融合特征增強(qiáng)模塊的小樣本農(nóng)業(yè)害蟲(chóng)識(shí)別

王祎 李旭偉 劉怡光 陳立平

摘 要 ???：基于深度學(xué)習(xí)的圖像識(shí)別技術(shù)在具體應(yīng)用前必須先經(jīng)過(guò)大量帶標(biāo)簽樣本的訓(xùn)練，然而在實(shí)際場(chǎng)景中目標(biāo)域樣本可能非常稀缺，小樣本圖像識(shí)別技術(shù)應(yīng)運(yùn)而生.為了提升小樣本場(chǎng)景下的圖像識(shí)別準(zhǔn)確率，本文提出一個(gè)通用的兩階段訓(xùn)練模型以融合現(xiàn)行主流方法并增強(qiáng)其表現(xiàn).首先，針對(duì)訓(xùn)練時(shí)不同害蟲(chóng)種類(lèi)背景相似度過(guò)高的問(wèn)題提出融合雙注意力機(jī)制的特征加強(qiáng)模塊；其次，針對(duì)小樣本情況下預(yù)測(cè)可能產(chǎn)生的過(guò)擬合問(wèn)題提出基于高斯分布的特征生成模塊以提高泛化能力；最后，將三種典型小樣本識(shí)別方法統(tǒng)一成兩階段訓(xùn)練模型以融入提出的方法.將該思路及改進(jìn)首次應(yīng)用于傳統(tǒng)害蟲(chóng)分類(lèi)數(shù)據(jù)集IP102，識(shí)別準(zhǔn)確率可以在基準(zhǔn)方法上取得2.11%到6.87%的提升.為了進(jìn)一步驗(yàn)證本文方法的有效性，在小樣本領(lǐng)域公開(kāi)數(shù)據(jù)集Mini-Imagenet也進(jìn)行了相應(yīng)的實(shí)驗(yàn)，提升效果同樣顯著.

關(guān)鍵詞 ： 圖像識(shí)別； 小樣本； 特征增強(qiáng)； 農(nóng)業(yè)害蟲(chóng)

中圖分類(lèi)號(hào) ：S126 文獻(xiàn)標(biāo)識(shí)碼 ：A DOI ： ?10.19907/j.0490-6756.2023.042001

Few shot learning of agricultural pests classification ?fusion with enhanced feature model

WANG Yi ?1， LI ?Xu-Wei ?1， LIU Yi-Guang ?1， CHEN Li-Ping ?2

（1. College of Computer Science （College of Software）， Sichuan University， Chengdu 610065， China；

2. School of Information Engineering， Tarim University， Tarim 843300， China）

In order to achieve accurate image recognition in scenarios where the target domain samples are limited，such as agricultural pest Image recognition， few shot image classification methods have been developed as an extension of deep learning-based image classification .To further improve the accuracy in the few shot image classification， this paper proposes a general two-stage training model that integrates current mainstream methods and enhances their performance to improve the recognition accuracy in limited sample scenarios Firstly， a feature enhancement module incorporating dual attention mechanism is proposed to solve the problem that the background similarity of different pest species is too high during training. Secondly， a feature generation module based on Gaussian distribution is proposed to solve the problem of overfitting that may occur in prediction in the case of a single sample. to improve the generalization ability. Finally， three typical few-shot recognition methods are unified into a two-stage training model to incorporate the proposed method. This idea and improvement are applied to the traditional pest classification dataset IP102 for the first time， and the recognition accuracy can be improved by 2.11% to 6.87% over the benchmark method. In order to further verify the effectiveness of the method in this paper， corresponding experiments were also carried out on the public dataset Mini-Imagenet in the field of few shot learning， the improvement effect is also significant.

Imagine classification; Few shot learning; Feature enhancement; Agricultural pests

1 引 言

農(nóng)業(yè)問(wèn)題關(guān)乎民生大計(jì)，種類(lèi)繁多的害蟲(chóng)卻給糧食生產(chǎn)和作物安全帶來(lái)了巨大的挑戰(zhàn) ?［1］，因此安全高效地識(shí)別農(nóng)業(yè)害蟲(chóng)尤為重要.同時(shí)基于深度學(xué)習(xí)的圖像識(shí)別技術(shù)也取得了飛速的進(jìn)展，各種改進(jìn)的卷積神經(jīng)網(wǎng)絡(luò) ?［2-4］和Transformer機(jī)制 ?［5-6］在某些特定場(chǎng)景下的表現(xiàn)已經(jīng)超越人類(lèi)，面對(duì)經(jīng)濟(jì)與效率的取舍，有學(xué)者在農(nóng)業(yè)害蟲(chóng)識(shí)別領(lǐng)域使用機(jī)器視覺(jué)方法進(jìn)行了各種積極的嘗試.