in particular digital database images, in order to identify characteristic characteristics in a predetermined region of the carcasses or carcass areas, at least one input unit configured to sequentially provide the numerical data as input data to the computer unit, wherein the computer unit is configured to provide the data of the characteristic characteristics for display and / or forwarding to at least one downstream machine for evaluating or processing the carcasses or carcass areas on the basis of the determined characteristic characteristics. The device described here enables efficient analysis of carcasses or parts of carcasses. By combining individual components such as a conveyor device, recording unit and computer unit, numerical data including digital images can be generated quickly and precisely.The first acquisition unit is specifically designed for the acquisition of carcasses or carcass areas. This associated automated data acquisition ensures reliable and repeatable identification of trait expressions without relying on manual processes. The device enables the precise extraction of trait expressions in predefined regions of carcasses or carcass areas. This ensures precise identification and quantification of trait expressions relevant for evaluation or processing. The acquired numerical data of the carcasses or carcass areas can be compared with existing numerical database data, particularly digital database images, to form a reliable identification of trait expressions, which facilitates evaluation based on already known standards. The input unit enables theSequential provision of numerical data as input data for the computer unit. This process ensures orderly and efficient processing of the data to identify the trait characteristics. The provided trait characteristic data can be passed on to downstream machines for evaluating or processing the carcasses or carcass areas. This enables seamless integration into existing production or processing processes. Overall, significant time and resource savings can be achieved compared to conventional manual methods. This increases the efficiency and productivity of the entire processing plant. The presented device enables consistent and objective quality control of the carcasses or carcass areas. This contributes to ensuring consistently high product quality.

[0041] To avoid repetition, in connection with the device according to the invention for analyzing carcasses or parts thereof, particular reference is made to the advantages already described in detail in connection with the method according to the invention for analyzing carcasses or parts thereof, as well as to the advantages already described in detail in connection with the method according to the invention for training at least one neural network. These also apply analogously to the device according to the invention for analyzing carcasses or parts thereof.

[0042] In a further advantageous embodiment of the invention, the numerical data can be provided as input data to a first neural network comprising the computer unit, wherein the first neural network has been trained to localize and / or evaluate feature characteristics in a predetermined region of the carcasses or carcass areas.

[0043] An expedient embodiment of the invention is characterized in that the numerical data can be provided as input data to the first neural network comprising the computer unit and / or to a second neural network comprising the computer unit, wherein the first neural network and / or the second neural network is provided for the local detection and / or segmentation of the carcass, in particular of predetermined regions of the carcass or carcass areas.

[0044] A further preferred embodiment of the invention is characterized in that the device further comprises at least one further detection unit, wherein the at least one further detection unit is arranged downstream of the first detection unit in the conveying direction in order to detect each of the carcasses or carcass regions guided past the at least one further detection unit in order to generate, by means of the at least one further detection unit and / or by means of the computer unit, further numerical data, in particular further digital images, of the carcasses or carcass regions detected by the at least one further detection unit.

[0045] A further expedient embodiment of the invention is characterized in that the further numerical data generated by means of the at least one further recording unit can be further processed by means of the computer unit, wherein the carcasses or carcass areas can be selected for the extraction of feature characteristics in a predetermined region, or wherein the further numerical data of the carcasses or carcass areas recorded by the at least one further recording unit can be compared by means of the computer unit with numerical database data contained in a database, in particular digital database images, in order to identify feature characteristics in a predetermined region of the carcasses or carcass areas, or wherein the further numerical data, in particular the feature characteristics,the carcasses or carcass areas recorded by the at least one further recording unit are comparable by means of the computer unit with the numerical data, in particular the characteristic characteristics, of the first recording unit, whereby the characteristic characteristics can be made available for display and / or forwarding to a downstream machine for the evaluation or processing of the carcasses or carcass areas on the basis of the determined characteristic characteristics.

[0046] A preferred development of the invention is characterized in that at least one of the detection units comprises at least one sensor element for detecting carcasses or carcass areas with numerical data, in particular digital images.

[0047] An expedient embodiment of the invention is characterized in that the device further comprises at least one discharge device which is designed and configured to discharge carcasses or carcass regions from the conveyor device depending on the characteristic characteristics and / or an anomaly status.

[0048] An advantageous further development is characterized in that the device comprises at least one further detection unit which is designed and configured to detect further characteristic characteristics, such as geometry, color information and / or texture properties of the carcass or carcass area.

[0049] Furthermore, the object is also achieved by a further non-volatile computer-readable storage medium, comprising a program which includes instructions for causing the computer to carry out the method according to the invention for analyzing carcasses or parts thereof, in particular poultry carcasses or fish carcasses.

[0050] Further preferred and / or expedient features and embodiments of the invention emerge from the dependent claims and the description. Particularly preferred embodiments are explained in more detail with reference to the accompanying drawings. The drawing shows:

[0051] Fig. 1 is a first block diagram of a first embodiment of the method according to the invention for analyzing carcasses or parts thereof,

[0052] Fig. 2 is a second block diagram of a second embodiment of the method according to the invention for analyzing carcasses or parts thereof, - TI -

[0053] Fig. 3 is a simplified further block diagram of an embodiment of the method according to the invention for training at least one neural network,

[0054] Fig. 4a is a schematic view of the device according to the invention for analyzing carcasses or parts thereof in a first embodiment,

[0055] Fig. 4b is a schematic view of the device according to the invention shown in Fig. 4a with a further detection unit,

[0056] Fig. 5a is a schematic view of the device according to the invention for analyzing carcasses or parts thereof in a second embodiment,

[0057] Fig. 5b is a schematic view of the device according to the invention shown in Fig. 5a with a further detection unit and

[0058] Fig. 5c is a schematic view of the device according to the invention shown in Fig. 5b with a machine for processing the carcasses or parts thereof.

[0059] The method according to the invention, the storage medium according to the invention and the device according to the invention are explained in more detail below with reference to the above-mentioned figures.

[0060] The device 10 shown in the drawings is designed and configured for analyzing carcasses 11 or parts thereof, in particular poultry or parts thereof. For simplicity, only carcasses 11 are referred to overall, but corresponding parts thereof or individual regions are also intended to be included. The methods 100, 200 according to the invention are explained in particular by the corresponding diagrams and graphs in the above-mentioned drawings, each of which already contains some preferred method steps.

[0061] The method 100 according to the invention for analyzing carcasses 11 or parts thereof, in particular poultry carcasses or fish carcasses, is schematically illustrated in the diagrams of Fig. 1 and Fig. 2. The method 100 comprises the steps of conveying 101 the carcasses 11 in a conveying direction 12 by means of a conveying device 13. A conveying device 13 is shown schematically in Fig. 4a, Fig. 4b and in Fig. 5a to Fig. 5c. The conveying device 13 is intended to convey the carcasses 11 at least in sections during the method 100. This is followed by a detection 102 of the carcasses 11 or a carcass region by means of a first detection unit 14, for each of the carcasses 11 or the carcass region that are guided past the first detection unit 14.The next step is the generation 103 of numerical data, in particular digital images, of the carcasses 11 or carcass regions recorded by the first recording unit 14 by means of the first recording unit 14 and / or by means of a computer unit 15. The integration and positioning of the computer unit 15 during the method or within the system is not primarily relevant to the invention. For example, several computer units 15 can be provided or integrated into other components, in particular the recording unit 14, in order to perform individual computing operations separately from one another. After the generation 103 of the numerical data, the numerical data of the recorded carcasses 11 or carcass regions are processed 104 by means of a computer unit 15, wherein the carcasses 11 or carcass regions are selected for the extraction of characteristic characteristics in a predetermined region 105.Optionally, the numerical data of the recorded carcasses 11 or carcass areas are compared 106 by means of a computer unit 15 with numerical database data, in particular digital database images, contained in a database 16 in order to identify 107 characteristic expressions in a predetermined region of the carcasses 11 or carcass areas. This is followed by providing 108 the characteristic expressions for display 109 and / or forwarding 110 to at least one downstream machine 20 for evaluation 111 or processing 112 of the carcasses 11 or carcass areas on the basis of the determined characteristic expressions.Preferably, the numerical data are provided as input data to a first, in particular pre-trained, neural network 17 (not shown in detail in the figures), wherein the first neural network 17 has been trained to localize and / or evaluate feature characteristics in a predetermined region of the carcasses 11 or carcass areas. The structure of such neural networks 17 is sufficiently known, so that further explanations are omitted here. However, the numerical data are preferably provided as input data in the first neural network 17 and / or a second neural network 18, wherein the first neural network 17 and / or the second neural network 18 are provided for the local detection and / or segmentation of the carcass 11, in particular of predetermined regions 105 of the carcass 11 or carcass areas. In Fig.Figure 2 shows a further preferred embodiment of the method 100 according to the invention for analyzing carcasses 11 or parts thereof, in which a further neural network 18 is provided. Within the method 100, the second neural network 18 is configured analogously to the first neural network 17, but, for example, different regions or carcass areas are evaluated accordingly. In preferred embodiments, the predetermined region 105 is selected depending on the carcass 11 or the carcass area and / or depending on the design and / or positioning 113 of the detection unit 14.

[0062] According to a further advantageous embodiment, the method 100 according to the invention further comprises at least one further detection unit 19, 21, wherein the at least one further detection unit 19, 21 is arranged downstream of the first detection unit 14 in the conveying direction 12 in order to detect each of the carcasses 11 or carcass regions guided past the at least one further detection unit 19, 21 in order to generate, by means of the at least one further detection unit 19, 21 and / or by means of the computer unit 15, further numerical data, in particular further digital images, of the carcasses 11 or carcass regions detected by the at least one further detection unit 19, 21. In Fig. 2 and Fig. 3, a second detection unit 19 and, by way of example, an n-th detection unit 21 are shown. The n-th further detection unit 21 is shown in Fig. 2 and Fig.3 is shown by way of example to show that, depending on the configuration, more than two or three acquisition units 14, 19, 21 can be provided. The numerical data generated by the additional acquisition units 19, 21 can be further processed using a single-stage analysis method (cf. Fig. 1), in particular using the first neural network 17, or using a two-stage analysis method (cf. Fig. 2 and Fig. 3), in particular using at least one additional neural network 18.

[0063] In the embodiment of the method 100 with at least one further recording unit 19, 21, the generated additional numerical data are preferably further processed by the computer unit 15, wherein the carcasses 11 or carcass areas are selected for the extraction of feature characteristics in a predetermined region 105. Optionally, the additional numerical data of the carcasses 11 or carcass areas recorded by the at least one additional recording unit 19, 21 are compared by the computer unit 15 with numerical database data contained in a database 16, in particular digital database images, in order to identify feature characteristics in a predetermined region 105 of the carcasses 11 or carcass areas.In a further alternative embodiment, the further numerical data, in particular the characteristic characteristics, of the carcasses 11 or carcass regions recorded by the at least one further recording unit 19, 21 are compared by means of the computer unit 15 with the numerical data, in particular the characteristic characteristics, of the first recording unit 14. Finally, the characteristic characteristics are provided for display and / or forwarding to a downstream machine for evaluating or processing the carcasses 11 or carcass regions based on the determined characteristic characteristics.

[0064] According to a further advantageous embodiment of the invention, at least one of the detection units 14, 19, 21 comprises at least one sensor element 22 for detecting carcasses 11 or carcass regions with numerical data, in particular digital images. As also schematically shown in Fig. 1 and Fig. 2, at least one of the detection units 14, 19, 21 more preferably comprises at least one sensor element 22 selected from sensors for detecting / generating a visible spectrum 23, a UV spectrum 24, an IR spectrum 25, depth / 3D information 26, or X-ray information. Preferably, the feature characteristics are detected and compared with reference features of known and defined anomalies stored in the database 16 and / or with the feature characteristics determined by the first detection unit 14 or at least one further detection unit 19, 21.For this purpose, in a preferred embodiment, an anomaly status of the feature characteristics is determined based on the comparison 106 of the feature characteristics, in particular using the first neural network 17. The process of determining the anomaly status is particularly integrated into or downstream of the method step of processing 112 the numerical data of the detected carcasses 11 or carcass regions by means of the computer unit 15. In further embodiments, more in-depth or more differentiated comparisons of the feature characteristics can be performed with the aid of the second neural network 18 or with additional neural networks in order to determine a corresponding anomaly status.Preferably, an action 114 is carried out depending on the characteristic characteristics and / or the anomaly status, wherein the action 114 particularly comprises the automatic discharge 115 of the carcass 11 or carcass region from the conveyor device 13. In a further preferred embodiment of the invention, the action 114 comprises the automatic sorting 116 of the carcasses 11 or the carcass regions into several categories based on the anomaly status and / or the characteristic characteristics. Further preferably, the action 114 comprises an assessment of the quality of the carcasses 11 or carcass regions based on the detected characteristic characteristics, the anomalies, and / or the anomaly status—not shown in detail in the diagrams of Fig. 1 and Fig. 2.The preferably multiple sensor elements 22, the first detection unit 14, and / or the at least one further detection unit 19, 21 are preferably used to detect further characteristic characteristics, such as geometry, color information, and / or texture properties of the carcass 11 or carcass region, in order to perform an additional assessment of the anomaly status using at least one further neural network 18. Preferably, the detected characteristic characteristics are processed or evaluated by a pre-trained machine learning model. Further preferably, the method 100 is carried out essentially in real time to enable continuous monitoring of the conveyance of the carcasses 11 or carcass regions.The inventive method 200 for training at least one neural network 17, 18 for the analysis of carcasses 11 or parts thereof, in particular poultry carcasses or fish carcasses, is described by way of example with reference to Fig. 3. The method 200 comprises the steps of providing 201 a plurality of carcasses 11 or parts thereof, in particular poultry carcasses or fish carcasses. The carcasses are preferably conveyed in a conveying direction 12 by means of a conveyor device 13. The method 200 further comprises detecting 202 the carcasses or a carcass region by means of a first detection unit 14. Generating 203 numerical data, in particular digital images, of the carcasses 11 or carcass regions detected by the first detection unit 14 by means of the first detection unit 14 and / or by means of a computer unit 15.This is followed by the determination 204 of at least one reference value 205 for identifying the positions and / or the variation of at least one characteristic expression in a predetermined region of the carcass 11 or carcass area based on the numerical data. Input 206 of the numerical data as input data 207 and of the at least one reference value 205 as target data as training data for the neural network 17, 18, and repeated adjustment 208 of the weights of the neural network 17, 18 based on the difference between the target data and the output data 209 generated by the neural network 17, 18. The determination 204 of at least one reference value 205 is carried out in particular by evaluating the numerical data from the at least one acquisition unit 14, 19, 21 by dividing it into the partial data relevant for the processing step by a human training person.The reference values ​​205 are then preferably used to identify and / or vary at least one characteristic expression in a predetermined region of the carcass based on the numerical data. The reference values ​​205 are accordingly used as training data for training the neural network 17, 18. Further preferably, the reference values ​​205 are sorted into known or defined anomalies, whereby an anomaly status of the characteristic expressions is formed and established. Preferably, the carcasses 11 or carcass regions are sorted, evaluated, and / or provided for further processing based on the detected characteristic expressions, anomalies, and / or the anomaly status.In an advantageous embodiment of the method 200 for training at least one neural network 17, 18 for the analysis of carcasses 11, after the detection 202 of the carcass 11 or the carcass region by the first detection unit 14, a modification 210 of the carcass 11 or the carcass region is carried out in the range of the at least one reference value 205. The modification 210 carried out provides for a change in state and a corresponding training of this change in state on the carcasses 11.In a further advantageous embodiment, the carcasses 11 or the carcass region are recorded 202 after the modification 210 by means of at least one further recording unit 19, 21 in order to generate 203, by means of the at least one further recording unit 19, 21 and / or by means of the computer unit, further numerical data, in particular further digital images, of the carcasses 11 or carcass regions recorded by the at least one further recording unit 19, 21. Fig. 3 shows three recording units 14, 19, 21 as an example, whereby training can already be implemented with just one recording unit 14. If there is more than one recording unit 14, 19, 21, the numerical data of carcasses 11 preferably supplement the qualitative modification 210 by (processing) machines 20 between the recording units 14, 19, 21.

[0065] In a further preferred embodiment, the method 200 according to the invention comprises determining a modification classification—not shown in detail in the figures—for evaluating the performed modification 210, in particular the qualitative modification, in the range of the at least one reference value 205 based on the further numerical data of the further acquisition unit 19, 21. Preferably, the method 200 comprises the further steps of inputting the further numerical data of the further acquisition unit 19, 21 as further input data 207 and modification classification as further target data in order to perform a comparison, in particular a qualitative comparison, between the performed modification classification and the performed modification 210, and repeatedly adjusting 208 the weights of the neural network 17, 18 based on the difference between the further target data and the output data 209 generated by the neural network 17, 18.

[0066] To train the at least one neural network 17, 18, a learning cycle control unit is preferably provided to repeatedly adjust the weights of the at least one neural network 17, 18 based on the difference between the target data and the output data 209 generated by the at least one neural network 17, 18. The design of the corresponding neural network 17, 18 can be diverse. In principle, all multi-layer networks are contemplated. A network structure with 29 layers has proven particularly advantageous in terms of recognition accuracy while simultaneously maintaining a reasonable algorithmic effort.

[0067] The at least one neural network 17, 18 is preferably a multi-layer neural network with a corresponding number of hidden layers. The weights are preferably adjusted during training using the stochastic gradient descent method. For example, the mean square error from the difference between the target data and the output data 209 generated by the at least one neural network 17, 18 is used as the loss function. The structure of such neural networks and the adjustment of the weights of the neural network based on the error between the desired output data and the target data is well known, so further explanations will also be omitted here.

[0068] Preferably, the at least one reference value 205 is at least one value that references the position or the qualitative nature of the feature characteristics and / or the anomaly status. Further preferably, the at least one reference value 205 for the qualitative nature is referenced based on database data stored in the neural network 17, 18 for classifying the qualitative assessment of the nature of the feature characteristics and / or the anomaly status.

[0069] The present invention also relates to a non-volatile computer-readable storage medium with a program comprising instructions for causing the computer to execute the above-described method for training the neural network. All common storage types are suitable as the storage medium, for example, CD-ROMs, DVDs, memory sticks, hard drives, or cloud storage services.

[0070] The present invention also encompasses a device 10 for analyzing carcasses 11 or parts thereof, in particular poultry carcasses or fish carcasses. The device 10 according to the invention is explained in more detail below with reference to Fig. 4a, Fig. 4b and Fig. 5a to Fig. 5c. The device 10 comprises a conveyor device 13 configured to convey the carcasses 11 or parts thereof in a conveying direction 12, as well as at least one detection unit 14, 19, 21 designed and configured to detect the carcasses 11 or a carcass region, for each of the carcasses or the carcass region that can be guided past the at least one detection unit 14, 19, 21. The conveyor device 13 of Fig. 4a and Fig. 4b is preferably designed as a hanging conveyor, in particular for the hanging conveyance of poultry carcasses, while the conveyor device 13 of Fig. 5a to Fig.5c are designed as belt conveyors, which are provided in particular for the horizontal conveyance of carcasses 11. Preferably, the device 10 according to the invention, as shown in Figs. 4a and 5a, comprises a first detection unit 14, wherein numerical data, in particular digital images, of the carcasses 11 or carcass regions detected by the first detection unit 14 can be generated by means of the first detection unit 14 and / or by means of a computer unit 15. Furthermore, a computer unit 15 is provided for processing the numerical data. According to the device 10 according to the invention, the carcasses 11 or carcass regions can be selected for the extraction of characteristic characteristics in a predetermined region.Optionally, the numerical data of the recorded carcasses 11 or carcass areas can be compared by means of a computer unit 15 with numerical database data, in particular digital database images, contained in a database 16 in order to identify feature characteristics in a predetermined region of the carcasses or carcass areas. The device 10 further comprises at least one input unit (not shown in detail in the figures) configured to sequentially provide the numerical data as input data to the computer unit 15, wherein the computer unit 15 is configured to provide the data of the feature characteristics for display and / or forwarding to at least one downstream machine 20 for evaluating or processing the carcasses 11 or carcass areas based on the determined feature characteristics. The downstream machine 20 is shown in Fig.5c is shown by way of example as a cutting station for processing the carcasses 11. The machine 20 can alternatively also be designed as a display or similar. Preferably, the numerical data can be provided as input data to a first neural network 17 comprising the computer unit, wherein the first neural network 17 has been trained to localize and / or evaluate feature characteristics in a predetermined region of the carcasses 11 or carcass areas. More preferably, the numerical data can be provided as input data to the first neural network 17 comprising the computer unit 15 and / or to a second neural network 18 comprising the computer unit 15, wherein the first neural network 17 and / or the second neural network 18 is provided for the local detection and / or segmentation of the carcass 11, in particular of predetermined regions of the carcass or carcass areas. In Figs. 4a, 4b,4b and Fig. 5a to Fig. 5c, two neural networks 17, 18 are each schematically identified in the computer unit 15, wherein in preferred embodiments only one neural network 17 or more than two neural networks can be provided.

[0071] The device 10 according to the invention further comprises at least one further detection unit 19, 21, wherein the at least one further detection unit 19, 21 is arranged downstream of the first detection unit 14 in the conveying direction 12 in order to detect each of the carcasses 11 or carcass regions guided past the at least one further detection unit 19, 21 in order to generate, by means of the at least one further detection unit 19, 21 and / or by means of the computer unit 15, further numerical data, in particular further digital images, of the carcasses 11 or carcass regions detected by the at least one further detection unit 19, 21.

[0072] In a further preferred embodiment of the device 10, the additional numerical data generated by the at least one additional acquisition unit 19, 21 can be further processed by the computer unit 15, wherein the carcasses 11 or carcass areas can be selected for the extraction of characteristic expressions in a predetermined region. Optionally, the additional numerical data of the carcasses 11 or carcass areas acquired by the at least one additional acquisition unit 19, 21 can be compared by the computer unit 15 with numerical database data, in particular digital database images, contained in a database 16 in order to identify characteristic expressions in a predetermined region of the carcasses 11 or carcass areas.In a further alternative, the further numerical data, in particular the characteristic characteristics, of the carcasses 11 or carcass areas recorded by the at least one further recording unit 19, 21 are compared by means of the computer unit 15 with the numerical data, in particular the characteristic characteristics, of the first recording unit 14. The characteristic characteristics compared in this way can each be made available for display and / or forwarding to a downstream machine 20 for evaluating or processing the carcasses 11 or carcass areas based on the determined characteristic characteristics.

[0073] Preferably, at least one of the detection units 14, 19, 21 comprises at least one sensor element 22 for detecting carcasses 11 or carcass regions with numerical data, in particular digital images. In a preferred embodiment, at least one of the sensor elements 22 is selected from sensors for detecting / generating a visible spectrum 23, a UV spectrum 24, an IR spectrum 25, depth / 3D information 26, and / or X-ray data. Further preferably, at least one further detection unit 19, 21 is provided, which is designed and configured to detect further feature characteristics, such as geometry, color information, and / or texture properties of the carcass 11 or carcass region.

[0074] The device 10 according to the invention preferably comprises at least one discharge device 27—indicated by a schematic arrow in Fig. 4a and Fig. 5c—which is designed and configured to discharge carcasses 11 or carcass regions from the conveyor device 13 depending on the characteristic characteristics and / or an anomaly status. By means of the discharge device 27, a corresponding carcass 11 or carcass region can be discharged from the conveying process. As shown schematically in the preferred embodiment in Fig. 5c, for example, only a region of a carcass 11 can be separated, in particular by means of a machine 20, preferably a cutting station. This separated region can then be discharged by means of the discharge device 27 without discharging the entire carcass 11 from the conveyor device 13.

Claims

1. Method (100) for the analysis of carcasses (11) or parts thereof, in particular poultry carcasses or fish carcasses, comprising the steps - conveying (101) the carcasses (11) in a conveying direction (12) by means of a conveying device (13), - detecting (102) the carcasses (11) or a carcass area by means of a first detection unit (14), for each of the carcasses (11) or the carcass area which are passed by the first detection unit (14), - generating (103) numerical data, in particular digital images, of the carcasses (11) or carcass areas detected by the first detection unit (14) by means of the first detection unit (14) and / or by means of a computer unit (15), - processing (104) the numerical data of the recorded carcasses (11) or carcass areas by means of a computer unit (15), wherein the carcasses (11) or carcass areas are selected (105) for the extraction of characteristic expressions in a predetermined region, or wherein the numerical data of the recorded carcasses (11) or carcass areas are compared (106) by means of a computer unit (15) with numerical database data contained in a database (16), in particular digital database images, in order to identify (107) characteristic expressions in a predetermined region of the carcasses (11) or carcass areas, - Providing (108) the characteristic expressions for display (109) and / or forwarding (110) to at least one downstream machine (20) for evaluating (111) or processing (112) the carcasses (11) or carcass areas on the basis of the determined characteristic expressions.

2. Method (100) according to claim 1, characterized in that the numerical data are provided as input data to a first, in particular pre-trained, neural network (17), wherein the first neural network has been trained to localize and / or evaluate feature characteristics in a predetermined region of the carcasses (11) or carcass areas.

3. Method (100) according to claim 1 or 2, characterized in that the numerical data are provided as input data to the first neural network (17) and / or a second neural network (18), wherein the first neural network (17) and / or the second neural network (18) is provided for the local detection and / or segmentation of the carcass (11), in particular of predetermined regions of the carcass (11) or carcass areas.

4. Method (100) according to one or more of claims 1 to 3, characterized in that the predetermined region is selected (105) depending on the carcass (11) or the carcass area and / or is selected (105) depending on the design and / or the positioning of the detection unit.

5. Method (100) according to one or more of claims 1 to 4, comprising at least one further detection unit (19, 21), wherein the at least one further detection unit (19, 21) is arranged downstream of the first detection unit (14) in the conveying direction (12) in order to detect each of the carcasses (11) or carcass regions guided past the at least one further detection unit (19, 21) in order to generate, by means of the at least one further detection unit (19, 21) and / or by means of the computer unit (15), further numerical data, in particular further digital images, of the carcasses (11) or carcass regions detected by the at least one further detection unit (19, 21).

6. Method (100) according to claim 5, characterized in that the further numerical data generated by the at least one further detection unit (19, 21) are further processed by the computer unit (15), wherein the carcasses (11) or carcass areas are selected for the extraction of feature characteristics in a predetermined region (105) or wherein the further numerical data of the carcasses (11) or carcass areas recorded by the at least one further recording unit (19, 21) are compared (106) by means of the computer unit (15) with numerical database data, in particular digital database images, contained in a database (16) in order to identify feature characteristics in a predetermined region of the carcasses (11) or carcass areas, or wherein the further numerical data, in particular the feature characteristics, of the carcasses (11) or carcass areas recorded by the at least one further recording unit (19, 21) are compared by means of the computer unit (15) with the numerical data, in particular the feature characteristics, of the first recording unit (14),wherein the characteristic characteristics are provided for display (109) and / or forwarding (110) to a downstream machine (20) for evaluation (111) or processing (112) of the carcasses (11) or carcass areas on the basis of the determined characteristic characteristics.

7. Method (100) according to one or more of claims 1 to 6, characterized in that at least one of the detection units (14, 17, 21) comprises at least one sensor element (22) for detecting (102) carcasses (11) or carcass areas with numerical data, in particular digital images.

8. Method (100) according to one or more of claims 1 to 7, characterized in that the feature characteristics are recorded and compared (106) with reference features of known and defined anomalies stored in the database (16) and / or with the feature characteristics determined by the first recording unit (14) or at least one further recording unit (19, 21).

9. Method (100) according to claim 8, characterized in that based on the comparison (106) of the feature characteristics, in particular Using the first neural network (17), an anomaly status of the feature values ​​is determined.

10. Method (100) according to claim 9, characterized in that an action (114) is carried out depending on the characteristic values ​​and / or the anomaly status, wherein the action (114) comprises in particular the automatic discharge (115) of the carcass (11) or carcass region from the conveyor device (13).

11. The method (100) according to claim 10, characterized in that the action (114) comprises automatically sorting (116) the carcasses (11) or the carcass areas into a plurality of categories based on the anomaly status and / or the characteristic values.

12. Method (100) according to claim 10 or 11, characterized in that the action (114) comprises an assessment (111) of the quality of the carcasses (11) or carcass areas on the basis of the detected characteristic values, the anomalies and / or the anomaly status.

13. Method (100) according to one or more of claims 1 to 12, characterized in that at least one further detection unit (19, 21) is used to detect (102) further feature characteristics, such as geometry, color information and / or texture properties of the carcass (11) or carcass region, in order to carry out an additional evaluation (111) of the anomaly status by means of at least one further neural network (18).

14. Method (100) according to one or more of claims 1 to 13, characterized in that a processing (112) or an evaluation (111) of the detected feature values ​​is carried out by a pre-trained machine learning model.

15. Method (100) according to one or more of claims 1 to 14, characterized in that the method (100) is carried out essentially in real time carried out to enable continuous monitoring of the movement of carcasses (11) or carcass areas.

16. Method (200) for training at least one neural network (17, 18) for the analysis of carcasses (11) or parts thereof, in particular poultry carcasses or fish carcasses, comprising the following steps: - Providing (201) a plurality of carcasses (11) or parts thereof, in particular poultry carcasses or fish carcasses, - detecting (202) the carcasses (11) or a carcass area by means of a first detection unit (14), - generating (203) numerical data, in particular digital images, of the carcasses (11) or carcass areas detected by the first detection unit (14) by means of the first detection unit (14) and / or by means of a computer unit (15), - determining (204) at least one reference value (205) for identifying the positions and / or the variation of at least one characteristic expression in a predetermined region of the carcass (11) or carcass area on the basis of the numerical data, - inputting (206) the numerical data as input data (207) and the at least one reference value (205) as target data as training data for the neural network (17, 18), - repeatedly adjusting (208) the weights of the neural network (17, 18) based on the difference between the target data and the output data (209) generated by the neural network (17, 18).

17. The method (200) according to claim 16, characterized in that the reference values ​​(205) are sorted into known or defined anomalies, wherein an anomaly status of the feature values ​​is formed and established.

18. Method (200) according to claim 16 or 17, characterized in that after the detection (202) of the carcass (11) or the carcass area by the first detection unit (14), a modification (210) of the carcass (11) or the carcass area in the range of at least one reference value (205).

19. Method (200) according to claim 17 or 18, characterized in that the carcasses (11) or carcass areas are sorted, evaluated and / or provided for further processing on the basis of the detected characteristic characteristics, anomalies and / or the anomaly status.

20. Method (200) according to claim 18 or 19, characterized by detecting (202) the carcass (11) or the carcass region after the modification (210) by means of at least one further detection unit (19, 21) in order to generate (203) further numerical data, in particular further digital images, of the carcass (11) or carcass regions detected by the at least one further detection unit (19, 21) by means of the at least one further detection unit (19, 21) and / or by means of the computer unit (15).

21. Method (200) according to claim 20, characterized by determining a modification classification for evaluating the modification (210) carried out, in particular the qualitative modification, in the range of the at least one reference value (205) on the basis of the further numerical data of the at least one further detection unit (19, 21).

22. Method (200) according to claim 21, characterized by - inputting (206) the further numerical data of the at least one further acquisition unit (19, 21) as further input data (207) and modification classification as further target data in order to carry out a comparison, in particular a qualitative comparison, between the modification classification carried out and the modification carried out (210), - repeatedly adjusting (208) the weights of the neural network (17, 18) on the basis of the difference between the further target data and the output data (209) generated by the neural network (17, 18).

23. Method (200) according to one or more of claims 16 to 22, characterized in that the at least one reference value (205) is at least one value which respectively references the position or the qualitative nature of the feature characteristics and / or the anomaly status.

24. Method (200) according to claim 23, characterized in that the at least one reference value (205) for the qualitative condition is referenced on the basis of database data stored in the neural network (17, 18) for classifying the qualitative assessment of the condition of the feature characteristics and / or the anomaly status.

25. A non-transitory computer-readable storage medium comprising a program including instructions for causing the computer to perform the method (200) of any one of claims 16 to 24.

26. Device (10) for analyzing carcasses (11) or parts thereof, in particular poultry carcasses or fish carcasses, comprising a conveyor device (13) configured to convey the carcasses (11) or parts thereof in a conveying direction (12), a first detection unit (14) designed and configured to detect the carcasses (11) or a carcass area, for each of the carcasses (11) or the carcass area that can be guided past the first detection unit (14), wherein numerical data, in particular digital images, of the carcasses (11) or carcass areas detected by the first detection unit (14) can be generated by means of the first detection unit (14) and / or by means of a computer unit (15), a computer unit (15) provided for processing the numerical data,wherein the carcasses (11) or carcass areas are selectable for extracting characteristic expressions in a predetermined region, or wherein the numerical data of the recorded carcasses (11) or carcass areas are compared by means of a computer unit (15) with numerical database data contained in a database (16), in particular, digital database images, in order to identify feature characteristics in a predetermined region of the carcasses (11) or carcass areas, at least one input unit configured to sequentially provide the numerical data as input data to the computer unit (15), wherein the computer unit (15) is configured to provide the data of the feature characteristics for display and / or forwarding to at least one downstream machine (20) for evaluating or processing the carcasses (11) or carcass areas on the basis of the determined feature characteristics.

27. Device (10) according to claim 26, characterized in that the numerical data can be provided as input data to a first neural network (17) comprising the computer unit (15), wherein the first neural network (17) has been trained to localize and / or evaluate feature characteristics in a predetermined region of the carcasses (11) or carcass areas.

28. Device (10) according to claim 26 or 27, characterized in that the numerical data can be provided as input data to the first neural network (17) comprising the computer unit (15) and / or to a second neural network (18) comprising the computer unit (15), wherein the first neural network (17) and / or the second neural network (18) is provided for the local detection and / or segmentation of the carcass (11), in particular of predetermined regions of the carcass (11) or carcass areas.

29. Device (10) according to one or more of claims 26 to 28, further comprising at least one further detection unit (19, 21), wherein the at least one further detection unit (19, 21) is arranged downstream of the first detection unit (14) in the conveying direction (12) in order to detect each of the carcasses (11) or carcass areas guided past the at least one further detection unit (19, 21), in order to use the at least one further detection unit (19, 21) and / or to generate further numerical data, in particular further digital images, of the carcasses (11) or carcass areas recorded by the at least one further recording unit (19, 21) by means of the computer unit (15).

30. Device (10) according to one or more of claims 26 to 29, characterized in that the further numerical data generated by the at least one further detection unit (19, 21) can be further processed by the computer unit (15), wherein the carcasses (11) or carcass areas can be selected for the extraction of feature characteristics in a predetermined region, or wherein the further numerical data of the carcasses (11) or carcass areas detected by the at least one further detection unit (19, 21) can be compared by the computer unit (15) with numerical database data contained in a database (16), in particular digital database images, in order to identify feature characteristics in a predetermined region of the carcasses (11) or carcass areas, or wherein the further numerical data, in particular the feature characteristics,the carcasses (11) or carcass areas recorded by the at least one further recording unit (19, 21) are comparable by means of the computer unit (15) with the numerical data, in particular the characteristic characteristics, of the first recording unit (14), wherein the characteristic characteristics can be provided for display and / or forwarding to a downstream machine (20) for evaluating or processing the carcasses (11) or carcass areas on the basis of the determined characteristic characteristics.

31. Device (10) according to one or more of claims 26 to 30, characterized in that at least one of the detection units (14, 17, 21) comprises at least one sensor element (22) for detecting carcasses (11) or carcass areas with numerical data, in particular digital images.

32. Device (10) according to one or more of claims 26 to 31, further comprising at least one discharge device (27) which is designed and configured to discharge carcasses (11) or carcass regions from the conveyor device (13) depending on the characteristic characteristics and / or an anomaly status.

33. Device (10) according to one or more of claims 26 to 32, further comprising at least one further detection unit (19, 21) which is designed to detect further feature characteristics, such as geometry, colour information and / or texture properties of the carcass (11) or Carcass area is trained and equipped.

34. A non-transitory computer-readable storage medium comprising a program including instructions for causing the computer to perform the method (100) according to one or more of claims 1 to 15.

Images