Learning method and learning device for setting parameters of CNN in which remaining networks are provided for meta-learning, and testing method and testing device using thereof

Description

[0001] The present disclosure relates to a computer implemented convolutional neural network (CNN)-based learning method for a CNN to perform image segmentation. The present disclosure further relates to a learning device for a CNN to perform image segmentation .

[0002] Deep learning is a technology used to cluster or classify objects or data. For example, computers cannot distinguish dogs and cats from photographs alone. But a human can easily distinguish those two. To this end, a method called "machine learning" was devised. It is a technique to allow a computer to classify similar things among lots of data inputted thereto. When a photo of an animal similar to a dog is inputted, the computer may classify it as a dog photo.

[0003] There have already been many machine learning algorithms to classify various data. For example, a decision tree, a Bayesian network, a support vector machine (SVM), an artificial neural network, etc. have been developed. The deep learning is a descendant of the artificial neural network.

[0004] Deep Convolution Neural Networks (Deep CNNs) are the heart of the remarkable development in deep learning. CNNs have already been used in the 90's to solve the problem of character recognition, but their use has become as widespread as it is now thanks to recent researches. These deep CNNs won the 2012 ImageNet image classification tournament, crushing other competitors. Then, the convolution neural network has become a very useful tool in the field of the machine learning.

[0005] Image segmentation is a way to receive an image as an input and produce a labeled image as an output. As deep learning technology has recently become popular, the image segmentation also adopts the deep learning. According to a conventional technology, the image segmentation was performed by (i) applying one or more convolution operations to an input image to thereby generate a feature vector and (ii) applying one or more fully connected operations to the feature vector to thereby generate a label image. According to another conventional technology, an encoder-decoder configuration was designed to extract features on an input image by using an encoder and reconstruct a label image by using a decoder.

[0006] Fig. 1 is a drawing schematically illustrating a process of performing the segmentation by using a general convolutional neural network (CNN).

[0007] Referring to Fig. 1, according to a conventional lane detection method, a learning device receives an input image and applies a plurality of convolution operations and non-linear operations such as ReLU at a plurality of convolutional layers to the input image, to thereby obtain a plurality of feature maps, and applies a plurality of deconvolution operations and SoftMax operations at a plurality of deconvolutional layers to the feature maps, to thereby obtain segmentation results.

[0008] Figs. 2A and 2B are drawings illustrating dropout in a neural network.

[0009] Dropout is disclosed for example in Nitish Srivastava ET AL: "Dropout: a simple way to prevent neural networks from overfilling", Journal of Machine Learning Research, vol. 15, no. 1, 1 June 2014 (2014-06-01), pages 1929-1958, XP055193568, ISSN: 1533-7928.

[0010] The dropout is frequently used in a fully connected (FC) layer, and since all nodes of the FC layer are connected, overfitting may frequently occur, which is a problem. For example, while performing a learning process of recognizing a dog, if training images include images of white dogs, only a white dog may be recognized as a dog. This is a problem generated due to a small number of the training images or a large capacity of a network.

[0011] Figs. 2A and 2B illustrate a configuration of a neural network without the dropout and that with the dropout. Herein, the dropout is one of methods that are introduced to solve such a problem. Fig. 2A is a drawing illustrating a configuration of a neural network having two FC layers. Referring to Fig. 2A, each of values of nodes in one FC layer is calculated by applying its corresponding weight to (i) each of element values of each of inputs or (ii) each of element values of each of nodes in its corresponding previous FC layer.

[0012] Fig. 2B illustrates a configuration of a neural network in which some nodes are dropped out. Herein, the neural network performs a learning process while one or more arbitrary nodes in one or more FC layers are dropped out. The nodes denoted as "X" in Fig. 2B are dropped out.

[0013] Likewise, not the entire weight may be set to be contributed to the learning process, but only some nodes may be set to be contributed to the learning process. The selected nodes are randomly changed for each learning process.

[0014] However, a method for improving a performance of the learning process by using one or more residual networks, without dropping out inner nodes of each layer of the neural network, is required.

[0015] It is an object of the present disclosure to solve all the aforementioned problems.

[0016] It is another object of the present disclosure to provide a method for dropping out a layer itself without dropping out some nodes in a layer.

[0017] It is still another object of the present disclosure to provide a method for resolving an over-fitting problem without increasing an amount of learning data or reducing the number of features.

[0018] In accordance with one aspect of the present disclosure, there is provided a computer implemented convolutional neural network (CNN)-based learning method for a CNN to perform image segmentation as claimed in claim 1.

[0019] Embodiments of the CNN-based learning method of claim 1 are claimed in claims 2 and 3.

[0020] In accordance with another aspect of the present disclosure, there is provided a method for testing as claimed in claim 4.

[0021] Embodiments of the testing method of claim 4 are claimed in claims 5,6 and 7.

[0022] In accordance with still another aspect of the present disclosure, there is provided a learning device as claimed in claim 8

[0023] In accordance with still yet another aspect of the present disclosure, there is provided a testing device as claimed in claim 9.

[0024] The above and other objects and features of the present disclosure will become apparent from the following description of preferred embodiments given in conjunction with the accompanying drawings, in which: Fig. 1 is a drawing schematically illustrating a segmentation process by using a conventional convolutional neural network (CNN). Figs. 2A and 2B are drawings illustrating a configuration of a neural network without the dropout and that with the dropout. Fig. 3 is a drawing illustrating a configuration of a CNN including residual networks and bridge layers in accordance with the present disclosure. Fig. 4 is a drawing illustrating a configuration of a plurality of residual networks in a convolutional layer of a reference number 400 in Fig. 3. Fig. 5 is a flowchart illustrating a learning process by dropping out one or more layers in accordance with the present disclosure. Fig. 6 is a drawing illustrating a transformed CNN in which one or more residual networks and one or more bridge layers are dropped out in accordance with the present disclosure.

[0025] Detailed explanation on the present disclosure to be made below refer to attached drawings and diagrams illustrated as specific embodiment examples under which the present disclosure may be implemented to make clear of purposes, technical solutions, and advantages of the present disclosure. These embodiments are described in sufficient detail to enable those skilled in the art to practice the disclosure.

[0026] Besides, in the detailed description and claims of the present disclosure, a term "include" and its variations are not intended to exclude other technical features, additions, components or steps. Other objects, benefits, and features of the present disclosure will be revealed to one skilled in the art, partially from the specification and partially from the implementation of the present disclosure. The following examples and drawings will be provided as examples but they are not intended to limit the present disclosure.

[0027] Moreover, the present disclosure covers all possible combinations of example embodiments indicated in this specification. It is to be understood that the various embodiments of the present disclosure, although different, are not necessarily mutually exclusive. For example, a particular feature, structure, or characteristic described herein in connection with one embodiment may be implemented within other embodiments without departing from the scope of the claims. In addition, it is to be understood that the position or arrangement of individual elements within each disclosed embodiment may be modified without departing from the spirit and scope of the claims. The following detailed description is, therefore, not to be taken in a limiting sense, and the scope of the present disclosure is defined only by the appended claims, appropriately interpreted, according to the protocol under Article 69 EPC. In the drawings, like numerals refer to the same or similar functionality throughout the several views.

[0028] Any images referred to in the present disclosure may include images related to any roads paved or unpaved, in which case the objects on the roads or near the roads may include vehicles, persons, animals, plants, buildings, flying objects like planes or drones, or any other obstacles which may appear in a road-related scene, but the scope of the present disclosure is not limited thereto. As another example, said any images referred to in the present disclosure may include images not related to any roads, such as images related to alleyway, land lots, sea, lakes, rivers, mountains, forests, deserts, sky, or any indoor space, in which case the objects in said any images may include vehicles, persons, animals, plants, buildings, flying objects like planes or drones, ships, amphibious planes or ships, or any other obstacles which may appear in a scene related to alleyway, land lots, sea, lakes, rivers, mountains, forests, deserts, sky, or any indoor space, but the scope of the present disclosure is not limited thereto.

[0029] To allow those skilled in the art to the present disclosure to be carried out easily, the example embodiments of the present disclosure by referring to attached diagrams will be explained in detail as shown below.

[0030] Fig. 3 is a drawing illustrating a configuration of a CNN including residual networks and bridge layers in accordance with the present disclosure.

[0031] Referring to Fig. 3, a learning device based on a convolutional neural network, i.e., CNN, in accordance with the present disclosure includes a first convolutional layer 100_1, a second convolutional layer 100_2, ... , an N-th convolutional layer 100_N, and an N-th deconvolutional layer 200_N, ... , a second deconvolutional layer 200_2, a first deconvolutional layer 200_1. Further, an intermediate layer, i.e., the bridge layer, exits between each convolutional layer and its corresponding deconvolutional layer. Herein, in the configuration of the CNN in accordance with the present disclosure, the bridge layers are not essential elements. A first bridge layer 300_1 is arranged between the first convolutional layer 100_1 and the first deconvolutional layer 200_1. A second bridge layer 300_2 is arranged between the second convolutional layer 100_2 and the second deconvolutional layer 200_2. An N-th bridge layer 300_N is arranged between the N-th convolutional layer 100_N and the N-th deconvolutional layer 200_N, and the like. Further, each of the convolutional layers includes each of residual networks 110_1 to 110_N capable of feed-forwarding an input image or its corresponding encoded input feature maps to its corresponding next convolutional layer through bypassing at least one of the convolutional layers or at least one of sub-convolutional layers included in at least one of the convolutional layers.

[0032] Fig. 3 illustrates that each of the residual networks 110_1 to 110_N is arranged between its corresponding convolutional layer and its corresponding deconvolutional layer, but this is just briefly illustrated. A configuration of one of the residual networks will be described in detail by referring to Fig. 4. Further, though not illustrated in Fig. 3, at least one of the residual networks bypassing at least one of sub-deconvolutional layers also exist in each of the deconvolutional layers. Herein, the bridge layer may be a dilated convolutional layer.

[0033] Fig. 4 is a drawing illustrating a configuration of the residual networks in a convolutional layer, i.e., a layer with a reference number 400 in Fig. 3.

[0034] Fig. 4 illustrates a configuration of the third convolutional layer 100_3 in which three sub-convolutional layers 100_3_a, 100_3_b, and 100_3_c, and three ReLU layers are arranged. Namely, in the third convolutional layer 100_3, the first sub-convolutional layer 100_3_a, a first ReLU layer, the second sub-convolutional layer 100_3_b, a second ReLU layer, the third sub-convolutional layer 100_3_c, and a third ReLU layer are sequentially arranged.

[0035] Further, the third convolutional layer 100_3 includes (i) a first residual network 110_3_a for feed-forwarding a second encoded output feature map, outputted from the second convolutional layer 100_2, through bypassing the first sub-convolutional layer 100_3_a, the first ReLU, the second sub-convolutional layer 100_3_b, the second ReLU and the third sub-convolutional layer 100_3_c, (ii) a second residual network 110_3_b for feed-forwarding a first non-linear output feature map, outputted from the first ReLU, through bypassing the second sub-convolutional layer 100_3_b, the second ReLU and the third sub-convolutional layer 100_3_c, and (iii) a third residual network 110_3_c for feed-forwarding a second non-linear output feature map, outputted from the second ReLU, through bypassing the third sub-convolutional layer 100_3_c.

[0036] At an input end of the third ReLU, (i) a feature map outputted from the second convolutional layer 100_2, i.e., the second encoded output feature map, (ii) a feature map outputted from the first ReLU, i.e., the first non-linear output feature map, (iii) a feature map outputted from the second ReLU, i.e., the second non-linear output feature map, and (iv) a feature map outputted from the third sub-convolutional layer 100_3_c are concatenated. Further, the concatenated feature map is fed into the third ReLU, to thereby generate a third non-linear output feature map. Herein, the third non-linear output feature map functions as a third encoded output feature map. The third encoded output feature map may be delivered to the third deconvolutional layer via the third bridge layer 300_3.

[0037] Further, locations of the ReLU layers and bypassing routes may be variously changed depending on the network design.

[0038] Through such a configuration, various feature maps, i.e., a feature map generated by passing through all the sub-convolutional layers, a feature map generated by passing through two sub-convolutional layers, a feature map generated by passing through only one sub-convolutional layer and a feature map generated without passing through any sub-convolutional layer, can be used.

[0039] Fig. 5 is a flowchart illustrating a learning process by dropping out one or more layers in accordance with the present disclosure.

[0040] Referring to Fig. 5, a process of learning the CNN by dropping out one or more layers in accordance with the present disclosure is described as follows. First of all, if an input image, e.g., a training image, is fed into the learning device, the learning device randomly selects a specific residual network and a specific bridge layer to be dropped out among the residual networks and the bridge layers at a step of S01.

[0041] Thereafter, the learning device configures a transformed CNN in which the specific residual network and the specific bridge layer are dropped out at a step of S02.

[0042] Fig. 6 is a drawing illustrating a transformed CNN in which one or more residual networks and one or more bridge layers are dropped out in accordance with the present disclosure.

[0043] Referring to Fig. 6, in the third convolutional layer 100_3, a first residual network 110_3 _a and a second residual network 110_3 _b are selected as the specific residual networks to be dropped out, and in the (1-4)-th convolutional layer 100_4, a first residual network 110_4_a and a third residual network 110_4_c are selected as the specific residual networks to be dropped out. Further, the fourth bridge layer 300_4 is also selected as the specific bridge layer to be dropped out. That is, the transformed CNN is configured such that routes or a layer denoted as "X" is dropped out as shown in Fig. 6.

[0044] When considering the entire learning device, each of the N convolutional layers and each of the N deconvolutional layers include L sub-convolutional layers and L sub-deconvolutional layers, respectively. Each of the N convolutional layers and each of the N deconvolutional layers include M sub-residual networks having each different bypassing route through which at least one of the L sub-convolutional layer and at least one of the L sub-deconvolutional layer is bypassed respectively. Further, when the number of the intermediate layers is K, at least one specific intermediate layer and at least one specific sub-residual networks to be dropped out are randomly selected among the K intermediate layers and the 2xNxM sub-residual networks, at a step of S02.

[0045] As another example, at the step of S02, only the residual network(s) may be selected as the layer(s) to be dropped out.

[0046] As specified herein, the N convolutional layers each include L sub-convolutional layers, and the N convolutional layers each include M sub-residual networks having each different bypassing route for bypassing at least one of the L sub-convolutional layer, and at the step of S02, the specific sub-residual network to be dropped out are randomly selected among the NxM sub-residual networks. This may be applied to both the case that the learning device includes the bridge layers and the case that the learning device does not include the bridge layers.

[0047] Referring to Fig. 5 again, the learning device generates adjusted encoded output feature maps from the input image, based on the transformed CNN to which the dropout has been applied and generates a CNN output by using the adjusted encoded output feature maps at a step of S03. Herein, the learning device calculates one or more losses by using the CNN output and its corresponding GT and adjusts at least one parameter of the transformed CNN by backpropagating the losses at a step of S04.

[0048] Meanwhile, the learning method may be applied to a same input image by repeating the above-mentioned processes or may be applied to various input images by changing the configuration of the transformed CNN.

[0049] Further, the learning method illustrated in Figs. 2 to 6 may be applied to the CNN testing method in the same manner.

[0050] For reference, in the description below, the phrase "for training" is added for terms related to the learning processes, and the phrase "for testing" is added for terms related to testing processes, to avoid possible confusion.

[0051] Namely, a method for testing a test image based on a convolutional neural network (CNN) includes steps of: (a) on condition that, assuming that a learning device includes (i) one or more convolutional layers capable of applying one or more convolution operations to a training image or its corresponding one or more encoded input feature maps for training to thereby sequentially generate one or more encoded output feature maps for training and (ii) one or more residual networks capable of feed-forwarding the training image or its corresponding encoded input feature maps for training to its corresponding next convolutional layer through bypassing at least one of the convolutional layers or at least one of sub-convolutional layers included in at least one of the convolutional layers, (1) the learning device has performed a process of randomly selecting a specific residual network to be dropped out among the residual networks, (2) the learning device (i) has fed the training image into at least one transformed CNN in which the specific residual network is dropped out, to thereby generate adjusted encoded output feature maps for training and (ii) has generated a CNN output for training by using the adjusted encoded output feature maps for training, and (3) the learning device has calculated one or more losses by using the CNN output for training and its corresponding GT and has adjusted at least one parameter of the transformed CNN by backpropagating the losses, a testing device acquiring the test image; and (b) the testing device generating one or more encoded output feature maps for testing based on the test image and generating a CNN output for testing by using the encoded output feature maps for testing.

[0052] The present disclosure has an effect of improving a learning capability by randomly dropping out some layers within the CNN.

[0053] The present disclosure has another effect of performing a learning process by using the diversified feature maps.

[0054] Further, the CNN-based method for meta learning, i.e., learning to learning, in accordance with the present disclosure can be also applied to layer-wise dropout, stochastic ensemble, virtual driving, and the like.

[0055] It would be understood by one of ordinary skill in the art that a transmission and / or a reception of the above-described images such as the training image and the test image can be performed by communication parts of the learning device and the testing device, and processes of the convolution operation, the deconvolution operation, and the loss value operation can be mainly performed by processors of the learning device and the testing device, but the present disclosure is not limited to these examples. In addition, the learning device and the testing device may further include memories capable of storing computer readable instructions for performing the above-described processes. As one example, the processor, the memory, a medium, etc. may be integrated with an integrated processor.

[0056] The embodiments of the present disclosure as explained above can be implemented in a form of executable program command through a variety of computer means recordable to computer readable media. The computer readable media may include solely or in combination, program commands, data files, and data structures. The program commands recorded to the media may be components specially designed for the present disclosure or may be usable to a skilled human in a field of computer software. Computer readable media include magnetic media such as hard disk, floppy disk, and magnetic tape, optical media such as CD-ROM and DVD, magneto-optical media such as floptical disk and hardware devices such as ROM, RAM, and flash memory specially designed to store and carry out program commands. Program commands include not only a machine language code made by a complier but also a high level code that can be used by an interpreter etc., which is executed by a computer. The aforementioned hardware device can work as more than a software module to perform the action of the present disclosure and they can do the same in the opposite case.

[0057] As seen above, the present disclosure has been explained by specific matters such as detailed components, limited embodiments, and drawings. They have been provided only to help more general understanding of the present disclosure. It, however, will be understood by those skilled in the art that various changes and modification may be made from the description without departing from the scope of the following claims.

Claims

1. A computer implemented convolutional neural network (CNN)-based learning method for a CNN to perform image segmentation to an image provided at its input to produce a labeled image as an output by using a learning device, the CNN including a number of N convolutional layers (100_1, 100_2, 100_3, ...,100_N) capable of applying one or more convolution operations to an input image or its corresponding one or more encoded input feature maps to thereby sequentially generate one or more encoded output feature maps, the method comprising steps of: (a) the learning device, if the input image is acquired, performing a process of randomly selecting a unit of the CNN to be dropped out among the units; (b) the learning device (i) feeding the input image into at least one transformed CNN in which the specific unit is dropped out, to thereby generate adjusted encoded output feature maps and (ii) generating a CNN output by using the adjusted encoded output feature maps; and (c) the learning device calculating one or more losses by using the CNN output and its corresponding ground truth (GT) and adjusting at least one parameter of the transformed CNN by backpropagating the losses, characterized in that (i) each of the N convolutional layers of the CNN has L sub-convolutional layers (100_3_A, 100_3_B, 100_3_C; 100_4_A, 100_4_B, 100_4_C) and M sub-residual networks (110_3_A, 110_3_B, 110_3_C; 110_4_A, 110_4_B, 110_4_C) having each different bypassing route for bypassing at least one of the L sub-convolutional layers, (ii) the CNN further includes a number of N deconvolutional layers (200_1, 200_2, 200_3,...,200_N) capable of applying one or more deconvolutional operations to at least part of the adjusted encoded output feature maps or its corresponding adjusted decoded input feature maps to thereby sequentially generate one or more adjusted decoded output feature maps and each of the N deconvolutional layers includes L sub-deconvolutional layers, each of the N deconvolutional layers includes M sub-residual networks having each different bypassing route through which at least one of the L sub-deconvolutional layers is bypassed, wherein the NxM sub-residual networks include residual networks capable of feed-forwarding the adjusted decoded input feature maps to its corresponding next deconvolutional layer through bypassing at least one of the plurality of deconvolutional layers, (iii) the CNN further includes a plurality of K intermediate layers (300_1, 300_2, 300_3,...,300_N), which are located between at least one of the convolutional layers and at least one of the deconvolutional layers, capable of applying one or more convolution operations to one or more inputs fed thereto and then feeding one or more outputs therefrom into at least one of the deconvolutional layers, and in that, at the step of (a), the randomly selected unit to be dropped out among the units is at least one specific intermediate layer and at least one specific sub-residual network among the K intermediate layers and the 2xNxM sub-residual networks, respectively.

2. The CNN-based learning method of Claim 1, wherein the steps of (a) to (c) are repeated for the same input image.

3. The CNN-based learning method of Claim 1, wherein, at the step of (c), the learning device backpropagates the losses to thereby adjust at least one parameter of the deconvolutional layers (200_1, 200_2, 200_3,... ,200_N), the intermediate layers (300_1, 300_2, 300_3,...,300_N), and the convolutional layers (100_1, 100_2, 100_3, ...,100_N).

4. A computer implemented method for testing a test image based on a convolutional neural network (CNN) obtained with the method of either of the claims 1-3, comprising the steps of: a testing device acquiring the test image; and the testing device generating one or more encoded output feature maps for testing based on the test image and generating a CNN output for testing by using the encoded output feature maps for testing.

5. The method of Claim 4, wherein, on condition that the test device includes (i) the convolutional layers (100_1, 100_2, 100_3, ...,100_N) capable of applying the convolution operations to the test image or its corresponding one or more encoded input feature maps for testing to thereby sequentially generate the encoded output feature maps for testing and (ii) the residual networks (110_1, 110_2, 110_3, ...) capable of feed-forwarding the test image or its corresponding encoded input feature maps for testing to its corresponding next convolutional layer through bypassing at least one of the convolutional layers or at least one of the sub-convolutional layers (e.g. 100_3_A, 100_3_B, 100_3_C; 100_4_A, 100_4_B, 100_4_C), included in at least one of the convolutional layers (100_3, 100_4), at the step of (b), the testing device (b-i) performs a process of randomly selecting a certain residual network (110_3_A, 110_3_B; 110_4_A, 110_4_C) to be dropped out among the residual networks, and (b-ii) feeds the test image into the transformed CNN in which the certain residual network is dropped out, to thereby generate adjusted encoded output feature maps for testing, and then generates the CNN output for testing by using the adjusted encoded output feature maps for testing.

6. The method of Claim 5, wherein the testing device further includes: (i) one or more deconvolutional layers (200_1, 200_2, 200_3,... ,200_N) capable of applying one or more deconvolutional operations to at least part of the adjusted encoded output feature maps for testing or its corresponding adjusted decoded input feature maps for testing to thereby sequentially generate one or more adjusted decoded output feature maps for testing and (ii) one or more intermediate layers (300_1, 300_2, 300_3,...,300_N), which are located between at least one of the convolutional layers (100_1, 100_2, 100_3,...,100_N) and at least one of the deconvolutional layers, capable of applying one or more convolution operations to one or more inputs fed thereto and then feeding one or more outputs therefrom into at least one of the deconvolutional layers, wherein, at the step of (b-i), the testing device performs a process of randomly selecting the certain residual network and at least one certain intermediate layer to be dropped out among the residual networks and the intermediate layers, and wherein, at the step of (b-ii), the testing device controls the transformed CNN such that the certain residual network and the certain intermediate layer are dropped out.

7. The method of Claim 6, wherein the testing device further includes additional residual networks capable of feed-forwarding the adjusted decoded input feature maps for testing to its corresponding next deconvolutional layer through bypassing at least one of the plurality of deconvolutional layers, wherein, assuming that (i) each of the N convolutional layers includes L sub-convolutional layers (e.g. 100_3_A, 100_3_B, 100_3_C; 100_4_A, 100_4_B, 100_4_C) and (ii) each of the N deconvolutional layers (200_1, 200_2, 200_3,... ,200_N) includes L sub-deconvolutional layers, and the number of the intermediate layers is K, (1) each of the N convolutional layers includes M sub-residual networks (110_3_A, 110_3_B, 110_3_C; 110_4_A, 110_ 4_B, 110_4_C) having each different bypassing route through which at least one of the L sub-convolutional layers is bypassed and (2) each of the N deconvolutional layers includes M sub-residual networks having each different bypassing route through which at least one of the L sub-deconvolutional layers is bypassed, and wherein, at the step of (b-i), at least one certain intermediate layer and at least one certain sub-residual network to be dropped out are randomly selected among the K intermediate layers and the 2xNxM sub-residual networks, respectively.

8. A learning device for a CNN to perform image segmentation to an image provided at its input to produce a labeled image as an output in which the CNN comprise a number of N convolutional layers (100_1, 100_2, 100_3, ... ,100_N) capable of applying one or more convolution operations to an input image or its corresponding one or more encoded input feature maps to thereby sequentially generate one or more encoded output feature maps, comprising: at least one memory that stores instructions; and at least one processor configured to execute the instructions to: perform processes of (I) randomly selecting a unit of the CNN to be dropped out among a plurality of units, (II) (i) feeding the input image into at least one transformed CNN in which the specific unit is dropped out, to thereby generate adjusted encoded output feature maps and (ii) generating a CNN output by using the adjusted encoded output feature maps, and (III) calculating one or more losses by using the CNN output and its corresponding ground truth (GT) and adjusting at least one parameter of the transformed CNN by backpropagating the losses, characterized in that (i) each of the N convolutional layers of the CNN has L sub-convolutional layers (100_3_A, 100_3_B, 100_3_C; 100_4_A, 100_4_B, 100_4_C) and M sub-residual networks (110_3_A, 110_3_B, 110_3_C; 110_4_A, 110_4_B, 110_4_C) having each different bypassing route for bypassing at least one of the L sub-convolutional layers, (ii) the CNN further includes a number of N deconvolutional layers (200_1, 200_2, 200_3,...,200_N) capable of applying one or more deconvolutional operations to at least part of the adjusted encoded output feature maps or its corresponding adjusted decoded input feature maps to thereby sequentially generate one or more adjusted decoded output feature maps and each of the N deconvolutional layers includes L sub-deconvolutional layers, each of the N deconvolutional layers includes M sub-residual networks having each different bypassing route through which at least one of the L sub-deconvolutional layers is bypassed, wherein the NxM sub-residual networks include residual networks capable of feed-forwarding the adjusted decoded input feature maps to its corresponding next deconvolutional layer through bypassing at least one of the plurality of deconvolutional layers, (iii) the CNN further includes a plurality of K intermediate layers (300_1, 300_2, 300_3,...,300_N), which are located between at least one of the convolutional layers and at least one of the deconvolutional layers, capable of applying one or more convolution operations to one or more inputs fed thereto and then feeding one or more outputs therefrom into at least one of the deconvolutional layers, and in that the randomly selected unit to be dropped out among the units is at least one specific intermediate layer and at least one specific sub-residual network among the K intermediate layers and the 2xNxM sub-residual networks respectively.

9. A testing device for testing a test image based on a convolutional neural network (CNN) obtained with the method of either of claims 1-3, comprising: at least one memory that stores instructions; and at least one processor, configured to execute the instructions to: perform processes of generating one or more encoded output feature maps for testing based on the test image and generating a CNN output for testing by using the encoded output feature maps for testing.